INTRODUCTION OF A NOVEL ANOMALOUS SOUND DETECTION METHODOLOGY

International Journal on Cybernetics & Informatics (IJCI) Vol. 11, No.4, August 2022
David C. Wyld et al. (Eds): AIAD, EDU, COMSCI, IOTCB - 2022
pp. 31-51, 2022. IJCI - 2022 DOI:10.5121/ijci.2022.110403
INTRODUCTION OF A NOVEL ANOMALOUS
SOUND DETECTION METHODOLOGY
Xiao Tan and Prof. S M Yiu
Dept of Computer Science, University of Hong Kong, Hong Kong
ABSTRACT
This paper is to introduce a novel semi-supervised methodology, the enhanced incremental principal
component analysis (“IPCA”) based deep convolutional neural network autoencoder (“DCNN-AE) for
Anomalous Sound Detection (“ASD”) with high accuracy and computing efficiency. This hybrid
methodology is to adopt Enhanced IPCA to reduce the dimensionality and then to use DCNN-AE to extract
the features of the sample sound and detect the anomality. In this project, 228 sets of normal sounds and
100 sets of anomaly sounds of same machine are used for the experiments. And the sound files of machines
(stepper motors) for the experiments are collected from a plant site. 50 random test cases are executed to
evaluate the performance of the algorithm with AUC, PAUC, F measure and Accuracy Score. IPCA Based
DCNN-AE shows high accuracy with the average AUC of 0.815793282, comparing with that of Kmeans++
of 0.499545351, of Incremental PCA based DBSCAN clustering of 0.636348073, of Incremental based
PCA based One-class SVM of 0.506749433 and of DCGAN of 0.716528104. From the perspective of
computing efficiency, because of the dimensions-reduction by the IPCA layer, the average execution time
of the new methodology is 15 minutes in the CPU computing module of 2.3 GHz quad-core processors,
comparing with that of DCGAN with 90 minutes in GPU computing module of 4 to 8 kernels.
KEYWORDS
Internet of things (“IoT”), Anomalous Sound Detection (“ASD”), Audio Frames, Deep learning, Deep
Convolutional Neural Network, Autoencoder.
1. INTRODUCTION
1.1. Anomalous Sound Detection and Issues of Implementation to Industry
Internet of things (“IoT”) and industry 4.0 are becoming more and more important in both
research and industrial domains. Industry 4.0 is a paradigm within automation and data
transformation in manufacturing technologies, which help to build up the intelligent networking
of devices, machines, and processes in the enterprises. Industrial Internet of Things (“IoT”), as
one of the key components of industry 4.0, is to adopt the devices, including sensors, actuators,
computers or “smart” objects to collect the operation data in the manufacturing process, to
monitor the operation status of the machines and to detect the anomalous behaviours of the
machines. The machine learning/deep learning algorithms are adopted widely to deploy the
artificial intelligent agents for IoT data analysis.
Anomalous Sound Detection (“ASD”) is one prominent approach of IoT, which is to adopt
unsupervised or semi-supervised ML/DL methodologies to extract the features of the sound data
of the equipment, and then to observe and detect the anomalous sound emitted from the target
machinery as the basis of the warning of the operation status of the targets and send it to the
human operators. All machinery in factories is subject to breakdown, which causes potential risk
of loss for the companies [1]. Experienced technicians are able to diagnose such failures and the

32
machines’ operational conditions by listening to the machinery [2]. As a major type of Predictive
Maintenance Technologies, acoustic monitoring sensor could detect the anomalous changes of
the machine’s condition before the breakdown really happens. Based on the statistics result
published by US Department of Energy in 2020, the functional predictive maintenance program
reduces the maintenance costs of 25%-30%, eliminates the breakdown of 70%-75%, reduces the
downtime of 35%-45% and increases the production of 20%-25% [5]. However, based on the
studies in United States in the winter of 2020, only 12% maintenance resources and activities of
average facilities are predictive [3]. Therefore, the need of the predictive maintenance is huge and
on a high priority. especially when the machines work in an environment which is difficult for
humans to enter or work continuously in (eg: the Deep Ocean Drifters), or there is lack of the
labours in the companies which are quite normal in the developed countries or regions, the
quality of the maintenance service will be impacted significantly due to the shortage of human-
friendly environment or skilled workers. To resolve this issue, the technology of “smart”
anomalous sound detection with high precise and sensitivity on operating acoustic data is
developed.
Except for the lower cost of predictive maintenance, the reduced risk of deaths and severe
injuries due to the equipment incidents is another important reason to implement the Anomalous
Sound Detection technology in the maintenance practices. For example, incidents involving lifts,
elevators and escalators each year kill 31 people and seriously injure about 17,000 people in the
United States, according to the Centre to Protect Workers’ Rights (“CPWR”) analysis of data
provided by the BLS (1992-2009) and the Consumer Product Safety Commission (“CPSC”)
(1997-2010) [4]. In Hong Kong, the annually Reported Lift Incident Records are 366 on average
(2011-2021) [5]. With Anomalous Sound Detection technology, the working conditions of the
lifts or elevators are capable to be monitored and predicted in time before the incidents really
happen and therefore reduce the loss of the peoples’ lives and weaken the harms to the safety of
equipment located environments.
However, there are quite a few challenges of the ASD, although the importance of its role is
increasing with the development of industry 4.0 and IoT. And those challenges are mainly related
to the limit of the data collections, which includes:
Imbalanced Training Dataset. For most IoT applications, the anomaly events with a long time-
series data are rare. Generally, the IoT generates a large volume of normal data with 24*7, and
the actual anomalous sounds rarely occur and are highly diverse [6] because of the changing and
complex operating conditions [7][8]. The ratio of anomalous data files to normal ones depends on
the different types of machines, maintenance conditions and external environments eg: changes
of seasons etc. In practice, the range of such ratio varies from 5% or less to more than 20% when
the machines’ conditions are getting damaged. Therefore, to identify the unknown anomalous
sounds from the given exhaustive continuous normal data is a challenge for traditional ASD
technology.
Noise. In most occasions, the sound data is collected in the real manufacturing environments
where multiple machines are placed together to emit different types of sounds, therefore, the
environmental noise is a potential issue to impact the data quality significantly[9]. Especially, the
sudden changes of the noise characteristics increase the difficulty of the detection of the
anomalous sound in the noise contaminated environment, and lead to cognitive performance
decline [10]. Therefore, a machine learning methodology with noise-tolerance is crucial for the
application for the ASD. However, the traditional methodologies, including both classification
and clustering, are not capable to handle this issue from the nature.

