SlideShare a Scribd company logo
IAES International Journal of Artificial Intelligence (IJ-AI)
Vol. 13, No. 1, March 2024, pp. 1038~1049
ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i1.pp1038-1049  1038
Journal homepage: http://ijai.iaescore.com
Multi-channel microseismic signals classification with
convolutional neural networks
Hongmei Shu1,2
, Ahmad Yahya Dawod1
, Worawit Tepsan1
, Lei Mu1,3
, Zheng Tang2
1
International College of Digital Innovation, Chiang Mai University, Chiang Mai, Thailand
2
School of Resources and Safety Engineering, Central South University, Changsha, China
3
Office of International Collaboration and Exchange, Chengdu University, Chengdu, China
Article Info ABSTRACT
Article history:
Received May 11, 2023
Revised Aug 16, 2023
Accepted Sep 8, 2023
Identifying and classifying microseismic signals is essential to warn of mines’
dangers. Deep learning has replaced traditional methods, but labor-intensive
manual identification and varying deep learning outcomes pose challenges.
This paper proposes a transfer learning-based convolutional neural network
(CNN) method called microseismic signals-convolutional neural network
(MS-CNN) to automatically recognize and classify microseismic events and
blasts. The model was instructed on a limited sample of data to obtain an
optimal weight model for microseismic waveform recognition and
classification. A comparative analysis was performed with an existing CNN
model and classical image classification models such as AlexNet,
GoogLeNet, and ResNet50. The outcomes demonstrate that the MS-CNN
model achieved the best recognition and classification effect (99.6%
accuracy) in the shortest time (0.31 s to identify 277 images in the test set).
Thus, the MS-CNN model can efficiently recognize and classify microseismic
events and blasts in practical engineering applications, improving the
recognition timeliness of microseismic signals and further enhancing the
accuracy of event classification.
Keywords:
Classification
Convolutional neural network
Microseismic events
Multi-channel waveform
Recognition
Transfer learning
This is an open access article under the CC BY-SA license.
Corresponding Author:
Ahmad Yahya Dawod
International College of Digital Innovation, Chiang Mai University
239 Huay Kaew Rd, Suthep, Mueang Chiang Mai District, Chiang Mai 50200, Thailand
Email: ahmadyahyadawod.a@cmu.ac.th
1. INTRODUCTION
Microseismic monitoring technology (MMT) has found extensive applications in underground
engineering for disasters and safety monitoring [1]. Specifically, it has been utilized for location
monitoring [2], [3], as well as forecasting and providing early warning systems for rock bursts [4], and mine
earthquake disasters during mining operations [5], [6]. The basic principle involves identifying microseismic
events by analyzing prominent features within the monitoring data [7]. Subsequently, the relevant parameters
of these events are analyzed to facilitate informed decision-making [8]. Due to the complex geological
environment of mines [9], many interference signals frequently mix with the recorded microseismic signals,
such as blasting, rock drilling, fan vibration, and other noises generated during engineering operations [10].
Therefore, the basis of MMT is to identify microseismic events quickly and accurately, which determines the
timeliness and effectiveness of mine safety early warning [11].
Earlier studies have investigated several strategies, such as conventional signal processing methods
and machine learning algorithms [12], [13], to identify and classify microseismic signals [14]. However, these
techniques have limitations regarding their efficiency and accuracy. Recently, convolutional neural
networks (CNNs) [15], [16] have shown great promise in identifying and categorizing images. Based on this
Int J Artif Intell ISSN: 2252-8938 
Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu)
1039
success, some researchers employed CNNs to successfully recognize and categorize microseismic signals [17].
However, these studies have primarily focused on single-channel signals, and further research is necessary to
explore the application of CNNs to multichannel signals. The higher dimensionality of multichannel signals
presents unique challenges, and the effectiveness of CNNs for these signals is an active research area.
Traditional methods for identifying microseismic events and blasts primarily rely on manual or
engineering experience, which often results in significant and parameter analysis errors as well as time delays.
Extensive research has been conducted by experts and scholars in the field of recognition and classification, as
evidenced by previous studies [18], [19]. Traditional methods typically rely on statistical analysis of waveform
characteristics to classify microseismic events and blasts. However, these methods often require manual
experience for selecting characteristic parameters, resulting in limited generalizability and robustness due to
their specific applicability to certain cases or mines. For example, researchers conducted an analysis on the
characteristic patterns of source parameters associated with microseismic events and blasts [20], [21]. They
developed a statistical model for the automatic identification of such events, taking into account the probability
density distribution of each parameter and its impact on identification accuracy. However, the recognition
effect of this method on different mine data samples is quite different, and the robustness is not high.
With the explosive growth of data, the use of machine learning for image classification has become
popular [22], [23]. Deep learning [24]–[26] is a widely used method at present, which automatically extracts
features through data-driven, and obtains the specific feature representation of the dataset based on the learning
and training of many samples. It is more efficient and accurate in the expression of data sets without manual
intervention. Lin et al. and Lin et al. [27], [28] have proposed two methods, deep convolutional neural network
and spatial pyramidal pool (DCNN-SPP) and DCNN and support vector machine (DCNN-SVM), to
automatically identify and classify multi-channel microseismic waveforms, with accuracy improved from
91.13% to 98.18%. Li et al. [10] introduced a deep learning method for microseismic waveform classification
based on computer vision, which analyzed both the waveform of the microseismic signal and the corresponding
spectrogram of the waveform. Four deep learning models were used for experiments, including VGG-16 [29],
ResNet18 [30], AlexNet [31], and their ensemble models. The results showed that these models perform well
in recognizing the features of waveform graphs and can accurately predict data categories, among which the
ensemble model had the highest recognition accuracy of 98%. The end-to-end automatic classification of the
original waveform was realized, and it proved that the identification effect of the original waveform and the
spectrogram was basically the same.
This study proposes a CNN-based model named microseismic signals convolutional neural
network (MS-CNN) to enhance the precision and effectiveness of microseismic event recognition and
classification. This study's goal is to assess how well CNNs and related transfer learning models work. The
researchers [32], [33] in accurately classifying microseismic events and blasts in mines. We developed and
trained multiple deep learning models, comprising three conventional models (AlexNet, GoogLeNet, and
ResNet50) and one CNN model to accomplish this. We assessed the classification performance of these models
using five metrics based on an identical dataset. After analyzing the experimental results, we pinpointed the
most effective model for this purpose, hoping that this research will assist users in making informed decisions.
2. METHOD
The application of CNNs for microseismic waveform recognition and classification is guided by
several fundamental principles. These include the capacity of deep learning algorithms to extract intricate
features from multi-channel data, the utilization of supervised learning to train the model, and the iterative
optimization of the model's parameters to minimize classification error. The CNN model is trained using
labeled data, where the model's parameters and weights are adjusted to minimize classification error. This
optimization process often involves iterative algorithms like stochastic gradient descent, which update the
model's parameters based on the discrepancy between expected and actual labels. Convolutional layers play a
crucial role in this process by applying filters to the input data to extract local features [34]. Additionally,
pooling layers are utilized to down-sample the feature maps, simplifying the computational complexity of the
model. Once trained, the CNN can classify new and unseen data by inputting it into the network and leveraging
the learned parameters to predict the type of microseismic event present in each segment.
2.1. Microseismic signals convolutional neural network
Figure 1 depicts the MS-CNN model's hierarchical structure, which comprises an input layer, two
convolutional layers, a pooling layer, two fully connected layers, and an output layer. It abstracts the image
into data layer by layer to provide the necessary feature representation, and then maps the features to
the task objective. Humans do not need to intervene in the feature selection process during CNN training. The
neural network automatically learns image local features through convolution and backpropagates the error
through the loss function to obtain the optimal convolution coefficients (weights and biases). The identification
 ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 1, March 2024: 1038-1049
1040
and classification principle of the MS-CNN model was mainly composed of two parts, one part featured
learning (convolution, activation function, and pooling layers), and the other part was classification (full
connection layer).
Figure 1. Structure of MS-CNN, a multichannel microseismic waveform recognition and classification model
2.1.1. Feature extraction
The convolutional and pooling layers are crucial to CNN's capacity to automatically extract image
features. The weight and bias parameters in the convolution operation are effectively solved during the CNN
learning process utilizing the dataset and loss function [35]. The value of the feature map is determined using
(1), where 𝑥 represents the input image, f denotes the convolution kernel, and 𝑚, 𝑛 represent the row and
column indices, respectively, of the calculation result [36].
𝐷[𝑚, 𝑛] = (𝑥 ∗ 𝑓)[𝑚, 𝑛] = ∑ ∑ 𝑓[𝑖, 𝑗]𝑥[𝑚 − 𝑖, 𝑛 − 𝑗]
𝑗
𝑖 (1)
The pooling layer, also known as the downsampling layer, reduces the dimension of the feature map
by pooling each feature map, thus reducing the number of parameters in CNN, and avoiding overfitting [37].
The window size of the pooling operation can be specified as any size. There are two main pooling operations:
max pooling and average pooling. The MS-CNN model used max pooling. The maximum pooling function
uses the maximum value in the small block as the feature output of the block. When the image undergoes small
changes such as translation, scaling, and rotation, it is still possible to obtain the maximum value at the same
position, which is the same as the response before the change, thus realizing the spatial invariance feature.
2.1.2. Classification
The classification of the MS-CNN was achieved by fully connected layers. Since MS-CNN was a
supervised learning, the model was trained based on the labeled training samples to obtain the weights of the
fully connected layer. When using the model for outcome recognition, the weights obtained from the training
of the model and the results calculated from the deep network after the previous convolution, activation
function, and pooling, were weighted and summed to obtain the predicted value of each outcome. Then the
maximum value was taken as the result of the recognition. As shown in Figure 1, the probability of a
microseismic event was finally calculated to be 1.00, and the probability of a blasting event was 0.00.
Therefore, the event corresponding to this figure was finally determined to be a microseismic event.
The output layer of the model computes the error using the cross-entropy loss function, which assesses
the similarity between the predicted output and the desired output. In the back-propagation process, this
information is then used to optimize or update the parameters of each layer. In (2) [38] represents the
mathematical expression for the cross-entropy loss function. Where N represents the sample size, K denotes
the number of categories, and tij =1 if sample i belongs to the jth category, otherwise tij=0. The variable yij
represents the probability that sample i belongs to the jth category.
𝐿𝑜𝑠𝑠 = −
1
𝑁
∑ ∑ 𝑡𝑖𝑗𝑙𝑛𝑦𝑖𝑗
𝐾
𝑗=1
𝑁
𝑖=1 (2)
Int J Artif Intell ISSN: 2252-8938 
Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu)
1041
The main advantages of MS-CNN are its ability to handle multichannel data. In microseismic
applications, multiple sensors are often used to record microseismic waves, and MS-CNN can use all the
available channels to improve the accuracy of the recognition and classification task. MS-CNN is trained using
a dataset of microseismic event waveforms with labeled classes. The model learns to recognize patterns in the
data and associate them with specific classes. Once trained, MS-CNN can be used to classify new microseismic
waveforms based on their features and similarities to the learned patterns.