33
Hardcoded Architecture. Because most different local environments have their own unique
acoustic signatures, therefore the localized models are required for each different environment or
domain. Lack of automatic self-adaptability and hardcoded framing are the defects of most
algorithms because of the requirement to set different parameters for the different characteristics
of the original acoustic data and the external environments. And such settings impact the
efficiency and accuracy of ASD.
Computation Cost. To achieve the satisfactory performance of features extractions of audio
files, the enormous size of the training data and the architecture of the machine learning or deep
learning algorithms require high computation capability, and thus the computation cost keeps
high. For example, for some generative adversarial network or autoencoder, the continuous
computation capability of one or multiple GPUs are required when keeping successful batch-
running. The computing cost can reach thousands of dollars per machine per month.
To resolve these four issues, different with the methodologies introduced in this chapter, as the
semi-supervised algorithm, the Deep Convolutional Neural Network Autoencoder has the
capability to train the labelled normal dataset and classify the unlabelled anomalous sound when
lacking the known abnormal scenarios. To aware the noise robustness, a noise is collected in the
environment for experiments purpose. Hence, the encoder inputs are the clean sample values plus
the noise for training and prediction.
And the audio segments in the pre-processing phrase are referred as one input parameters of the
DCNN Auto Encoder, therefore the adaptability of the deep learning algorithm is relatively
higher than other unsupervised or semi-supervised algorithms with high accuracy and computing
efficiency.
In the next chapter, several well-known unsupervised and supervised algorithms are introduced.
The pros and cons of each algorithm are discussed.
1.2. Limitation of Existing Solutions
As one of the hottest topics in the research field of industry 4.0, Anomalous Sound Detection
develops quite fast during the past decades. As defined by V. Chandola, A. Banerjee, and V.
Kumar in 2009[11], anomalies refer to the behaviour patterns that do not conform to a well-
defined notion of normal behaviour. Based on the definition, the characteristics of the anomalies
to differ them from the normal ones includes (1) the scarcity as the anomaly event occurs quite
less frequently than normal ones; (2) the features which are able to be identified or extracted from
the normal ones;(3) the meaning to be represented by the anomality [12]. Because of the
characteristics of the anomalies, many algorithms, mainly unsupervised and semi-supervised, are
created for the fault diagnose and anomaly prediction. Several of them are discussed here.
Classification is one type of the earliest unsupervised machine learning (“ML”) methodology, in
which Principal Component Analysis (“PCA”), and Linear Discriminant Analysis (“LDA”) are
widely used to reduce the dimensions by looking for the linear combinations of variables [13].
The main challenge of such classification is that the characteristics of the data should be known
in advance, eg: the count of the variables, the class should be Gaussian etc. Another disadvantage
is the accuracy when deploying such algorithms in practice, especially when predicting the
changes of the temporal shapes after the long-time training of the normal dataset.
Clustering is another type of the unsupervised ML methodology in ASD. In clustering, K-means
is the most popular unsupervised algorithm [14]. The advantage of K-means is the high efficiency
because of its linear complexity. The disadvantages include that the count of clusters should be

34
pre-set as an input and that the inconsistent clustering results due to its random starting point
selection. Agglomerative Hierarchical Clustering is another widely-used clustering
algorithm[15]. The advantage of the hierarchical algorithm is that it doesn’t require the count of
clusters as an input. The disadvantages include the low efficiency and the low accuracy when
processing a large volume of data because of its dendrogram [16].
Both classification and clustering algorithms are less noise-tolerant because of their architecture.
That means, these two types of algorithms perform weak when deploying to noise contaminated
environments.
Support Vector Machine (“SVM”) and Convolutional Neural Network (“CNN”) are two well-
known supervised ML algorithms [17][18]. The accuracy of supervised ML algorithms can be
much high. But such methodology cannot be deployed to imbalanced training datasets because of
the lack of known anomalous samples.
Except for the traditional algorithms discussed above, the latest developments in the ASD
algorithms include the Variational Autoencoder (“VAE”)[19] and the Deep Convolutional
Generative Adversarial Network (“DCGAN”)[20], and other generative adversarial networks for
the unsupervised deep learning. With appropriate architecture design, such unsupervised DL
algorithms can achieve quite high accuracy for with imbalanced training datasets. The
disadvantage is that those algorithms require relatively higher computation capability and costs.
In this paper, the incremental principal component analysis based deep convolutional neural
network autoencoder is introduced. The details of the new DL algorithm will be described in the
section B,C, D, E F and G.
1.3. Methodology Overview
The proposed autoencoder was designed in Tensorflow and Pytorch which are Google and
Facebook open-source software for machine learning respectively. With the dimensionality
reduction by Enhanced IPCA, the autoencoder has the capability to learn efficient feature
representations of the input data with sole label, i.e. the training requires only one label for the
normal sound datasets, and then reconstruct the input to generate the output based on the useful
features learned by the autoencoder. It consists of two parts. One is an encoder 𝑔 = 𝑒𝑛𝑐(𝑥),
which encodes the input x to a learned feature representation 𝑔. The encoder extracts meaningful
features at the bottleneck layer, known as the code or latent space, which typically has a much
lower dimensionality than the input data. This ensures that, instead of copying the input data,
only the useful features are adopted in the calculation scope to reconstruct the output. The other is
the decoder 𝑥
̂ = 𝑑𝑒𝑐(𝑔) which attempts to recapture the original input by decoding the useful
features stored in the bottleneck layer. For a deep convolutional AE model, the encoder 𝑒𝑛𝑐𝜃(𝑥)
and decoder 𝑑𝑒𝑐𝛾(𝑔) are modelled as deepconvolutional networks with parameters 𝜃 and 𝛾,
respectively. Therefore, the training of the deep convolutional neural network autoencoder is
described as the formula below:
min
𝜃,𝛾
∑ 𝐸𝑟𝑟𝑜𝑟𝑟𝑒𝑐𝑛(𝑥, 𝑥
̂
𝑥
),
𝑤ℎ𝑒𝑟𝑒 𝑥
̂ = 𝑑𝑒𝑐𝛾(𝑒𝑛𝑐𝜃(𝑥))
For 𝐸𝑟𝑟𝑜𝑟𝑟𝑒𝑐𝑛(𝑥, 𝑥
̂),, we can use mean-squared error (“MSE”) as the evaluation of the
reconstruction errors between the input x and the output 𝑥
̂.

35
This methodology is capable to achieve high performance without the requirement on high
computing capability especially for the large volume and high dimensional machines’ sound files.
The details of the encoder and decoder are described in the next section C.
1.4. Imbalanced Dataset
IPCA based DCNN-AE, as a semi-supervised machine learning algorithm, is capable to train the
large-sized normal dataset and detect the anomality when a small-sized abnormous data enters.
DCNN-AE utilizes the encoder-decoder architecture to train the input data to estimate the output
data to be similar to the input. Therefore, when the input data is anormal, with the trained
architecture, the output generated will be anormal, theoretically.
Because of its ignorance of the normality of input dataset, unbalance is not an issue to impact the
performance of the algorithm in nature.
1.5. Noise Awareness and Tolerance
Noise identification, noise removal and noise accommodation are the major topics to dealwith the
unwanted background noise when analyzing anomalous sound. The architecture of the new
methodology adopts the noise accommodation which to immune the background noise from the
anomalous sound observations[11][12]. Considering the complex environments to collect the
sounds files, i.e. factories or construction sites, in an attempt to improve the robustness against
noise of the algorithm, the noise awareness is included in the data pre-processing phase with a
Gaussian white noise to be added in the original audio files and the signal to noise ratio(“SNR”)
is 6.9897 (SNR=10*Log(1/0.2) while 0.2 is the ratio of noise to real sound, or the noise
significancy factor). Hence, the encoder input actually is the clean sample values plus the noise:
𝑥
̃ = 𝑥 + 𝜖,
𝑤ℎ𝑒𝑟𝑒 𝜖 ~ 𝒩(−1,1) ∗ 𝑛𝑜𝑖𝑠𝑒 𝑠𝑖𝑔𝑛𝑖𝑓𝑖𝑐𝑎𝑛𝑐𝑦 𝑓𝑎𝑐𝑡𝑜𝑟. And the decoded output is redefined as
𝑑𝑒𝑐𝛾(𝑒𝑛𝑐𝜃(𝑥
̃)).
This is to increase the noise awareness in the algorithm and was shown to results in better
generalizing representations, which als0o appears to capture semantic information of the input.
And based on the experiments, the maximum noise tolerance (measured by the noise significancy
factor) of the IPCA based DCNN-AE is 0.5.
1.6. Parameterized Framing Technology
DCNN-AE in this paper adopts the parameterized architecture instead of hardcoded when
defining the parameters of the Incremental PCA layer and the input data shapes of autoencoder
layer. Different with the hardcoded input shapes, eg: 256*256 or 512*512, the temporal of input
data is decided by the size of the audio data files. Each audio file is framed to the shape of
Maximum Squared Root of the audio file data size (which is calculated to be 210) * Maximum
Squared Root of the audio file data size (same as 210). The enhanced IPCA is capable to reduce
the dimensionality to the calculated framing for DCNN-AE to proceed. When the environment
changes and the raw sound data is changed, the algorithm will change the temporal of the input
sound data automatically.

36
Such design, on some level, helps to build the auto-localized models to improve the self-
adaptability of the different environments.
1.7. Less Requirement of Computing Capability and Computing Cost
Enhanced IPCA based DCNN-AE includes the two layers of IPCA and DCNN-AE, which allows
the dimensions of the audio data to be reduced to adapt to the optimized architecture of the
Autoencoder, and thus requires less computing capability of GPU.
The high-dimensional audio data files can be trained and predicted in normal CPU instead of
GPU or TPU, after the technique of dimensions-reduction by enhanced IPCA, with minimum
impacts on the performance. The computing cost because of the continuous GPU(s)
consumptions can be reduced significantly as well.
The details of the algorithm will be introduced in Chapter II.
2. ENHANCED INCREMENTAL PRINCIPAL COMPONENT ANALYSIS BASED
DEEP CONVOLUTIONAL NEURAL NETWORK-AUTOENCODER
2.1. Dimensionality Reduction by Enhanced Incremental Principal Component
Analysis
Principal Component Analysis (“PCA”) is a classical statistical method to reduce the dimensions
in a dataset with many interrelated variables. Since being raised by Pearson as early as in 1901
and Hotelling in 1930’s, PCA has been widely used in science and business domains as an
unsupervised method [21]. The principal of PCA is to seek the subspace of largest variance in
dataset. For data given as a set of n vectors in Rd, 𝑥1, 𝑥2, … 𝑥𝑛, denote by 𝑋 which is the
normalized matrix: 𝑋 =
1
𝑛
∑ 𝑥𝑖𝑥𝑖
𝑇
𝑛
𝑖=1 [22]. The PCA is to find the k-dimensional subspace so that
the projected data on the subspace has largest variance. Assume 𝑊 ∈ 𝑅𝑑 ∗ 𝑘 , PCA could be
formulated as the following optimization problem[23]:
max
𝑊∈𝑅𝑑∗𝑘,𝑊𝑇𝑊=𝐼
||𝑋𝑊||𝐹
2
Where ∥⋅∥𝐹 is the Frobenius norm. The above optimization is an inherently non-convex
optimization problem even for k=1. However, in the classical PCA algorithm, when finding the
leading principal component, there are two methods: one is the singular value decomposition
(SVD), which requires super-linear time and potential 𝑂(𝑑2
) space so that it is prohibitive for
large scale machine learning problems. The other is Power method which is to estimate the
matrix X. The approximation method requires the data to be well-conditioned, and the spectral
gap, i.e. the distance between the largest and second largest eigenvalues of X, to be bounded
away from zero. This method can improve the computing efficiency but might loss the useful
data as the pass-over from the leading components.
In 2002 a new PCA algorithm for incremental computation of PCA is introduced by Arta𝑐̃ et al.,
2002[23]. This method is to make the simultaneous learning and recognition possible. This
algorithm is to update the original eigenspace and mean continuously with the learning rate, and
store only the data projected onto it. It releases the restriction of the classical PCA that all the
sample data are required to be available as a priority, and therefore more adaptive for the demand
online learning.