2.2. Model construction
Figure 1 shows the parameter settings of each layer for the MS-CNN model, and the number of
learning parameters is shown in Table 1. Before the image was fed into the model, the color
image was preprocessed by transforming its size from 432×288 pixels to 100×100 pixels. After the image
(e.g., X0) is inputted into the model, first, the Input layer would be normalized to X0 (size
width*height*channels=100×100×3). The C1 layer used a (3×3) convolution kernel to perform 32
convolutions on the X0 to form a (32×98×98) tensor. Then the C2 layer continued to use the (3×3) convolution
kernel to perform 32 convolutions on the output data of the C1 layer, and finally obtained a (32×48×48)
structure tensor. In the C1 and C2 layers, the stride is (1×1) and (2×2), respectively. To facilitate feature
extraction and reduce the amount of computation, a pooling operation was performed on the output data after
the C2 layer by pooling a maximum of (3×3) P1 to form a pooling tensor of (32×46×46). After pooling, a fully
connected layer was added to flatten the P1 data into a one-dimensional vector, and the number of output
neurons in the Fc1 was set to 256. Finally, the output size of the Fc2 is 2, and the softmax function was added
after Fc2 to calculate the probability value of each waveform belonging to each category. After obtaining the
probability values of the two categories, the category corresponding to the maximum probability value was
selected as the output value in the output layer. The activation functions of MS-CNN all used ReLU (rectified
linear unit), which existed between C1 and C2, and between C2 and P1, respectively. Additionally, to hasten
CNN's training and lessen the sensitivity of the initialization of the network, we added a batch normalization
layer between the C1 and the ReLU layer. Mini-batch processing of all observations for each channel was
implemented at this layer. And adopted the dropout method to alleviate overfitting.
Table 1. Learning params numbers for the layers of the MS-CNN model
Layer type Learnable Params
Convolution layer C1 Weights (3×3×3×32); Bias (1×1×32) 896
Batch normalization layer Offset (1×1×32); Scale (1×1×32) 64
Convolution layer C2 Weights (3×3×32×32); Bias (1×1×32) 9248
Max pooling layer P1  0
Fully Connected layer Fc1 Weights (256×80000); Bias (256×1) 20480256
Fully Connected layer Fc2 Weights (2×256); Bias (2×1) 514
3. RESULTS AND ANALYSIS
To comprehensively evaluate the performance of CNN models, we conducted experimental tests using
existing field data. The training process and test results of the MS-CNN model were longitudinally analyzed
and compared to CNN [1], AlexNet, GoogLeNet, and ResNet50 models. Our research findings demonstrate
that the MS-CNN model surpassed other models in terms of accuracy and computational efficiency. This was
observed by achieving both high classification accuracy and low training time. These results suggest that the
MS-CNN model has significant potential for use in real-world applications.
3.1. Data
To proper data preparation is essential for the successful recognition and classification of multi-
channel microseismic waveforms using CNN-based models. Careful attention should be paid to data collection,
preprocessing, labeling, and splitting to ensure that the CNN is trained and tested on high-quality data.
Real-time data from a mine microseismic monitoring system in China's Shanxi province was used in this study.
A 26-channel microseismic monitoring system was installed by the engineers based on the mine's geological
setting and mining strategy. Due to the large volume of data and the similarity among each type of event, data
from March 29 to April 19, 2022, were selected for this study. This was a total of 22 days of data and a database
size of 6.85 GB, chosen for computational performance reasons.
Firstly, we filtered all event data from the database with at least 6 triggered sensors. Then, we used
the Python plotting library Matplotlib to draw a 2D waveform graph of the event based on the sample time
magnitude data, which was saved as an event waveform with six subplots as input. Next, we manually screened
and marked the noise events, such as rock drilling and mechanical vibration, to filter them out. Finally, we
obtained the multi-channel waveform of two types of typical blasts and microseismic events. Examples of
 ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 1, March 2024: 1038-1049
1042
waveforms for a microseismic event are shown in Figure 2(a), and those for a blast event can be seen in
Figure 2(b). The vertical axis shows the signal amplitude, and the horizontal axis depicts a time in
microseconds. The experimental sample contained a total of 1386 images, including 512 waveform images of
microseismic events and 874 waveform images of blasts.
The classification performance of various models using the same dataset will be compared and
examined, control variables were employed to ensure fair training and testing. The distribution of sample sizes
for each event type in the dataset is presented in Table 2. The dataset was split into two chronological sections:
the first 80% were used as the training set, while the final 20% were utilized as the test set. During the training
phase, the model was trained using waveform features. To mitigate the risk of overfitting, a validation set
comprising 20% (222 images) of the training data was randomly selected. This validation set was used for fine-
tuning hyperparameters and obtaining the optimal model. Finally, The model's capacity for prediction and
generalization was evaluated using the test set.
(a)
(b)
Figure 2. Example waveforms of microseismic event and blasting (a) microseismic event waveform and (b)
blasting waveform
Table 2. Data sets and their divisions
Type All Training set (80%) Testing set (20%)
Microseismic 512 410 102
Blast 874 699 175
Sum 1386 1109 277
Int J Artif Intell ISSN: 2252-8938 
Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu)
1043
3.2. Experiment
For the hyperparameter settings of CNN model training, different models had different initial learning
rates, and the best parameter values need to be selected after several network training sessions. The following
parameters were chosen for the MS-CNN model: i) pick stochastic gradient descent with momentum (SGDM)
as your optimization function; ii) define the first learning rate as 0.001; iii) choose six epochs as the maximum
number; iv) decide on 16 as the mini-batch size; v) select a 10-period validation frequency. Next, we examine
the manner in which the MS-CNN model performed in terms of the training results and the test outcomes,
respectively.
3.2.1. Training results and analysis
The loss and accuracy indicators are essential in assessing the progress of the learning process. During
training, the loss typically decreases over time as the model learns to better fit the training data. To monitor the
accuracy and loss changes during training, these values are often computed at regular intervals, such as after
each epoch or after a certain number of training batches. The accuracy and loss curves can then be plotted over
time to visualize how they change during training. In this study, a five-fold cross-validation was utilized during
the training process, and the changes in loss and accuracy of the MS-CNN model during training and validation
were depicted in Figure 3. The results show that accuracy increased continuously with training time and
eventually stabilized at 1.0, while the loss gradually decreased and converged to approximately 0.0. The loss
and accuracy values of the validation set were consistent with those of the training set. The accuracy of the
validation set eventually stabilized at 98.6%, with a loss value of 0.05. These trends in loss and accuracy are
in accordance with the changes expected during neural network training. The accuracy and loss metrics are
frequently employed to monitor the model's performance during the MS-CNN training process and decide
when to stop training.
Figure 3. Accuracy and loss changes in MS-CNN training process
The accuracy of MS-CNN refers to the percentage of correctly classified samples in the training or
validation set. As the model learns from the training data, the accuracy typically increases, although it may
eventually plateau or even decrease if the model overfits the training data. The loss function employed in MS-
CNN quantifies the disparity between the predicted labels and the actual labels for each input sample. During
the training process, the objective is to minimize this loss function, thereby enhancing the model's capacity to
generate precise predictions.
3.2.2. Test results and analysis
After the MS-CNN model training was completed, a new test dataset was used for prediction and
classification, thus enabling analysis of the model performance. The identification results and time
consumption of a CNN model can vary depending on the specific architecture, dataset, and hardware used to
run the model. However, in general, the performance and time consumption of CNN models can be evaluated
using the following metrics: The accuracy calculates the proportion of test set samples that were properly
categorized. Better performance is indicated by higher accuracy. Precision counts the proportion of real
positives among all correctly predicted positive outcomes. Less false positives are indicated by increased
precision. The F1 score serves as a balanced metric, considering both precision and recall, and is calculated as
 ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 1, March 2024: 1038-1049
1044
the harmonic mean of these two measures. To assess the classification prediction performance of the model,
we employed five evaluation metrics: precision, recall, F1 score, accuracy, and testing time. The evaluation
measures for single-category classification include precision, recall, and F1 score. The overall performance of
the model is assessed using the accuracy metric and testing time.
As shown in Table 3, the MS-CNN model was very accurate in predicting microseismic events and
blasts, with an overall recognition accuracy of 99.6%. In terms of the classification results of different events,
those predicted as blast events were equal to the true category labels, with 100% precision. All microseismic
events were successfully recalled, but one blast event was misjudged as a microseismic event, resulting in a
99% precision for microseismic events.
The receiver operating characteristic (ROC) curve is a visual representation that effectively illustrates
the performance of a classification model. To generate the ROC curve, the model is first trained on a labeled
dataset, and then the model's output probabilities are calculated for each input sample. After the probabilities
are thresholded, binary predictions are created and compared to the true labels to calculate the true positive
rate (TPR) and the false positive rate (FPR) at various thresholds. Plotting the TPR and FPR for different
thresholds, as shown in Figure 4, allows us to construct the receiver operating characteristic (ROC) curve for
the MS-CNN model.
The area under the ROC curve (AUC) is a measure of the model's overall performance, with a higher
AUC indicating better performance. In the context of microseismic waveform recognition and classification,
the ROC curve can be used to evaluate the performance of MS-CNN in distinguishing between different types
of microseismic events or blasts. By analyzing the ROC curve, the optimal threshold can be selected to achieve
a desired balance between sensitivity and specificity, depending on the specific application requirements. As
shown in Figure 4, the MS-CNN model predicted microseismic events and blasts with high accuracy, reaching
a value of 0.999 for the AUC, which was close to a perfect classifier.
Table 3. Testing results of MS-CNN
Type Precision Recall F1 score Accuracy Testing time (s)
Microseismic events 0.990 1.000 0.990
0.996 0.310
Blasts 1.000 0.989 0.994
Figure 4. ROC curve of MS-CNN
4. DISCUSSION
Convolutional neural networks offer significant advantages over conventional techniques for the
recognition and classification of multi-channel microseismic waveforms. One of the main advantages is the
ability to extract more detailed information from the data using deep learning algorithms. This can result in
more accurate classification of microseismic events, including the ability to distinguish between different types
of events such as blasts, rock microfracture events, and noise. Another advantage is that using multi-channel
microseismic waveforms helps to capture a more complete picture of microseismic activity, which can improve
the accuracy of the classification process. In addition, as new data becomes available, CNN can update its
parameters and improve its accuracy without requiring manual intervention.
Int J Artif Intell ISSN: 2252-8938 
Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu)
1045
One potential limitation of using CNNs for microseismic waveform recognition and classification is
the need for a substantial volume of labeled training data. However, there are several techniques available to
address this challenge, including transfer learning [39], data augmentation, and semi-supervised learning.
Overall, the use of CNNs for microseismic waveform identification is a promising approach that has significant
potential to enhance our understanding of rockburst activity and improve our ability to detect and respond to
microseismic events. With the ongoing advancements in the field of deep learning, we can expect continuous
enhancements in both the accuracy and efficiency of these models.