37
In this paper, an enhanced incremental PCA is adopted for dimension reduction. The summary of
the new algorithm is:
1. Locate the target object in the first frame, and initialize the eigen basis U to be empty, and the
mean μ to be the appearance of the target in the first frame. The effective number of
observations so far is n=1.
2. Then advance to the next frame. Draw particles from the particle filter, according to the
dynamical model. And the dynamics between states in this space is modeled by Brownian
motion. Each parameter in the affine transformation of 𝑋 is modeled independently by a
Gaussian distribution around its counterpart in 𝑋𝑡−1, and therefore the motion between frames
is itself and an affine transformation. Specifically, 𝑝(𝑋𝑡 |𝑋𝑡−1) = 𝒩(𝑋𝑡 ; 𝑋𝑡−1, 𝜓), where 𝜓 is
a diagonal covariance matrix whose elements are the corresponding variances of affirm
parameters.
3. For each particle, extract the corresponding window from the current frame and calculate its
weight, which is its likelihood under the observation model.
4. Perform an incremental update (with a forgetting factor) of the eigen basis, mean, and effective
number of observations.
5. Go to step 3.
The enhanced incremental algorithm is based on the Sequential Karhunen-Loeve (“SKL”)
algorithm of Levy and Lindenbaum (2000)[24], which takes QR decomposition and computes the
SVD of basis vectors. The computational advantage of SKL algorithm is that it has space and
time complexity constant in n. The space complexity is reduced to 𝑂(𝑑(𝑘 + 𝑚)), and the
computational requirements are reduced to 𝑂(𝑑𝑚2
) for recomputing the entire SVD[25].
However, this algorithm does not account for the sample mean of the training data, which
changes over time as a new data arrive. The enhanced incremental PCA is to augment the new
training data with an additional vector carefully chosen to correct for the time-varying mean. The
principal computation process is:
Given U and ∑ from the SVD of (A – 𝐼𝐴
̅ ), as well as 𝐼𝐴
̅, n, and B, compute 𝐼𝐶
̅ as well as U’ and
∑’ from the SVD of (C-𝐼𝐶
̅):
1. Compute the mean vector
1
𝑚
Σ𝑖=𝑛+1
𝑛+𝑚
𝐼𝑖, and 𝐼𝐶
̅ =
𝑛
𝑛+𝑚
𝐼𝐴
̅ +
𝑚
𝑛+𝑚
𝐼𝐵
̅ .
2. Form the matrix 𝐵
̂ = [(𝐼𝑚+1 − 𝐼𝐵
̅ )…(𝐼𝑛+𝑚 − 𝐼𝐵
̅)√
𝑛𝑚
𝑛+𝑚
(𝐼𝐵
̅ − 𝐼𝐴
̅)].
3. Compute 𝐵
̃ = 𝑜𝑟𝑡ℎ(𝐵
̂ − 𝑈𝑈⊺
𝐵
̂) and 𝑅 = [
∑ 𝑈⊺
𝐵
̂
0 𝐵
̃(𝐵
̂ − 𝑈𝑈⊺
𝐵
̂ )
] , and𝐵
̂will be one column
larger than in the SKL algorithm.
4. Compute the SVD of R: R = 𝑈
̃∑
̃𝑉⊺
̃
5. Finally𝑈′
= [𝑈, 𝐵
̃]𝑈
̃ and ∑ = ∑
̃
′
.
In this algorithm, an important consideration is the effect of the forgetting factor on the mean of
eigen basis when reducing the contribution of each block of data to the overall covariance
modelled by an additional factor of f2 at each SVD update. And the forgetting factor are
multiplied with the incremental update of the eigen basis, mean and effective number of
observations to generate the new frame.
SVD

38
Another enhancement is in the Sequential Inference Model. When calculating the affine motion
parameters (and thereby the location) of the target at time t, the Condensation algorithm
(proposed by Isard and Blake in 1996)[26], based on factor sampling, is adopted to approximate
an arbitrary distribution of observations with a stochastically generated set of weighted samples,
as it evolves over time.
The new PCA algorithm above is adopted to improve the accuracy with computing efficiency. By
Enhanced IPCA, the components for DCNN-AE to extract features can be reduced to fit for the
optimized architecture designs of DCNN-AE based on the parameterized framing shapes.
2.2. Deep Convolutional Neural Network Autoencoder
An autoencoder is a feed-forward neural network and is an unsupervised or semi-supervised
algorithm that applies backpropagation, setting the target value to be equal to the input[27].
Generally speaking, it is trained to approximate the identification so that the output 𝑥
̂is similar to
the input 𝑥.
The autoencoder is consisted of encoder which obtains a compressed encoder from the input and
of decoder which reconstructs the data from encoding. The encoder maps the input information to
the compressed nodes and creates the bottleneck which forces the network to extract the
compressed low-dimensional representations from high-dimensional data [28]. And based on the
extracted information in the bottleneck, or, hidden layer, the decoder reconstructs the input data,
with the backpropagation to minimize the reconstruction error which is defined as Mean Squared
Error (“MSE”). The structure can be described as the two formulas below [29][30].
Encoder: ℎ = 𝜎(𝑊𝑥ℎ𝑥 + 𝑏𝑥ℎ).
Here ℎ is the hidden layer, or bottleneck, 𝜎 is the non-linear mapping relationship between the
compressed nodes in the hidden layer and the original inputs. 𝑊𝑥ℎ are the weights when mapping
the original inputs to the nodes of the hidden layers. 𝑏𝑥ℎ are the biases from the original inputs to
the nodes of the hidden layers.
Decoder: 𝑥
̂ = 𝜎
̂(𝑊ℎ𝑥ℎ + 𝑏ℎ𝑥).
Here 𝑥
̂ is the output of the decoder, and 𝜎
̂ is the non-linear mapping relationship between the
compressed nodes in hidden layers and the final output. 𝑊ℎ𝑥 are the weights to map the
compressed nodes in the hidden layers to the outputs. And 𝑏ℎ𝑥are the biases from the compressed
data in the hidden layers to the output [31][32][33].
The architecture of standard autoencoder is shown in Fig 1.