Using the same dataset, the existing CNN [1] model, AlexNet, GoogLeNet, and ResNet50
classic image classification models were first trained for modeling. Figure 5 shows the comparison results of
the loss rate and accuracy of different CNN models during the training and validation process. Specifically,
Figures 5(a) and 5(b) show the variation of loss rate and accuracy for the five models on the training set
respectively. Figures 5(c) and 5(d) display the variation of loss rate and accuracy for the five models on the
validation set respectively. The training and validation loss and accuracy trends of the five models were
basically the same, and all of them were consistent with the variation of neural network training. Among
them, the accuracy of the validation ranged from 97.75% to 98.65%, and the loss was basically in the range of
0.041-0.074. The CNN [1] model fluctuated the most in training and validation, while ResNet50 was the most
stable and the best in validation. MS-CNN performed moderately in the training and validation process among
these five models.
(a) (b)
(c) (d)
Figure 5. Training and validation results of five CNN models (a) training loss, (b) training accuracy,
(c) validation loss, and (d) validation accuracy.
Based on the five selected evaluation metrics, the performance and effectiveness of the 5 CNN models
can be assessed by comparing their time requirements and recognition results. The best model will be
determined by the requirements of the particular application, such as accuracy, speed, and memory utilization.
 ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 1, March 2024: 1038-1049
1046
For real-time applications, for instance, a model with high accuracy and little inference time can be favored,
whereas a model requiring more memory and more training time might be chosen for applications where
precision is crucial. Ultimately, the architecture, hyperparameters, dataset, and hardware employed will all
depend on the identification results and processing time of CNN models, which must be assessed on a case-
by-case basis shown in Figure 6.
(a)
(b)
Figure 6. Identification results and time consumption for five CNN models (a) test results
and (b) time consumption
Compared to the existing CNN model [1], we reduced the number of convolutional layers, modified
the feature maps' parameters, and set the dropout value to 0.4. After multiple rounds of training and
hyperparameter tuning, we obtained the optimal MS-CNN model. Experimental results indicate that MS-CNN
is simpler and more efficient than the original CNN model [1] and transfer models such as AlexNet,
GoogLeNet, and ResNet50 because it has fewer convolutional layers while maintaining good recognition
accuracy on a limited dataset. From the comparative analysis of test results in Figure 6, it is clear that MS-
CNN outperforms other models in terms of classification accuracy (99.6%) and recognition efficiency (which
takes only 0.31 seconds to recognize 277 images in the test set) for the classification and recognition of
microseismic events.
5. CONCLUSION
The recognition and classification of multi-channel microseismic waveforms using CNNs is a
promising approach for accurately identifying microseismic signals. CNN-based models leverage the power of
deep learning algorithms to effectively extract and analyze features from complex microseismic data, enabling
accurate event classification based on waveform characteristics. The use of multi-channel data allows for a
more comprehensive analysis of microseismic events, leading to improved classification accuracy and reduced
false positives. Moreover, the adaptability of CNNs to new data makes them suitable for real-time applications,
such as rockburst monitoring and early warning systems. Developing CNN-based models for microseismic
waveform identification has significant potential to enhance our understanding of mining seismic activity and
improve our ability to detect and respond to microseismic events. Advancements in deep learning will further
increase the accuracy and efficiency of these models, making them increasingly valuable tools for rockburst
analysis and prediction. This paper proposed the MS-CNN model, which demonstrated outstanding
Int J Artif Intell ISSN: 2252-8938 
Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu)
1047
performance in recognizing and classifying multichannel microseismic signal waveforms with high speed and
accuracy. Compared to existing CNN models and classical image recognition and classification models, such
as AlexNet, GoogLeNet, and ResNet50, the MS-CNN model achieved the best classification results
with 99.6% accuracy and the shortest dataset identification time. Therefore, the MS-CNN method is
practical for engineering applications where automatic recognition and classification of microseismic events
and blasts are necessary.
ACKNOWLEDGEMENTS
The writers express their sincere appreciation for the reviewers' valuable time dedicated to reading
this paper. Furthermore, we would like to extend our gratitude to the International College of Digital Innovation
at Chiang Mai University for providing us with the opportunity to carry out this study.
REFERENCES
[1] L. Dong, Z. Tang, X. Li, Y. Chen, and J. Xue, “Discrimination of mining microseismic events and blasts using convolutional neural
networks and original waveform,” Journal of Central South University, vol. 27, no. 10, pp. 3078–3089, Oct. 2020, doi:
10.1007/s11771-020-4530-8.
[2] Q. Feng, L. Han, B. Pan, and B. Zhao, “Microseismic source location using deep reinforcement learning,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 60, pp. 1–9, 2022, doi: 10.1109/TGRS.2022.3182991.
[3] K. Ma, X. Sun, Z. Zhang, J. Hu, and Z. Wang, “Intelligent location of microseismic events based on a fully convolutional neural
network (FCNN),” Rock Mechanics and Rock Engineering, vol. 55, no. 8, pp. 4801–4817, Aug. 2022, doi: 10.1007/s00603-022-
02911-x.
[4] X. Yin, Q. Liu, X. Huang, and Y. Pan, “Real-time prediction of rockburst intensity using an integrated CNN-Adam-BO algorithm
based on microseismic data and its engineering application,” Tunnelling and Underground Space Technology, vol. 117, p. 104133,
Nov. 2021, doi: 10.1016/j.tust.2021.104133.
[5] L. Dong and Q. Luo, “Investigations and new insights on earthquake mechanics from fault slip experiments,” Earth-Science
Reviews, vol. 228, p. 104019, May 2022, doi: 10.1016/j.earscirev.2022.104019.
[6] Y. Pu, D. B. Apel, V. Liu, and H. Mitri, “Machine learning methods for rockburst prediction-state-of-the-art review,” International
Journal of Mining Science and Technology, vol. 29, no. 4, pp. 565–570, Jul. 2019, doi: 10.1016/j.ijmst.2019.06.009.
[7] M. Chakraborty, M. Das, and S. Aruchamy, “Micro-Seismic Event Detection using statistical feature extraction and machine
learning techniques,” in 2022 IEEE 7th International conference for Convergence in Technology (I2CT), IEEE, Apr. 2022, pp. 1–
5. doi: 10.1109/I2CT54291.2022.9824819.
[8] J. X. Wang, S. B. Tang, M. J. Heap, C. A. Tang, and L. X. Tang, “An auto-detection network to provide an automated real-time
early warning of rock engineering hazards using microseismic monitoring,” International Journal of Rock Mechanics and Mining
Sciences, vol. 140, p. 104685, Apr. 2021, doi: 10.1016/j.ijrmms.2021.104685.
[9] D. Wamriew, R. Pevzner, E. Maltsev, and D. Pissarenko, “Deep neural networks for detection and location of microseismic events
and velocity model inversion from microseismic data acquired by distributed acoustic sensing array,” Sensors, vol. 21, no. 19, p.
6627, Oct. 2021, doi: 10.3390/s21196627.
[10] J. Li et al., “Automatic recognition and classification of microseismic waveforms based on computer vision,” Tunnelling and
Underground Space Technology, vol. 121, p. 104327, Mar. 2022, doi: 10.1016/j.tust.2021.104327.
[11] J. He, L. Dou, S. Gong, J. Li, and Z. Ma, “Rock burst assessment and prediction by dynamic and static stress analysis based on
micro-seismic monitoring,” International Journal of Rock Mechanics and Mining Sciences, vol. 93, pp. 46–53, Mar. 2017, doi:
10.1016/j.ijrmms.2017.01.005.
[12] Y. Pu, D. B. Apel, and R. Hall, “Using machine learning approach for microseismic events recognition in underground excavations:
comparison of ten frequently-used models,” Engineering Geology, vol. 268, p. 105519, Apr. 2020, doi:
10.1016/j.enggeo.2020.105519.
[13] K. Thamprasert, A. Y. Dawod, and N. Chakpitak, “Simulated trial and error experiments on productivity,” IAES International
Journal of Artificial Intelligence (IJ-AI), vol. 11, no. 4, p. 1570, Dec. 2022, doi: 10.11591/ijai.v11.i4.pp1570-1578.
[14] H. Wu, B. Zhang, and N. Liu, “Self-adaptive denoising net: self-supervised learning for seismic migration artifacts and random
noise attenuation,” Journal of Petroleum Science and Engineering, vol. 214, p. 110431, Jul. 2022, doi:
10.1016/j.petrol.2022.110431.
[15] Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, “A survey of convolutional neural networks: analysis, applications, and prospects,”
IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 12, pp. 6999–7019, Dec. 2022, doi:
10.1109/TNNLS.2021.3084827.
[16] R. M. Abd-ElGhaffar, M. El-Zalabany, H. E.-D. Moustafa, and M. El-Seddek, “Classification of focal liver disease in egyptian
patients using ultrasound images and convolutional neural networks,” Indonesian Journal of Electrical Engineering and Computer
Science, vol. 27, no. 2, p. 793, Aug. 2022, doi: 10.11591/ijeecs.v27.i2.pp793-802.
[17] S. Tang, J. Wang, and C. Tang, “Identification of microseismic events in rock engineering by a convolutional neural network
combined with an attention mechanism,” Rock Mechanics and Rock Engineering, vol. 54, no. 1, pp. 47–69, Jan. 2021, doi:
10.1007/s00603-020-02259-0.
[18] D. Anikiev et al., “Machine learning in microseismic monitoring,” Earth-Science Reviews, vol. 239, p. 104371, Apr. 2023, doi:
10.1016/j.earscirev.2023.104371.
[19] K. Peng, Z. Tang, L. Dong, and D. Sun, “Machine learning based identification of microseismic signals using characteristic
parameters,” Sensors, vol. 21, no. 21, p. 6967, Oct. 2021, doi: 10.3390/s21216967.
[20] L.-J. Dong, J. Wesseloo, Y. Potvin, and X.-B. Li, “Discriminant models of blasts and seismic events in mine seismology,”
International Journal of Rock Mechanics and Mining Sciences, vol. 86, pp. 282–291, Jul. 2016, doi: 10.1016/j.ijrmms.2016.04.021.
[21] L. Dong, J. Wesseloo, Y. Potvin, and X. Li, “Discrimination of mine seismic events and blasts using the fisher classifier, naive
bayesian classifier and logistic regression,” Rock Mechanics and Rock Engineering, vol. 49, no. 1, pp. 183–211, Jan. 2016, doi:
10.1007/s00603-015-0733-y.
[22] C. Ma et al., “A novel microseismic classification model based on bimodal neurons in an artificial neural network,” Tunnelling and
 ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 1, March 2024: 1038-1049
1048
Underground Space Technology, vol. 131, p. 104791, Jan. 2023, doi: 10.1016/j.tust.2022.104791.
[23] A. H. Wilkins, A. Strange, Y. Duan, and X. Luo, “Identifying microseismic events in a mining scenario using a convolutional neural
network,” Computers & Geosciences, vol. 137, p. 104418, Apr. 2020, doi: 10.1016/j.cageo.2020.104418.
[24] L. Alzubaidi et al., “Review of deep learning: concepts, CNN architectures, challenges, applications, future directions,” Journal of
Big Data, vol. 8, no. 1, p. 53, Mar. 2021, doi: 10.1186/s40537-021-00444-8.
[25] S. M. Mousavi and G. C. Beroza, “Deep-learning seismology,” Science, vol. 377, no. 6607, Aug. 2022, doi:
10.1126/science.abm4470.
[26] Y. Pratama, E. Marbun, Y. Parapat, and A. Manullang, “Deep convolutional neural network for hand sign language recognition
using model E,” Bulletin of Electrical Engineering and Informatics, vol. 9, no. 5, pp. 1873–1881, Oct. 2020, doi:
10.11591/eei.v9i5.2027.
[27] B. I. Lin, X. I. E. Wei, Z. H. A. O. Junjie, and Z. H. A. O. Hui, “Automatic classification of multi-channel microseismic waveform
based on DCNN-SPP,” Journal of Applied Geophysics, vol. 159, pp. 446–452, Dec. 2018, doi: 10.1016/j.jappgeo.2018.09.022.
[28] B. Lin, X. Wei, and Z. Junjie, “Automatic recognition and classification of multi-channel microseismic waveform based on DCNN
and SVM,” Computers & Geosciences, vol. 123, pp. 111–120, Feb. 2019, doi: 10.1016/j.cageo.2018.10.008.
[29] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” Sep. 2014, [Online].
Available: http://arxiv.org/abs/1409.1556
[30] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” Dec. 2015, [Online]. Available:
http://arxiv.org/abs/1512.03385
[31] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” Communications
of the ACM, vol. 60, no. 6, pp. 84–90, May 2017, doi: 10.1145/3065386.
[32] J. Wang and S. Tang, “Novel transfer learning framework for microseismic event recognition between multiple monitoring
projects,” Rock Mechanics and Rock Engineering, vol. 55, no. 6, pp. 3563–3582, Jun. 2022, doi: 10.1007/s00603-022-02790-2.
[33] A. S. Hatem, M. S. Altememe, and M. A. Fadhel, “Identifying corn leaves diseases by extensive use of transfer learning: a
comparative study,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 29, no. 2, p. 1030, Feb. 2023, doi:
10.11591/ijeecs.v29.i2.pp1030-1038.
[34] W. Setiawan, M. I. Utoyo, and R. Rulaningtyas, “Reconfiguration layers of convolutional neural network for fundus patches
classification,” Bulletin of Electrical Engineering and Informatics, vol. 10, no. 1, pp. 383–389, Feb. 2021, doi:
10.11591/eei.v10i1.1974.
[35] D.-X. Zhou, “Theory of deep convolutional neural networks: Downsampling,” Neural Networks, vol. 124, pp. 319–327, Apr. 2020,
doi: 10.1016/j.neunet.2020.01.018.
[36] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning (adaptive computation and machine learning series). Cambridge
Massachusetts. The MIT Press, 2017.
[37] R. Ali, J. H. Chuah, M. S. A. Talip, N. Mokhtar, and M. A. Shoaib, “Structural crack detection using deep convolutional neural
networks,” Automation in Construction, vol. 133, p. 103989, Jan. 2022, doi: 10.1016/j.autcon.2021.103989.
[38] A. Bahri, S. G. Majelan, S. Mohammadi, M. Noori, and K. Mohammadi, “Remote sensing image classification via improved cross-
entropy loss and transfer learning strategy based on deep convolutional neural networks,” IEEE Geoscience and Remote Sensing
Letters, vol. 17, no. 6, pp. 1087–1091, Jun. 2020, doi: 10.1109/LGRS.2019.2937872.
[39] W. Setiawan, M. I. Utoyo, and R. Rulaningtyas, “Transfer learning with multiple pre-trained network for fundus classification,”
TELKOMNIKA (Telecommunication Computing Electronics and Control), vol. 18, no. 3, p. 1382, Jun. 2020, doi:
10.12928/telkomnika.v18i3.14868.
BIOGRAPHIES OF AUTHORS
Hongmei Shu received a B.E. degree (2016) and the M.E. degree (2019) from
School of Computer Science and Technology at Sichuan Normal University. From July 2019
to February 2022, she worked as a lab technician in the School of Information and
Engineering at Sichuan Tourism University. She is currently a Ph.D. candidate in Digital
Innovation and Financial Technology at Chiang Mai University. From March 2022 to the
present, she is also a research assistant in the School of Resource Security and Engineering
at Central South University. Her research interests include image/signal processing, machine
learning, data mining, and artificial intelligence. She can be contacted at email:
hongmei_shu@cmu.ac.th.
Dr. Ahmad Yahya Dawod is currently a lecturer at International College of
Digital Innovation at Chiang Mai University, Thailand. He received his Ph.D. degree in
Machine Learning and Artificial Intelligence from the National University of Malaysia in
2018 with the topic “Hand gesture recognition based on isolated and continuous sign
language”. He also graduated his master’s degree in computing and informatics from
Multimedia University of Malaysia and had his bachelor’s degree in computer science from
The University of Mustansirya of Iraq. His research includes machine learning, pattern
recognition, computer vision, robotics, and artificial intelligence. He had published 20
articles up to date with more than a hundred citations. He can be contacted at email:
ahmadyahyadawod.a@cmu.ac.th.
Int J Artif Intell ISSN: 2252-8938 
Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu)
1049
Dr. Worawit Tepsan is currently a lecturer at International College of Digital
Innovation at Chiang Mai University, Thailand, since December 2020. He received his Ph.D.
degree in Mathematics from the University of Houston, US in 2020. His research includes
mathematical operator theory, functional analysis, applied mathematics, machine learning,
and data analysis. He can be contacted at email: worawit.tepsan@cmu.ac.th.
Dr. Lei Mu has been a full lecturer of Chengdu University, Mainland China
since 2003. He received his BE and ME from the University of Electronic Science and
Technology of China. In 2011, he obtained a full scholarship from Erasmus Mundus Action
2-a Sustainable e-Tourism Project and got his PhD from the University of Grenoble Alps,
France. He was appointed Deputy Dean of the School of Foreign Languages at Chengdu
University in December 2014. He is responsible for research, faculty, training and
international cooperation. Since December 2020 till now, he works as Deputy Director of the
Office of International Collaboration and Exchange, Chengdu University. His main research
areas include "technology application and innovation in international tourism" and
"computer-assisted translation". He can be contacted at email: mulei@cdu.edu.cn.
Zheng Tang holds a Bachelor of Engineering (B.Eng.) in Mining Engineering,
Master of Science (M.Sc.) in Mining Engineering, besides several professional certificates
and skills. He is currently studying for a doctorate in the School of Resource and Safety
Engineering at the Central South University. His research areas of interest include disaster
warning, artificial intelligent, and signal identification. He can be contacted at email:
tzheng@csu.edu.cn.

More Related Content

PDF
FAULT DETECTION IN INDUCTION MOTORPD.pdf
PDF
A Time Series ANN Approach for Weather Forecasting
PDF
Applying convolutional neural networks for limited-memory application
PDF
sensors-23-04512-v3.pdf
PDF
TBerger_FinalReport
PDF
Predicting and detecting fires on multispectral images using machine learning...
PDF
Ensemble Learning Approach for Digital Communication Modulation’s Classification
PDF
Attention Mechanism for Attacks and Intrusion Detection
FAULT DETECTION IN INDUCTION MOTORPD.pdf
A Time Series ANN Approach for Weather Forecasting
Applying convolutional neural networks for limited-memory application
sensors-23-04512-v3.pdf
TBerger_FinalReport
Predicting and detecting fires on multispectral images using machine learning...
Ensemble Learning Approach for Digital Communication Modulation’s Classification
Attention Mechanism for Attacks and Intrusion Detection

Similar to Multi-channel microseismic signals classification with convolutional neural networks (20)

PDF
Ensemble Learning Approach for Digital Communication Modulation’s Classification
PDF
ENSEMBLE LEARNING APPROACH FOR DIGITAL COMMUNICATION MODULATION’S CLASSIFICATION
PDF
IRJET - Detection of Malaria using Image Cells
PDF
Ubiquitous-cloud-inspired deterministic and stochastic service provider model...
PDF
Automatic fabric defect detection employing deep learning
PDF
1-s2.0-S1474034622002737-main.pdf
PDF
Detecting anomalies in security cameras with 3D-convolutional neural network ...
PDF
Redefining brain tumor segmentation: a cutting-edge convolutional neural netw...
PDF
Convolutional neural network-based real-time drowsy driver detection for acci...
PDF
ML_project_proposal for my university.pdf
PDF
G43043540
PDF
DiaMe: IoMT deep predictive model based on threshold aware region growing tec...
PDF
Accuracy study of image classification for reverse vending machine waste segr...
PDF
Enhancing machine failure prediction with a hybrid model approach
PDF
Preliminary Lung Cancer Detection using Deep Neural Networks
PDF
40120140507006
PDF
40120140507006
PDF
A deep convolutional structure-based approach for accurate recognition of ski...
PDF
A deep learning approach for brain tumor detection using magnetic resonance ...
PDF
Chaotic systems with pseudorandom number generate to protect the transmitted...
Ensemble Learning Approach for Digital Communication Modulation’s Classification
ENSEMBLE LEARNING APPROACH FOR DIGITAL COMMUNICATION MODULATION’S CLASSIFICATION
IRJET - Detection of Malaria using Image Cells
Ubiquitous-cloud-inspired deterministic and stochastic service provider model...
Automatic fabric defect detection employing deep learning
1-s2.0-S1474034622002737-main.pdf
Detecting anomalies in security cameras with 3D-convolutional neural network ...
Redefining brain tumor segmentation: a cutting-edge convolutional neural netw...
Convolutional neural network-based real-time drowsy driver detection for acci...
ML_project_proposal for my university.pdf
G43043540
DiaMe: IoMT deep predictive model based on threshold aware region growing tec...
Accuracy study of image classification for reverse vending machine waste segr...
Enhancing machine failure prediction with a hybrid model approach
Preliminary Lung Cancer Detection using Deep Neural Networks
40120140507006
40120140507006
A deep convolutional structure-based approach for accurate recognition of ski...
A deep learning approach for brain tumor detection using magnetic resonance ...
Chaotic systems with pseudorandom number generate to protect the transmitted...
Ad

More from IAESIJAI (20)

PDF
A comparative study of natural language inference in Swahili using monolingua...
PDF
Abstractive summarization using multilingual text-to-text transfer transforme...
PDF
Enhancing emotion recognition model for a student engagement use case through...
PDF
Automatic detection of dress-code surveillance in a university using YOLO alg...
PDF
Hindi spoken digit analysis for native and non-native speakers
PDF
Two-dimensional Klein-Gordon and Sine-Gordon numerical solutions based on dee...
PDF
Improved convolutional neural networks for aircraft type classification in re...
PDF
Primary phase Alzheimer's disease detection using ensemble learning model
PDF
Deep learning-based techniques for video enhancement, compression and restora...