39
Fig. 1. Standard Autoencoder
Deep Convolutional Neural Network Autoencoder (“DCNN-AE”) is the autoencoder composed
of multiple convolutional layers in encoder and inverse convolutional layers in decoder[19]. In
this paper, there are four convolutional layers in encoder part, each with batch-normalization,
Leaky ReLU and Max-pooling to construct a one typical compressed pattern. In decoder, there
are four convolutional transpose layers respectively, each with batch-normalization, Leaky ReLU
and Up-sampling to set up the reconstruction pattern. The architecture of the DCNN-AE is shown
as Fig 2.
Fig. 2. Architecture of Deep Convolutional Neural Network Autoencoder

40
The details of the DCNN-AE algorithm include [34]:
Convolutional and Inverse Convolutional Layers
Convolutional Layers are applied to extract the features based on the patches randomly sampled
from the input, being learned with the weighted matrices known as the filtering masks and
kernels, and then being processed by non-linear operations to generate the feature vector as the
output [35].The principle of the conventional layer of CNN is:
ℎ𝑖,𝑗,𝑘
𝑙
= 𝑤𝑘
𝑙 𝑇
𝑥𝑖,𝑗
𝑙
+ 𝑏𝑘
𝑙
Where ℎ𝑖,𝑗,𝑘
𝑙
is the feature value at location (𝑖, 𝑗) in the kth feature map of lth layer, 𝑤𝑘
𝑙 𝑇
and 𝑏𝑘
𝑙
are
the weight vector and bias of the kth filter of the lth layer respectively, and 𝑥𝑖,𝑗
𝑙
is the input patch
centered at location (𝑖, 𝑗) of the lth layer.
In this paper, 2-dimensional convolutional layers are deployed in the encoder part, and 2-
dimensional transpose convolutional layers are deployed in decoder part to reconstruct the
information transformed and compressed from the original inputs [36].
Batch Normalization Layers are added after the convolutional and inverse convolutional layers to
normalize the activations of the previous layers each batch [37]. Assume in the 𝑛 + 1𝑡ℎ
hiden
layer and ℎ𝑖
(𝑛)
, i=1,2,…,m be the values of the inputs from the previous layer for the a batch ℬ,
the principal of the batch normalization is described as the formulas shown as below:
ℎ𝑖
(𝑛+1)
= 𝐵𝑁𝛾,𝜀 (ℎ𝑖
(𝑛)
) = 𝛾ℎ̇
𝑖
(𝑛)
+ 𝜀
Where:
1. ℎ̇ (𝑛)
=
ℎ(𝑛)−𝜇ℬ
√𝜎ℬ
2+𝜖
2. 𝜇ℬ =
1
𝑚
∑ ℎ𝑖
(𝑛)
𝑚
𝑖=1
3. 𝜎ℬ
2
=
1
𝑚
∑ (ℎ𝑖
(𝑛)
− 𝜇ℬ)2
𝑚
𝑖=1
Here 𝐵𝑁𝛾,𝜑 refers to the Batch Normalization function to the two learning parameters of 𝛾, 𝜀. ℎ𝑖
(𝑛)
is the ith
compressed node in the nth
hidden layer. 1 describes the calculation of ℎ̇ (𝑛)
which is the
covariance of ℎ(𝑛)
.
From the principal formulas above, we can understand that Batch Normalization applies a
transformation that maintains the mean activation of every unit close to zero, then applies scaling
and shifting, parameterized by two learnable parameters 𝛾, 𝜀. By this, batch normalization
whitens the data which is well known for speeding up learning in neural networks and reduces
internal covariance shift, which is used for the change of a layer’s activations distribution over
time.
Pooling
Pooling layer is adopted to reduce the features extracted by convolutional layers. Four 2-
dimesional Max Pooling layers are applied among four convolutional layers to use maximum
values of each cluster of neurons of the prior layer. This is to reduce the dimensionality and the

41
overfitting by introducing the abstracted form of the low-level features generated by the first
convolutional layer. This pooling method is typically effective in this research because the
original input data are pre-processed and transformed to the categories with data type as of
integer.
Tab 1 shows the details of all the layers in this DCNN-AE algorithm.
Full Connected Layer and Classifier
The fully connected layer is utilized to associate neurons within outputs in the last inverse
convolutional layer, which are connected to the Softmax classifier.
Softmax classifier layer, which is used for classification according to the features, is after
multiple convolutional autoencoder layers, max pooling layers, Leaky ReLU layers and full
connected layers. In this paper, the anomalous sound detection results are divided into two
classes, therefore the classifier result can be written by the formula as below:
𝑝𝑖
̂ =
𝑒(𝑥
̂𝑖)
∑ 𝑒(𝑥
̂𝑘)
2
𝑘=1
, i = 0, 1
Here 𝑝𝑖
̂ is the classifier represents the probability of nodules and no nodules. 𝑥
̂𝑖 =
𝜎
̂ ∑ (𝑊ℎ𝑥ℎ + 𝑏ℎ𝑥)
𝑇
𝑡=1 represents the T output features generated through the full connected layer
where 𝑊ℎ𝑥and 𝑏ℎ𝑥represent the weight and error in decoder part respectively, 𝜎
̂ represents the
nonlinear function in decoder, in this paper it refers to ReLU and Leaky ReLU.
ReLU and Leaky ReLU
“Rectified Linear Unit” (“ReLU”) is one of the most notable non-saturated non-linear activation
functions. The principle of ReLU is shown as below:
a. With default values, it returns element-wise max(x, 0). Otherwise, it follows:
b. f(x) = max_value for x >= max_value,
f(x) = x for threshold <= x < max_value,
f(x) = alpha * (x – threshold) for otherwise.
The pro of ReLU is to prune the negative input to reduce the sparsity in the neural network, and
the con of ReLU is that such pruning might impact the accuracy of the computation because the
possible useful features contained in the negative inputs to build the discriminative high-level
features.
Leaky ReLU is a variant of ReLU by assigning none zero output for negative input. Instead of
mapping negative input to zero, the Leaky ReLU uses a predefined linear function to compress
negative input and therefore enables the features of negative parts contained[38]. Therefore
Leaky ReLU achieves the better trade-off between the network sparsity and its input information.
Both ReLU and Leaky ReLU are unbounded functions and solves the gradient vanishing.