PDF
Hybrid model detection and classification of lung cancer
PDF
Adaptive kernel integration in visual geometry group 16 for enhanced classifi...
PDF
Video forgery: An extensive analysis of inter-and intra-frame manipulation al...
PDF
Enhancing fall detection and classification using Jarratt‐butterfly optimizat...
PDF
Deep ensemble learning with uncertainty aware prediction ranking for cervical...
PDF
Event detection in soccer matches through audio classification using transfer...
PDF
Detecting road damage utilizing retinaNet and mobileNet models on edge devices
PDF
Optimizing deep learning models from multi-objective perspective via Bayesian...
PDF
Squeeze-excitation half U-Net and synthetic minority oversampling technique o...
PDF
A novel scalable deep ensemble learning framework for big data classification...
PDF
Exploring DenseNet architectures with particle swarm optimization: efficient ...
A comparative study of natural language inference in Swahili using monolingua...
Abstractive summarization using multilingual text-to-text transfer transforme...
Enhancing emotion recognition model for a student engagement use case through...
Automatic detection of dress-code surveillance in a university using YOLO alg...
Hindi spoken digit analysis for native and non-native speakers
Two-dimensional Klein-Gordon and Sine-Gordon numerical solutions based on dee...
Improved convolutional neural networks for aircraft type classification in re...
Primary phase Alzheimer's disease detection using ensemble learning model
Deep learning-based techniques for video enhancement, compression and restora...
Hybrid model detection and classification of lung cancer
Adaptive kernel integration in visual geometry group 16 for enhanced classifi...
Video forgery: An extensive analysis of inter-and intra-frame manipulation al...
Enhancing fall detection and classification using Jarratt‐butterfly optimizat...
Deep ensemble learning with uncertainty aware prediction ranking for cervical...
Event detection in soccer matches through audio classification using transfer...
Detecting road damage utilizing retinaNet and mobileNet models on edge devices
Optimizing deep learning models from multi-objective perspective via Bayesian...
Squeeze-excitation half U-Net and synthetic minority oversampling technique o...
A novel scalable deep ensemble learning framework for big data classification...
Exploring DenseNet architectures with particle swarm optimization: efficient ...
Ad

Recently uploaded (20)

PPTX
Spectroscopy.pptx food analysis technology
PDF
Architecting across the Boundaries of two Complex Domains - Healthcare & Tech...
PPTX
OMC Textile Division Presentation 2021.pptx
PDF
Network Security Unit 5.pdf for BCA BBA.
PDF
Per capita expenditure prediction using model stacking based on satellite ima...
PPTX
SOPHOS-XG Firewall Administrator PPT.pptx
PDF
Reach Out and Touch Someone: Haptics and Empathic Computing
PPTX
Machine Learning_overview_presentation.pptx
PDF
Mobile App Security Testing_ A Comprehensive Guide.pdf
PDF
Assigned Numbers - 2025 - Bluetooth® Document
PPTX
A Presentation on Artificial Intelligence
PDF
Heart disease approach using modified random forest and particle swarm optimi...
PPTX
Group 1 Presentation -Planning and Decision Making .pptx
PPTX
Tartificialntelligence_presentation.pptx
PDF
Diabetes mellitus diagnosis method based random forest with bat algorithm
PDF
Advanced methodologies resolving dimensionality complications for autism neur...
PPTX
KOM of Painting work and Equipment Insulation REV00 update 25-dec.pptx
PPTX
TechTalks-8-2019-Service-Management-ITIL-Refresh-ITIL-4-Framework-Supports-Ou...
PDF
Profit Center Accounting in SAP S/4HANA, S4F28 Col11
PPTX
cloud_computing_Infrastucture_as_cloud_p
Spectroscopy.pptx food analysis technology
Architecting across the Boundaries of two Complex Domains - Healthcare & Tech...
OMC Textile Division Presentation 2021.pptx
Network Security Unit 5.pdf for BCA BBA.
Per capita expenditure prediction using model stacking based on satellite ima...
SOPHOS-XG Firewall Administrator PPT.pptx
Reach Out and Touch Someone: Haptics and Empathic Computing
Machine Learning_overview_presentation.pptx
Mobile App Security Testing_ A Comprehensive Guide.pdf
Assigned Numbers - 2025 - Bluetooth® Document
A Presentation on Artificial Intelligence
Heart disease approach using modified random forest and particle swarm optimi...
Group 1 Presentation -Planning and Decision Making .pptx
Tartificialntelligence_presentation.pptx
Diabetes mellitus diagnosis method based random forest with bat algorithm
Advanced methodologies resolving dimensionality complications for autism neur...
KOM of Painting work and Equipment Insulation REV00 update 25-dec.pptx
TechTalks-8-2019-Service-Management-ITIL-Refresh-ITIL-4-Framework-Supports-Ou...
Profit Center Accounting in SAP S/4HANA, S4F28 Col11
cloud_computing_Infrastucture_as_cloud_p

Multi-channel microseismic signals classification with convolutional neural networks

  • 1. IAES International Journal of Artificial Intelligence (IJ-AI) Vol. 13, No. 1, March 2024, pp. 1038~1049 ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i1.pp1038-1049  1038 Journal homepage: http://ijai.iaescore.com Multi-channel microseismic signals classification with convolutional neural networks Hongmei Shu1,2 , Ahmad Yahya Dawod1 , Worawit Tepsan1 , Lei Mu1,3 , Zheng Tang2 1 International College of Digital Innovation, Chiang Mai University, Chiang Mai, Thailand 2 School of Resources and Safety Engineering, Central South University, Changsha, China 3 Office of International Collaboration and Exchange, Chengdu University, Chengdu, China Article Info ABSTRACT Article history: Received May 11, 2023 Revised Aug 16, 2023 Accepted Sep 8, 2023 Identifying and classifying microseismic signals is essential to warn of mines’ dangers. Deep learning has replaced traditional methods, but labor-intensive manual identification and varying deep learning outcomes pose challenges. This paper proposes a transfer learning-based convolutional neural network (CNN) method called microseismic signals-convolutional neural network (MS-CNN) to automatically recognize and classify microseismic events and blasts. The model was instructed on a limited sample of data to obtain an optimal weight model for microseismic waveform recognition and classification. A comparative analysis was performed with an existing CNN model and classical image classification models such as AlexNet, GoogLeNet, and ResNet50. The outcomes demonstrate that the MS-CNN model achieved the best recognition and classification effect (99.6% accuracy) in the shortest time (0.31 s to identify 277 images in the test set). Thus, the MS-CNN model can efficiently recognize and classify microseismic events and blasts in practical engineering applications, improving the recognition timeliness of microseismic signals and further enhancing the accuracy of event classification. Keywords: Classification Convolutional neural network Microseismic events Multi-channel waveform Recognition Transfer learning This is an open access article under the CC BY-SA license. Corresponding Author: Ahmad Yahya Dawod International College of Digital Innovation, Chiang Mai University 239 Huay Kaew Rd, Suthep, Mueang Chiang Mai District, Chiang Mai 50200, Thailand Email: ahmadyahyadawod.a@cmu.ac.th 1. INTRODUCTION Microseismic monitoring technology (MMT) has found extensive applications in underground engineering for disasters and safety monitoring [1]. Specifically, it has been utilized for location monitoring [2], [3], as well as forecasting and providing early warning systems for rock bursts [4], and mine earthquake disasters during mining operations [5], [6]. The basic principle involves identifying microseismic events by analyzing prominent features within the monitoring data [7]. Subsequently, the relevant parameters of these events are analyzed to facilitate informed decision-making [8]. Due to the complex geological environment of mines [9], many interference signals frequently mix with the recorded microseismic signals, such as blasting, rock drilling, fan vibration, and other noises generated during engineering operations [10]. Therefore, the basis of MMT is to identify microseismic events quickly and accurately, which determines the timeliness and effectiveness of mine safety early warning [11]. Earlier studies have investigated several strategies, such as conventional signal processing methods and machine learning algorithms [12], [13], to identify and classify microseismic signals [14]. However, these techniques have limitations regarding their efficiency and accuracy. Recently, convolutional neural networks (CNNs) [15], [16] have shown great promise in identifying and categorizing images. Based on this
  • 2. Int J Artif Intell ISSN: 2252-8938  Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu) 1039 success, some researchers employed CNNs to successfully recognize and categorize microseismic signals [17]. However, these studies have primarily focused on single-channel signals, and further research is necessary to explore the application of CNNs to multichannel signals. The higher dimensionality of multichannel signals presents unique challenges, and the effectiveness of CNNs for these signals is an active research area. Traditional methods for identifying microseismic events and blasts primarily rely on manual or engineering experience, which often results in significant and parameter analysis errors as well as time delays. Extensive research has been conducted by experts and scholars in the field of recognition and classification, as evidenced by previous studies [18], [19]. Traditional methods typically rely on statistical analysis of waveform characteristics to classify microseismic events and blasts. However, these methods often require manual experience for selecting characteristic parameters, resulting in limited generalizability and robustness due to their specific applicability to certain cases or mines. For example, researchers conducted an analysis on the characteristic patterns of source parameters associated with microseismic events and blasts [20], [21]. They developed a statistical model for the automatic identification of such events, taking into account the probability density distribution of each parameter and its impact on identification accuracy. However, the recognition effect of this method on different mine data samples is quite different, and the robustness is not high. With the explosive growth of data, the use of machine learning for image classification has become popular [22], [23]. Deep learning [24]–[26] is a widely used method at present, which automatically extracts features through data-driven, and obtains the specific feature representation of the dataset based on the learning and training of many samples. It is more efficient and accurate in the expression of data sets without manual intervention. Lin et al. and Lin et al. [27], [28] have proposed two methods, deep convolutional neural network and spatial pyramidal pool (DCNN-SPP) and DCNN and support vector machine (DCNN-SVM), to automatically identify and classify multi-channel microseismic waveforms, with accuracy improved from 91.13% to 98.18%. Li et al. [10] introduced a deep learning method for microseismic waveform classification based on computer vision, which analyzed both the waveform of the microseismic signal and the corresponding spectrogram of the waveform. Four deep learning models were used for experiments, including VGG-16 [29], ResNet18 [30], AlexNet [31], and their ensemble models. The results showed that these models perform well in recognizing the features of waveform graphs and can accurately predict data categories, among which the ensemble model had the highest recognition accuracy of 98%. The end-to-end automatic classification of the original waveform was realized, and it proved that the identification effect of the original waveform and the spectrogram was basically the same. This study proposes a CNN-based model named microseismic signals convolutional neural network (MS-CNN) to enhance the precision and effectiveness of microseismic event recognition and classification. This study's goal is to assess how well CNNs and related transfer learning models work. The researchers [32], [33] in accurately classifying microseismic events and blasts in mines. We developed and trained multiple deep learning models, comprising three conventional models (AlexNet, GoogLeNet, and ResNet50) and one CNN model to accomplish this. We assessed the classification performance of these models using five metrics based on an identical dataset. After analyzing the experimental results, we pinpointed the most effective model for this purpose, hoping that this research will assist users in making informed decisions. 