42
Table 1. Layers of the DCNN-AE Algorithm
Seq Layer Type Kernel
Size
Feature
Maps
Output Size Padding
1 Input - - None,210,210,1 -
2 Cov 1 2*2 32 None,210,210,32 Same
3 Batch
Normalization
4 Leaky ReLU
5 Max Pool 1 (2,2) None,105,105,32 Same
6 Drop Out 0.1
7 Cov 2 3*3 64 None,105,105,64 Same
8 Batch
Normalization
9 Leaky ReLU
11 Drop Out 0.1
12 Cov 2 5*5 128 None,35,35,128 Same
13 Batch
Normalization
14 Leaky ReLU
16 Drop Out 0.1
17 Cov 2 7*7 256 None,7,7,256 Same
18 Batch
Normalization
19 Leaky ReLU
21 Drop Out 0.1
22 Sequential
23 Cov Transpose 1 1*1 128 None, 1,1,128 Same
24 Batch
Normalization
25 Leaky ReLU
26 Up-sample 1 7 None,7,7,128 Same
27 Drop Out 0.1
28 Cov Transpose 2 7*7 64 None,7,7,64 Same
29 Batch
Normalization
30 Leaky ReLU
32 Drop Out 0.1
34 Batch
Normalization
35 Leaky ReLU
37 Drop Out 0.1
39 Batch
Normalization
40 Leaky ReLU
42 Drop Out 0.1
43 FC1 (Linear) 210*210 None,210,20

43
44 FC1 (ReLU) None,20
45 FC1 (Drop Out) 0.2
46 FC2 (Softmax) 2 None,2
47 Output None,2 Same
Regulatory and Other Hyperparameters
Gaussian Dropout. Gaussian noise is injected to hidden layers of neural networks to act as a
powerful regularity, and it is only active during training of the network.
Optimizer [39][40]. AdaMax is a method for stochastic optimization. It is to compute individual
adaptive learning rates for different parameters from estimates of first and second moments of the
gradients. Different with Adam which updates exponential moving average of the gradients and
the squared gradient based on the unlimited variants, AdaMax updates the gradients with a
simpler bound of maximum value of the calculated second momentum.
In this paper, the learning rate of AdaMax is set as 1e-4, beta1 is set as 0.5 and decay rate is set as
1e-9. “Mean Squared Error” and “Mean Absolute Error” are adopted as the parameters for the
cost evaluations.
The details of the hyperparameters are listed in Tab 2.
Table 2. Hyperparameters of the DCNN-AE Algorithm
Hyperparameter Setting
optimizer Admax (lr=1e-4, beta1=0.5)
dropout rate 0.1
batch normalization Momentum=0.8
Leaky ReLU Alpha=0.2
batch size 16
noise initializer Random Uniform (-1,1)
noise significancy factor 0.2
loss mse
monitor mae
epoch 400
3. EXPERIMENT SETUP
3.1. Dataset and Pre-processing
The data is collected from the stepper motors in a plant site in Changzhou City of Jiangsu
Province in China. The data consists of the normal/anomalous operating sounds of the real
machines. Each recording is a single-channel 2-sec length audio that includes both a target
machine's operating sound and added noise. The sample rate is standard as of 44,100.
In the experiment, the train dataset only includes the normal data from one machine type, and we
randomly pick up 100 normal wave files in the total 228 sample normal for consecutive 50 times.
The test dataset includes 10 anormal audio files, randomly selected from the total 120 sample
anormal data files. The prediction focuses on the turning point from consecutive 100 normal
audio files to 10 anomalous audio files.

44
The steps of data pre-processing include:
Random selection. 100 normal audio files and 10 anormal audio files are selected randomly
from the 228 respective sample normal and 120 anormal audio files.
Conversion. Convert the audio files to the arrays of readable digits.
Framing. Calculate the frame rates and reframe the files with adjusted frame rates.
Adding label. Add the label as of “0” as a column in train dataset only to indicate that the status
of all the data in the data frame are “normal”.
Scalarization and Normalization. Scale and normalize both the train and the test datasets.
For the experiments, the train data size is 88,200,000, and the prediction data size is 8,820,000.
The experiments run for consecutive 50 times for each algorithm with the datasets picked up
randomly in the total dataset. The null values of any attributes are filtered out during the pre-
processing.
3.2. Benchmark System and Result
A simple Deep Convolutional Neural Network (“DCGAN”) is adopted as the benchmark
performance of unsupervised anomaly detection for the experimental dataset. Generative
Adversarial Network is a deep neural net architecturecomprised of a pair of “adversarial” models
called the generator and the discriminator respectively [41][42]. The generator is to capture a
noise vector to map to the data distribution as the fake input of the discriminator. The
discriminator is to discriminate or identify the two inputs, one from the real data distribution and
the other from the fake data distribution generated by the generator, with some policies. Fig 1 is
the architecture of the general adversarial network.
Fig. 3. Architecture of Standard GAN
DCGAN is one type of GAN to apply several convolutional layers in the architecture of GAN.
The details are listed in the table 3.

45
Table 3. Parameters of the DCGAN
Hyperparameter Setting
optimizer Admax of Discriminator (lr=1e-5, beta1=0.5),
Admax of Generator (lr=1e-5, beta1=0.5),
Admax of Compiled (lr=2e-5, beta1=0.5).
dropout rate 0.1
batch normalization Momentum=0.8
Leaky ReLU Alpha=0.2
batch size 16
noise initializer Random Uniform (-1,1)
loss Mean Squared Error
monitor Mean Absolute Error
epoch 2000
With the same audio data, the experimental result is shown as table 4. From the experiments’
result, it can be seen that DCGAN is capable to achieve the accuracy of 0.716528104 which is
relatively satisfying. Besides, the average execution time of DCGAN is 90 minutes in the GPU
computing module. The computation cost of DCGAN is relatively high.
Table 4. Experimental Result of the DCGAN
AUC PAUC F Accuracy
Score
MSE Average
Execution
Time
0.716528104 0.706213763 0.51664777 0.483352227 90 mins
3.3. Noise Tolerance Test
Another series of experiments are conducted to test the maximum noise tolerance of the enhanced
IPCA based DCNN-AE. Based on the experiments results, the performance of the algorithm is
impacted when the SNR is 3.0103 (SNR=10*Log(1/0.5)), in which 0.5 is the noise significancy
factor.
3.4. Comparison of Hardcoded Architecture and Parameterized Architecture
The third series of experiments are to compare the prediction performance between the hardcoded
architecture and parameterized architecture. The experiments’ results show that the parameterized
architecture achieves higher accuracy within less training time. For example, when the shape of
framing is set as 256*256, the average AUC of the hardcoded architecture is around 0.7 with
average execution time of 35 minutes, and the average AUC of the parameterized architecture is
more than 0.8 with the average execution time of 15 minutes.
From the experiment result of one sample shown in Table 5, it can be observed that although the
AUC of hardcoded architecture is 0.8875, the MSE is much higher than that of parameterized
architecture. Therefore, the simulation result is not so satisfying comparing with parameterized
architecture.

46
Table 5. Single Example: Comparison Between Hardcoded Architecture and Parameterized Architecture
Evaluation Item Hardcoded
Architecture
Parameterized
Architecture
AUC 0.8875 1.0
pAUC 0.8026315789473684 1.0
MSE 0.46381313 0.31623548
Graph of the comparison
between actual and
predicted values
3.5. Computer Module and OS Parameters
All computations except basement system (“DCGAN”) were executed on 2.3 GHz quad-core
processors (Turbo Boost up to achieve 3.8 GHz) PC computer with 8 Gb RAM iOS.
DCGAN was executed in Google Cloud Platform with 1 NVIDIA P100 GPU.
4. RESULTS AND ANALYSIS
The experiment is to pick up 110 datasets with 100 normal audio files from total 228 normal files
as training dataset and 10 anormal audio files from total 100 anomalous files as test dataset to test
the IPCA Based DCNN-AE algorithm. Tab 6 shows the summary of the test results.
Table 6. Summary of Performance Evaluation
Number of
Test Cases
AUC PAUC MSE Average
Execution
Time
50 0.815793282 0.713118455 0.27791114 15 mins
From Table 6, it can be seen that IPCA based DCNN-AE shows high accuracy with the average
AUC and PAUC between 0.815793282 and 0.713118455. And the average execution time is
about 15 mins for total data size of 97,020,000 (normal audio dataset for training:88,200,000,
anomalous audio dataset for prediction: 8,820,000).
Fig 4 shows a sample of the prediction results when predicting the turning point changing from
normal status to anomalous status. The results show that the IPCA based DCNN-AE predicts the
turning point with high accuracy, and the AUC of the prediction reaches 1.0.
Fig. 4. ROC Curve of IPCA Based DCNN-AE Algorithms

47
Fig 5 shows the error rate and the loss rate of the sample training process of the IPCA based
DCNN-AE. These two graphs show that the validation error and loss error decreases significantly
in the first 50 batches and then keeps stable in the left batches.
Fig. 5. Error Rate and Loss Rate of One Sample of DCNN-AE
Table 7 shows the AUC and PAUC comparisons among 4 semi-supervised and unsupervised
machine learning algorithms: Kmeans++, one-class SVM[16], density-based spatial clustering of
applications with noise (“DBSCAN”)[43][44] and IPCA Based DCNN-AE.
From the experiments’ results in Table 7, it can be observed that Incremental PCA based DCNN-
AE shows the highest accuracy with the average AUC of 0.815793282, comparing with that of
Kmeans++ of 0.499545351, of Incremental PCA based DBSCAN clustering of 0.636348073, of
Incremental PCA based One-class SVM of 0.506749433 and of DCGAN of 0.716528104. From
the four measures, Incremental PCA based DCNN-AE performs best among the unsupervised or
semi-supervised algorithms.
Table 7. AUC Comparison Between Unsupervised and Semi-Supervised ML Algorithms
Machine Learning AUC PAUC F Accuracy
Score
MSE
Kmeans ++ 0.499545351 0.499960161 0.499545351 0.500454647
IPCA+DBSCAN 0.636348073 0.636325593 0.636348073 0.363651927
IPCA+One-class
SVM
0.506749433 0.500204267 0.497853267 0.493250568
DCGAN 0.716528104 0.706213763 0.51664777 0.483352227
IPCA+DCNN-AE 0.815793282 0.713118455 0.557364341 0.27791114
To compare the efficiency and computing cost of each unsupervised and semi-supervised
algorithms, we adopt the average execution time and the GPU/CPU usage as the measurements.
From Table 8, it can be observed that the IPCA based DBSCAN Clustering and the IPCA based
One-class SVM consumes least execution time and computing cost. Considering the
performance, the IPCA based DCNN-AE, although consumes more time as of 15 minutes, but
reaches the highest accuracy comparing with other algorithms. And because the DCNN-AE
doesn’t require the computing resource of GPU(s), its computing cost is higher than clustering
algorithms, and yet much lower than that of DCGAN.