2. METHOD The application of CNNs for microseismic waveform recognition and classification is guided by several fundamental principles. These include the capacity of deep learning algorithms to extract intricate features from multi-channel data, the utilization of supervised learning to train the model, and the iterative optimization of the model's parameters to minimize classification error. The CNN model is trained using labeled data, where the model's parameters and weights are adjusted to minimize classification error. This optimization process often involves iterative algorithms like stochastic gradient descent, which update the model's parameters based on the discrepancy between expected and actual labels. Convolutional layers play a crucial role in this process by applying filters to the input data to extract local features [34]. Additionally, pooling layers are utilized to down-sample the feature maps, simplifying the computational complexity of the model. Once trained, the CNN can classify new and unseen data by inputting it into the network and leveraging the learned parameters to predict the type of microseismic event present in each segment. 2.1. Microseismic signals convolutional neural network Figure 1 depicts the MS-CNN model's hierarchical structure, which comprises an input layer, two convolutional layers, a pooling layer, two fully connected layers, and an output layer. It abstracts the image into data layer by layer to provide the necessary feature representation, and then maps the features to the task objective. Humans do not need to intervene in the feature selection process during CNN training. The neural network automatically learns image local features through convolution and backpropagates the error through the loss function to obtain the optimal convolution coefficients (weights and biases). The identification
  • 3.  ISSN: 2252-8938 Int J Artif Intell, Vol. 13, No. 1, March 2024: 1038-1049 1040 and classification principle of the MS-CNN model was mainly composed of two parts, one part featured learning (convolution, activation function, and pooling layers), and the other part was classification (full connection layer). Figure 1. Structure of MS-CNN, a multichannel microseismic waveform recognition and classification model 2.1.1. Feature extraction The convolutional and pooling layers are crucial to CNN's capacity to automatically extract image features. The weight and bias parameters in the convolution operation are effectively solved during the CNN learning process utilizing the dataset and loss function [35]. The value of the feature map is determined using (1), where 𝑥 represents the input image, f denotes the convolution kernel, and 𝑚, 𝑛 represent the row and column indices, respectively, of the calculation result [36]. 𝐷[𝑚, 𝑛] = (𝑥 ∗ 𝑓)[𝑚, 𝑛] = ∑ ∑ 𝑓[𝑖, 𝑗]𝑥[𝑚 − 𝑖, 𝑛 − 𝑗] 𝑗 𝑖 (1) The pooling layer, also known as the downsampling layer, reduces the dimension of the feature map by pooling each feature map, thus reducing the number of parameters in CNN, and avoiding overfitting [37]. The window size of the pooling operation can be specified as any size. There are two main pooling operations: max pooling and average pooling. The MS-CNN model used max pooling. The maximum pooling function uses the maximum value in the small block as the feature output of the block. When the image undergoes small changes such as translation, scaling, and rotation, it is still possible to obtain the maximum value at the same position, which is the same as the response before the change, thus realizing the spatial invariance feature. 2.1.2. Classification The classification of the MS-CNN was achieved by fully connected layers. Since MS-CNN was a supervised learning, the model was trained based on the labeled training samples to obtain the weights of the fully connected layer. When using the model for outcome recognition, the weights obtained from the training of the model and the results calculated from the deep network after the previous convolution, activation function, and pooling, were weighted and summed to obtain the predicted value of each outcome. Then the maximum value was taken as the result of the recognition. As shown in Figure 1, the probability of a microseismic event was finally calculated to be 1.00, and the probability of a blasting event was 0.00. Therefore, the event corresponding to this figure was finally determined to be a microseismic event. The output layer of the model computes the error using the cross-entropy loss function, which assesses the similarity between the predicted output and the desired output. In the back-propagation process, this information is then used to optimize or update the parameters of each layer. In (2) [38] represents the mathematical expression for the cross-entropy loss function. Where N represents the sample size, K denotes the number of categories, and tij =1 if sample i belongs to the jth category, otherwise tij=0. The variable yij represents the probability that sample i belongs to the jth category. 𝐿𝑜𝑠𝑠 = − 1 𝑁 ∑ ∑ 𝑡𝑖𝑗𝑙𝑛𝑦𝑖𝑗 𝐾 𝑗=1 𝑁 𝑖=1 (2)
  • 4. Int J Artif Intell ISSN: 2252-8938  Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu) 1041 The main advantages of MS-CNN are its ability to handle multichannel data. In microseismic applications, multiple sensors are often used to record microseismic waves, and MS-CNN can use all the available channels to improve the accuracy of the recognition and classification task. MS-CNN is trained using a dataset of microseismic event waveforms with labeled classes. The model learns to recognize patterns in the data and associate them with specific classes. Once trained, MS-CNN can be used to classify new microseismic waveforms based on their features and similarities to the learned patterns. 2.2. Model construction Figure 1 shows the parameter settings of each layer for the MS-CNN model, and the number of learning parameters is shown in Table 1. Before the image was fed into the model, the color image was preprocessed by transforming its size from 432×288 pixels to 100×100 pixels. After the image (e.g., X0) is inputted into the model, first, the Input layer would be normalized to X0 (size width*height*channels=100×100×3). The C1 layer used a (3×3) convolution kernel to perform 32 convolutions on the X0 to form a (32×98×98) tensor. Then the C2 layer continued to use the (3×3) convolution kernel to perform 32 convolutions on the output data of the C1 layer, and finally obtained a (32×48×48) structure tensor. In the C1 and C2 layers, the stride is (1×1) and (2×2), respectively. To facilitate feature extraction and reduce the amount of computation, a pooling operation was performed on the output data after the C2 layer by pooling a maximum of (3×3) P1 to form a pooling tensor of (32×46×46). After pooling, a fully connected layer was added to flatten the P1 data into a one-dimensional vector, and the number of output neurons in the Fc1 was set to 256. Finally, the output size of the Fc2 is 2, and the softmax function was added after Fc2 to calculate the probability value of each waveform belonging to each category. After obtaining the probability values of the two categories, the category corresponding to the maximum probability value was selected as the output value in the output layer. The activation functions of MS-CNN all used ReLU (rectified linear unit), which existed between C1 and C2, and between C2 and P1, respectively. Additionally, to hasten CNN's training and lessen the sensitivity of the initialization of the network, we added a batch normalization layer between the C1 and the ReLU layer. Mini-batch processing of all observations for each channel was implemented at this layer. And adopted the dropout method to alleviate overfitting. Table 1. Learning params numbers for the layers of the MS-CNN model Layer type Learnable Params Convolution layer C1 Weights (3×3×3×32); Bias (1×1×32) 896 Batch normalization layer Offset (1×1×32); Scale (1×1×32) 64 Convolution layer C2 Weights (3×3×32×32); Bias (1×1×32) 9248 Max pooling layer P1  0 Fully Connected layer Fc1 Weights (256×80000); Bias (256×1) 20480256 Fully Connected layer Fc2 Weights (2×256); Bias (2×1) 514 3. RESULTS AND ANALYSIS To comprehensively evaluate the performance of CNN models, we conducted experimental tests using existing field data. The training process and test results of the MS-CNN model were longitudinally analyzed and compared to CNN [1], AlexNet, GoogLeNet, and ResNet50 models. Our research findings demonstrate that the MS-CNN model surpassed other models in terms of accuracy and computational efficiency. This was observed by achieving both high classification accuracy and low training time. These results suggest that the MS-CNN model has significant potential for use in real-world applications. 3.1. Data To proper data preparation is essential for the successful recognition and classification of multi- channel microseismic waveforms using CNN-based models. Careful attention should be paid to data collection, preprocessing, labeling, and splitting to ensure that the CNN is trained and tested on high-quality data. Real-time data from a mine microseismic monitoring system in China's Shanxi province was used in this study. A 26-channel microseismic monitoring system was installed by the engineers based on the mine's geological setting and mining strategy. Due to the large volume of data and the similarity among each type of event, data from March 29 to April 19, 2022, were selected for this study. This was a total of 22 days of data and a database size of 6.85 GB, chosen for computational performance reasons. Firstly, we filtered all event data from the database with at least 6 triggered sensors. Then, we used the Python plotting library Matplotlib to draw a 2D waveform graph of the event based on the sample time magnitude data, which was saved as an event waveform with six subplots as input. Next, we manually screened and marked the noise events, such as rock drilling and mechanical vibration, to filter them out. Finally, we obtained the multi-channel waveform of two types of typical blasts and microseismic events. Examples of
  • 5.  ISSN: 2252-8938 Int J Artif Intell, Vol. 13, No. 1, March 2024: 1038-1049 1042 waveforms for a microseismic event are shown in Figure 2(a), and those for a blast event can be seen in Figure 2(b). The vertical axis shows the signal amplitude, and the horizontal axis depicts a time in microseconds. The experimental sample contained a total of 1386 images, including 512 waveform images of microseismic events and 874 waveform images of blasts. The classification performance of various models using the same dataset will be compared and examined, control variables were employed to ensure fair training and testing. The distribution of sample sizes for each event type in the dataset is presented in Table 2. The dataset was split into two chronological sections: the first 80% were used as the training set, while the final 20% were utilized as the test set. During the training phase, the model was trained using waveform features. To mitigate the risk of overfitting, a validation set comprising 20% (222 images) of the training data was randomly selected. This validation set was used for fine- tuning hyperparameters and obtaining the optimal model. Finally, The model's capacity for prediction and generalization was evaluated using the test set. (a) (b) Figure 2. Example waveforms of microseismic event and blasting (a) microseismic event waveform and (b) blasting waveform Table 2. Data sets and their divisions Type All Training set (80%) Testing set (20%) Microseismic 512 410 102 Blast 874 699 175 Sum 1386 1109 277
  • 6. Int J Artif Intell ISSN: 2252-8938  Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu) 1043 3.2. Experiment For the hyperparameter settings of CNN model training, different models had different initial learning rates, and the best parameter values need to be selected after several network training sessions. The following parameters were chosen for the MS-CNN model: i) pick stochastic gradient descent with momentum (SGDM) as your optimization function; ii) define the first learning rate as 0.001; iii) choose six epochs as the maximum number; iv) decide on 16 as the mini-batch size; v) select a 10-period validation frequency. Next, we examine the manner in which the MS-CNN model performed in terms of the training results and the test outcomes, respectively. 3.2.1. Training results and analysis The loss and accuracy indicators are essential in assessing the progress of the learning process. During training, the loss typically decreases over time as the model learns to better fit the training data. To monitor the accuracy and loss changes during training, these values are often computed at regular intervals, such as after each epoch or after a certain number of training batches. The accuracy and loss curves can then be plotted over time to visualize how they change during training. In this study, a five-fold cross-validation was utilized during the training process, and the changes in loss and accuracy of the MS-CNN model during training and validation were depicted in Figure 3. The results show that accuracy increased continuously with training time and eventually stabilized at 1.0, while the loss gradually decreased and converged to approximately 0.0. The loss and accuracy values of the validation set were consistent with those of the training set. The accuracy of the validation set eventually stabilized at 98.6%, with a loss value of 0.05. These trends in loss and accuracy are in accordance with the changes expected during neural network training. The accuracy and loss metrics are frequently employed to monitor the model's performance during the MS-CNN training process and decide when to stop training. Figure 3. Accuracy and loss changes in MS-CNN training process The accuracy of MS-CNN refers to the percentage of correctly classified samples in the training or validation set. As the model learns from the training data, the accuracy typically increases, although it may eventually plateau or even decrease if the model overfits the training data. The loss function employed in MS- CNN quantifies the disparity between the predicted labels and the actual labels for each input sample. During the training process, the objective is to minimize this loss function, thereby enhancing the model's capacity to generate precise predictions. 3.2.2. Test results and analysis After the MS-CNN model training was completed, a new test dataset was used for prediction and classification, thus enabling analysis of the model performance. The identification results and time consumption of a CNN model can vary depending on the specific architecture, dataset, and hardware used to run the model. However, in general, the performance and time consumption of CNN models can be evaluated using the following metrics: The accuracy calculates the proportion of test set samples that were properly categorized. Better performance is indicated by higher accuracy. Precision counts the proportion of real positives among all correctly predicted positive outcomes. Less false positives are indicated by increased precision. The F1 score serves as a balanced metric, considering both precision and recall, and is calculated as
  • 7.  ISSN: 2252-8938 Int J Artif Intell, Vol. 13, No. 1, March 2024: 1038-1049 1044 the harmonic mean of these two measures. To assess the classification prediction performance of the model, we employed five evaluation metrics: precision, recall, F1 score, accuracy, and testing time. The evaluation measures for single-category classification include precision, recall, and F1 score. The overall performance of the model is assessed using the accuracy metric and testing time. As shown in Table 3, the MS-CNN model was very accurate in predicting microseismic events and blasts, with an overall recognition accuracy of 99.6%. In terms of the classification results of different events, those predicted as blast events were equal to the true category labels, with 100% precision. All microseismic events were successfully recalled, but one blast event was misjudged as a microseismic event, resulting in a 99% precision for microseismic events. The receiver operating characteristic (ROC) curve is a visual representation that effectively illustrates the performance of a classification model. To generate the ROC curve, the model is first trained on a labeled dataset, and then the model's output probabilities are calculated for each input sample. After the probabilities are thresholded, binary predictions are created and compared to the true labels to calculate the true positive rate (TPR) and the false positive rate (FPR) at various thresholds. Plotting the TPR and FPR for different thresholds, as shown in Figure 4, allows us to construct the receiver operating characteristic (ROC) curve for the MS-CNN model. The area under the ROC curve (AUC) is a measure of the model's overall performance, with a higher AUC indicating better performance. In the context of microseismic waveform recognition and classification, the ROC curve can be used to evaluate the performance of MS-CNN in distinguishing between different types of microseismic events or blasts. By analyzing the ROC curve, the optimal threshold can be selected to achieve a desired balance between sensitivity and specificity, depending on the specific application requirements. As shown in Figure 4, the MS-CNN model predicted microseismic events and blasts with high accuracy, reaching a value of 0.999 for the AUC, which was close to a perfect classifier. Table 3. Testing results of MS-CNN Type Precision Recall F1 score Accuracy Testing time (s) Microseismic events 0.990 1.000 0.990 0.996 0.310 Blasts 1.000 0.989 0.994 Figure 4. ROC curve of MS-CNN 4. DISCUSSION Convolutional neural networks offer significant advantages over conventional techniques for the recognition and classification of multi-channel microseismic waveforms. One of the main advantages is the ability to extract more detailed information from the data using deep learning algorithms. This can result in more accurate classification of microseismic events, including the ability to distinguish between different types of events such as blasts, rock microfracture events, and noise. Another advantage is that using multi-channel microseismic waveforms helps to capture a more complete picture of microseismic activity, which can improve the accuracy of the classification process. In addition, as new data becomes available, CNN can update its parameters and improve its accuracy without requiring manual intervention.
  • 8. Int J Artif Intell ISSN: 2252-8938  Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu) 1045 One potential limitation of using CNNs for microseismic waveform recognition and classification is the need for a substantial volume of labeled training data. However, there are several techniques available to address this challenge, including transfer learning [39], data augmentation, and semi-supervised learning. Overall, the use of CNNs for microseismic waveform identification is a promising approach that has significant potential to enhance our understanding of rockburst activity and improve our ability to detect and respond to microseismic events. With the ongoing advancements in the field of deep learning, we can expect continuous enhancements in both the accuracy and efficiency of these models. Using the same dataset, the existing CNN [1] model, AlexNet, GoogLeNet, and ResNet50 classic image classification models were first trained for modeling. Figure 5 shows the comparison results of the loss rate and accuracy of different CNN models during the training and validation process. Specifically, Figures 5(a) and 5(b) show the variation of loss rate and accuracy for the five models on the training set respectively. Figures 5(c) and 5(d) display the variation of loss rate and accuracy for the five models on the validation set respectively. The training and validation loss and accuracy trends of the five models were basically the same, and all of them were consistent with the variation of neural network training. Among them, the accuracy of the validation ranged from 97.75% to 98.65%, and the loss was basically in the range of 0.041-0.074. The CNN [1] model fluctuated the most in training and validation, while ResNet50 was the most stable and the best in validation. MS-CNN performed moderately in the training and validation process among these five models. (a) (b) (c) (d) Figure 5. Training and validation results of five CNN models (a) training loss, (b) training accuracy, (c) validation loss, and (d) validation accuracy. Based on the five selected evaluation metrics, the performance and effectiveness of the 5 CNN models can be assessed by comparing their time requirements and recognition results. The best model will be determined by the requirements of the particular application, such as accuracy, speed, and memory utilization.
  • 9.  ISSN: 2252-8938 Int J Artif Intell, Vol. 13, No. 1, March 2024: 1038-1049 1046 For real-time applications, for instance, a model with high accuracy and little inference time can be favored, whereas a model requiring more memory and more training time might be chosen for applications where precision is crucial. Ultimately, the architecture, hyperparameters, dataset, and hardware employed will all depend on the identification results and processing time of CNN models, which must be assessed on a case- by-case basis shown in Figure 6. (a) (b) Figure 6. Identification results and time consumption for five CNN models (a) test results and (b) time consumption Compared to the existing CNN model [1], we reduced the number of convolutional layers, modified the feature maps' parameters, and set the dropout value to 0.4. After multiple rounds of training and hyperparameter tuning, we obtained the optimal MS-CNN model. Experimental results indicate that MS-CNN is simpler and more efficient than the original CNN model [1] and transfer models such as AlexNet, GoogLeNet, and ResNet50 because it has fewer convolutional layers while maintaining good recognition accuracy on a limited dataset. From the comparative analysis of test results in Figure 6, it is clear that MS- CNN outperforms other models in terms of classification accuracy (99.6%) and recognition efficiency (which takes only 0.31 seconds to recognize 277 images in the test set) for the classification and recognition of microseismic events. 5. CONCLUSION The recognition and classification of multi-channel microseismic waveforms using CNNs is a promising approach for accurately identifying microseismic signals. CNN-based models leverage the power of deep learning algorithms to effectively extract and analyze features from complex microseismic data, enabling accurate event classification based on waveform characteristics. The use of multi-channel data allows for a more comprehensive analysis of microseismic events, leading to improved classification accuracy and reduced false positives. Moreover, the adaptability of CNNs to new data makes them suitable for real-time applications, such as rockburst monitoring and early warning systems. Developing CNN-based models for microseismic waveform identification has significant potential to enhance our understanding of mining seismic activity and improve our ability to detect and respond to microseismic events. Advancements in deep learning will further increase the accuracy and efficiency of these models, making them increasingly valuable tools for rockburst analysis and prediction. This paper proposed the MS-CNN model, which demonstrated outstanding
  • 10. Int J Artif Intell ISSN: 2252-8938  Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu) 1047 performance in recognizing and classifying multichannel microseismic signal waveforms with high speed and accuracy. Compared to existing CNN models and classical image recognition and classification models, such as AlexNet, GoogLeNet, and ResNet50, the MS-CNN model achieved the best classification results with 99.6% accuracy and the shortest dataset identification time. Therefore, the MS-CNN method is practical for engineering applications where automatic recognition and classification of microseismic events and blasts are necessary. ACKNOWLEDGEMENTS The writers express their sincere appreciation for the reviewers' valuable time dedicated to reading this paper. Furthermore, we would like to extend our gratitude to the International College of Digital Innovation at Chiang Mai University for providing us with the opportunity to carry out this study. REFERENCES [1] L. Dong, Z. Tang, X. Li, Y. Chen, and J. Xue, “Discrimination of mining microseismic events and blasts using convolutional neural networks and original waveform,” Journal of Central South University, vol. 27, no. 10, pp. 3078–3089, Oct. 2020, doi: 10.1007/s11771-020-4530-8. [2] Q. Feng, L. Han, B. Pan, and B. Zhao, “Microseismic source location using deep reinforcement learning,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–9, 2022, doi: 10.1109/TGRS.2022.3182991. [3] K. Ma, X. Sun, Z. Zhang, J. Hu, and Z. Wang, “Intelligent location of microseismic events based on a fully convolutional neural network (FCNN),” Rock Mechanics and Rock Engineering, vol. 55, no. 8, pp. 4801–4817, Aug. 2022, doi: 10.1007/s00603-022- 02911-x. [4] X. Yin, Q. Liu, X. Huang, and Y. Pan, “Real-time prediction of rockburst intensity using an integrated CNN-Adam-BO algorithm based on microseismic data and its engineering application,” Tunnelling and Underground Space Technology, vol. 117, p. 104133, Nov. 2021, doi: 10.1016/j.tust.2021.104133. [5] L. Dong and Q. Luo, “Investigations and new insights on earthquake mechanics from fault slip experiments,” Earth-Science Reviews, vol. 228, p. 104019, May 2022, doi: 10.1016/j.earscirev.2022.104019. [6] Y. Pu, D. B. Apel, V. Liu, and H. Mitri, “Machine learning methods for rockburst prediction-state-of-the-art review,” International Journal of Mining Science and Technology, vol. 29, no. 4, pp. 565–570, Jul. 2019, doi: 10.1016/j.ijmst.2019.06.009. [7] M. Chakraborty, M. Das, and S. Aruchamy, “Micro-Seismic Event Detection using statistical feature extraction and machine learning techniques,” in 2022 IEEE 7th International conference for Convergence in Technology (I2CT), IEEE, Apr. 2022, pp. 1– 5. doi: 10.1109/I2CT54291.2022.9824819. [8] J. X. Wang, S. B. Tang, M. J. Heap, C. A. Tang, and L. X. Tang, “An auto-detection network to provide an automated real-time early warning of rock engineering hazards using microseismic monitoring,” International Journal of Rock Mechanics and Mining Sciences, vol. 140, p. 104685, Apr. 2021, doi: 10.1016/j.ijrmms.2021.104685. [9] D. Wamriew, R. Pevzner, E. Maltsev, and D. Pissarenko, “Deep neural networks for detection and location of microseismic events and velocity model inversion from microseismic data acquired by distributed acoustic sensing array,” Sensors, vol. 21, no. 19, p. 6627, Oct. 2021, doi: 10.3390/s21196627. [10] J. Li et al., “Automatic recognition and classification of microseismic waveforms based on computer vision,” Tunnelling and Underground Space Technology, vol. 121, p. 104327, Mar. 2022, doi: 10.1016/j.tust.2021.104327. [11] J. He, L. Dou, S. Gong, J. Li, and Z. Ma, “Rock burst assessment and prediction by dynamic and static stress analysis based on micro-seismic monitoring,” International Journal of Rock Mechanics and Mining Sciences, vol. 93, pp. 46–53, Mar. 2017, doi: 10.1016/j.ijrmms.2017.01.005. [12] Y. Pu, D. B. Apel, and R. Hall, “Using machine learning approach for microseismic events recognition in underground excavations: comparison of ten frequently-used models,” Engineering Geology, vol. 268, p. 105519, Apr. 2020, doi: 10.1016/j.enggeo.2020.105519. [13] K. Thamprasert, A. Y. Dawod, and N. Chakpitak, “Simulated trial and error experiments on productivity,” IAES International Journal of Artificial Intelligence (IJ-AI), vol. 11, no. 4, p. 1570, Dec. 2022, doi: 10.11591/ijai.v11.i4.pp1570-1578. [14] H. Wu, B. Zhang, and N. Liu, “Self-adaptive denoising net: self-supervised learning for seismic migration artifacts and random noise attenuation,” Journal of Petroleum Science and Engineering, vol. 214, p. 110431, Jul. 2022, doi: 10.1016/j.petrol.2022.110431. [15] Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, “A survey of convolutional neural networks: analysis, applications, and prospects,” IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 12, pp. 6999–7019, Dec. 2022, doi: 10.1109/TNNLS.2021.3084827. [16] R. M. Abd-ElGhaffar, M. El-Zalabany, H. E.-D. Moustafa, and M. El-Seddek, “Classification of focal liver disease in egyptian patients using ultrasound images and convolutional neural networks,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 27, no. 2, p. 793, Aug. 2022, doi: 10.11591/ijeecs.v27.i2.pp793-802. [17] S. Tang, J. Wang, and C. Tang, “Identification of microseismic events in rock engineering by a convolutional neural network combined with an attention mechanism,” Rock Mechanics and Rock Engineering, vol. 54, no. 1, pp. 47–69, Jan. 2021, doi: 10.1007/s00603-020-02259-0. [18] D. Anikiev et al., “Machine learning in microseismic monitoring,” Earth-Science Reviews, vol. 239, p. 104371, Apr. 2023, doi: 10.1016/j.earscirev.2023.104371. [19] K. Peng, Z. Tang, L. Dong, and D. Sun, “Machine learning based identification of microseismic signals using characteristic parameters,” Sensors, vol. 21, no. 21, p. 6967, Oct. 2021, doi: 10.3390/s21216967. [20] L.-J. Dong, J. Wesseloo, Y. Potvin, and X.-B. Li, “Discriminant models of blasts and seismic events in mine seismology,” International Journal of Rock Mechanics and Mining Sciences, vol. 86, pp. 282–291, Jul. 2016, doi: 10.1016/j.ijrmms.2016.04.021. [21] L. Dong, J. Wesseloo, Y. Potvin, and X. Li, “Discrimination of mine seismic events and blasts using the fisher classifier, naive bayesian classifier and logistic regression,” Rock Mechanics and Rock Engineering, vol. 49, no. 1, pp. 183–211, Jan. 2016, doi: 10.1007/s00603-015-0733-y. [22] C. Ma et al., “A novel microseismic classification model based on bimodal neurons in an artificial neural network,” Tunnelling and
  • 11.  ISSN: 2252-8938 Int J Artif Intell, Vol. 13, No. 1, March 2024: 1038-1049 1048 Underground Space Technology, vol. 131, p. 104791, Jan. 2023, doi: 10.1016/j.tust.2022.104791. [23] A. H. Wilkins, A. Strange, Y. Duan, and X. Luo, “Identifying microseismic events in a mining scenario using a convolutional neural network,” Computers & Geosciences, vol. 137, p. 104418, Apr. 2020, doi: 10.1016/j.cageo.2020.104418. [24] L. Alzubaidi et al., “Review of deep learning: concepts, CNN architectures, challenges, applications, future directions,” Journal of Big Data, vol. 8, no. 1, p. 53, Mar. 2021, doi: 10.1186/s40537-021-00444-8. [25] S. M. Mousavi and G. C. Beroza, “Deep-learning seismology,” Science, vol. 377, no. 6607, Aug. 2022, doi: 10.1126/science.abm4470. [26] Y. Pratama, E. Marbun, Y. Parapat, and A. Manullang, “Deep convolutional neural network for hand sign language recognition using model E,” Bulletin of Electrical Engineering and Informatics, vol. 9, no. 5, pp. 1873–1881, Oct. 2020, doi: 10.11591/eei.v9i5.2027. [27] B. I. Lin, X. I. E. Wei, Z. H. A. O. Junjie, and Z. H. A. O. Hui, “Automatic classification of multi-channel microseismic waveform based on DCNN-SPP,” Journal of Applied Geophysics, vol. 159, pp. 446–452, Dec. 2018, doi: 10.1016/j.jappgeo.2018.09.022. [28] B. Lin, X. Wei, and Z. Junjie, “Automatic recognition and classification of multi-channel microseismic waveform based on DCNN and SVM,” Computers & Geosciences, vol. 123, pp. 111–120, Feb. 2019, doi: 10.1016/j.cageo.2018.10.008. [29] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” Sep. 2014, [Online]. Available: http://arxiv.org/abs/1409.1556 [30] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” Dec. 2015, [Online]. Available: http://arxiv.org/abs/1512.03385 [31] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” Communications of the ACM, vol. 60, no. 6, pp. 84–90, May 2017, doi: 10.1145/3065386. [32] J. Wang and S. Tang, “Novel transfer learning framework for microseismic event recognition between multiple monitoring projects,” Rock Mechanics and Rock Engineering, vol. 55, no. 6, pp. 3563–3582, Jun. 2022, doi: 10.1007/s00603-022-02790-2. [33] A. S. Hatem, M. S. Altememe, and M. A. Fadhel, “Identifying corn leaves diseases by extensive use of transfer learning: a comparative study,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 29, no. 2, p. 1030, Feb. 2023, doi: 10.11591/ijeecs.v29.i2.pp1030-1038. [34] W. Setiawan, M. I. Utoyo, and R. Rulaningtyas, “Reconfiguration layers of convolutional neural network for fundus patches classification,” Bulletin of Electrical Engineering and Informatics, vol. 10, no. 1, pp. 383–389, Feb. 2021, doi: 10.11591/eei.v10i1.1974. [35] D.-X. Zhou, “Theory of deep convolutional neural networks: Downsampling,” Neural Networks, vol. 124, pp. 319–327, Apr. 2020, doi: 10.1016/j.neunet.2020.01.018. [36] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning (adaptive computation and machine learning series). Cambridge Massachusetts. The MIT Press, 2017. [37] R. Ali, J. H. Chuah, M. S. A. Talip, N. Mokhtar, and M. A. Shoaib, “Structural crack detection using deep convolutional neural networks,” Automation in Construction, vol. 133, p. 103989, Jan. 2022, doi: 10.1016/j.autcon.2021.103989. [38] A. Bahri, S. G. Majelan, S. Mohammadi, M. Noori, and K. Mohammadi, “Remote sensing image classification via improved cross- entropy loss and transfer learning strategy based on deep convolutional neural networks,” IEEE Geoscience and Remote Sensing Letters, vol. 17, no. 6, pp. 1087–1091, Jun. 2020, doi: 10.1109/LGRS.2019.2937872. [39] W. Setiawan, M. I. Utoyo, and R. Rulaningtyas, “Transfer learning with multiple pre-trained network for fundus classification,” TELKOMNIKA (Telecommunication Computing Electronics and Control), vol. 18, no. 3, p. 1382, Jun. 2020, doi: 10.12928/telkomnika.v18i3.14868. BIOGRAPHIES OF AUTHORS Hongmei Shu received a B.E. degree (2016) and the M.E. degree (2019) from School of Computer Science and Technology at Sichuan Normal University. From July 2019 to February 2022, she worked as a lab technician in the School of Information and Engineering at Sichuan Tourism University. She is currently a Ph.D. candidate in Digital Innovation and Financial Technology at Chiang Mai University. From March 2022 to the present, she is also a research assistant in the School of Resource Security and Engineering at Central South University. Her research interests include image/signal processing, machine learning, data mining, and artificial intelligence. She can be contacted at email: hongmei_shu@cmu.ac.th. Dr. Ahmad Yahya Dawod is currently a lecturer at International College of Digital Innovation at Chiang Mai University, Thailand. He received his Ph.D. degree in Machine Learning and Artificial Intelligence from the National University of Malaysia in 2018 with the topic “Hand gesture recognition based on isolated and continuous sign language”. He also graduated his master’s degree in computing and informatics from Multimedia University of Malaysia and had his bachelor’s degree in computer science from The University of Mustansirya of Iraq. His research includes machine learning, pattern recognition, computer vision, robotics, and artificial intelligence. He had published 20 articles up to date with more than a hundred citations. He can be contacted at email: ahmadyahyadawod.a@cmu.ac.th.
  • 12. Int J Artif Intell ISSN: 2252-8938  Multi-channel microseismic signals classification with convolutional neural networks (Hongmei Shu) 1049 Dr. Worawit Tepsan is currently a lecturer at International College of Digital Innovation at Chiang Mai University, Thailand, since December 2020. He received his Ph.D. degree in Mathematics from the University of Houston, US in 2020. His research includes mathematical operator theory, functional analysis, applied mathematics, machine learning, and data analysis. He can be contacted at email: worawit.tepsan@cmu.ac.th. Dr. Lei Mu has been a full lecturer of Chengdu University, Mainland China since 2003. He received his BE and ME from the University of Electronic Science and Technology of China. In 2011, he obtained a full scholarship from Erasmus Mundus Action 2-a Sustainable e-Tourism Project and got his PhD from the University of Grenoble Alps, France. He was appointed Deputy Dean of the School of Foreign Languages at Chengdu University in December 2014. He is responsible for research, faculty, training and international cooperation. Since December 2020 till now, he works as Deputy Director of the Office of International Collaboration and Exchange, Chengdu University. His main research areas include "technology application and innovation in international tourism" and "computer-assisted translation". He can be contacted at email: mulei@cdu.edu.cn. Zheng Tang holds a Bachelor of Engineering (B.Eng.) in Mining Engineering, Master of Science (M.Sc.) in Mining Engineering, besides several professional certificates and skills. He is currently studying for a doctorate in the School of Resource and Safety Engineering at the Central South University. His research areas of interest include disaster warning, artificial intelligent, and signal identification. He can be contacted at email: tzheng@csu.edu.cn.