48
Table 8. Efficiency and Computing Capability Requirement Comparison Between Unsupervised and Semi-
Supervised ML Algorithms
ML/DL Algorithms Average
Execution Time
Per Round
(unit: minute)
CPU / GPU
(unit: core)
Computing Cost
Per Machine Per
Month (unit: USD)
Kmeans ++ 5 CPU N. A
IPCA+DBSCANClustering 2 CPU N. A
IPCA+One-class SVM 20 CPU N. A
DCGAN 90 GPU (8 to 64) 1000 to 8000
IPCA+DCNN-AE 15 CPU N. A
Fig 6 shows the ROC curves of those four semi-supervised and unsupervised machine learning
algorithms. From the graph, it can be observed that DCNN-AE reaches the highest AUC value as
of 0.815793282, one-class SVM and PCA+DBSCAN got the similar AUC of 0.65 and 0.54
respectively. The performances of Kmeans++ is the worst as of 0.5 respectively.
Fig. 6. ROC Curve of Unsupervised and Semi-supervised ML/DL Algorithms
5. CONCLUSION AND FUTURE WORK
The enhanced IPCA Based DCNN-AE for Anomalous Sound Detection, when analyzing and
detecting the machines’ sounds of high-dimensional and significant density distributions between
normal and anomaly, is proved to achieve high accuracy with lower computing cost because of
the dimensions-reduction before DCNN-AE layers. However, the stability of the algorithm to
keep the continuant high accuracy in each execution starting from a random point still needs large
quantities of trainings to improve. Therefore, how to find the optimized parameters to keep
balance among accuracy, efficiency and stability is important when considering deploying the
algorithm in the industry wide.
In the experiment, 100 from total 228 normal files are adopted for training and 10 from total 120
anomalous files as test dataset. The train data size is 88,200,000, and the prediction data size is
8,820,000. With the increase of the data size to 1000 audio files for training and 100 files for test,
the execution time is increased from 15 minutes to 120 minutes. However, the algorithm is still
capable to be executed in CPU computing model with the unimpacted accuracy.

49
Furthermore, the application of the IPCA based DCNN-AE could be extended from Industry 4.0
to the intrusion detection of acoustic data. For example, the detection of voiceprint for smart
appliances could adopt the new methodology with high computing efficiency and accuracy to
reduce the risk of intrusion.
ACKNOWLEDGEMENTS
This project is partially supported by the funding from the HKU-SCF FinTechAcademy.
REFERENCES
[1] Harsh Purohit, Ryo Tanabe, Kenji Ichige, Takashi Endo, Yuki Nikaido, Kaori Suefusa, and Yohei
Kawaguchi, “MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and
Inspection”, Proc. 4th Workshop on Detection and Classification of Acoustic Scenes and Events
(DCASE), 2019.
[2] Koizumi, Yuma, et al. "ToyADMOS: A dataset of miniature-machine operating sounds for
anomalous sound detection." 2019 IEEE Workshop on Applications of Signal Processing to Audio
and Acoustics (WASPAA). IEEE, 2019.
[3] Sullivan, Greg, et al. Operations & maintenance best practices-a guide to achieving operational
efficiency (release 3). No. PNNL-19634. Pacific Northwest National Lab. (PNNL), Richland, WA
(United States), 2020.
[4] McCann, Michael, and N. Zaleski. "Deaths And Injuries Involving Elevators and Escalators."
Retrieved on April 3 (2013): 2016.
[5] Electrical and Mechanical Services Department, “The Summaries of The Reported Lift Incidents.”
Year from 2011 to 2021.
[6] Yuma Koizumi, Yohei Kawaguchi, Keisuke Imoto, Toshiki Nakamura, Yuki Nikaido, Ryo Tanabe,
Harsh Purohit, Kaori Suefusa, Takashi Endo, Masahiro Yasuda, and Noboru Harada, "Description
and Discussion on DCASE2020 Challenge Task2: Unsupervised Anomalous Sound Detection for
Machine Condition Monitoring", arXiv e-prints: 2006.05822, 2020.
[7] Keith Grueneberg, Bongjun Ko, David Wood, Xiping Wang, Dean Steuer, Yeonsup Lim, “IoT Data
Management System for Rapid Development of Machine Learning Models”, IEEE International
Conference on Cognitive Computing (“ICCC”), 2019.
[8] Jae-joon Chung and Hyun-Jung Kim, “An Automobile Environment Detection System Based on
Deep Neural Network and its Implementation Using IoT-Enabled In-Vehicle Air Quality Sensors”,
Sustainability 2020, 12, 2475; doi:10.3390/su12062475, 2020.
[9] Hisashi Uematsu, Yuma Koizumi, Shoichiro Saito, Akira Nakagawa, and Noboru Harada, “Anomaly
Detection Technique in Sound to Detect Faulty Equipment”, NTT Technical Review, Vol. 15 No. 8
Aug. 2017.
[10] Nico Görnitz, Marius Kloft, Konrad Rieck, and Ulf Brefeld, “Anomaly Detection for IoT Time-Series
Data: A Survey”, DOI 10.1109/JIOT.2019.2958185, IEEE Internet of Things Journal.
[11] Chandola, Varun, Arindam Banerjee, and Vipin Kumar. "Anomaly Detection: A Survey." ACM
computing surveys (CSUR) 41.3 (2009): 1-58.
[12] Mnasri, Zied, Stefano Rovetta, and Francesco Masulli. "Anomalous sound event detection: A survey
of machine learning based methods and applications." Multimedia Tools and Applications 81.4
(2022): 5537-5586.
[13] Ricciardi, Carlo, et al. "Linear Discriminant Analysis and Principal Component Analysis to Predict
Coronary Artery Disease." Health informatics journal 26.3 (2020): 2181-2192.
[14] Justin Salamon and Juan Pablo Bello, “Unsupervised Feature Learning for Urban Sound
Classification”, 2015 IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP).
[15] Peng Shi, Zhen Zhao, Huaqiang Zhong, Hangyu Shen and Lianhong Ding, “Agglomerative
Hierarchical Clustering”, Concurrency Computat Pract Exper. 2021; 33:e6077.

50
[16] Shuangfei Zhai, Yu Cheng, Weining Lu, and Zhongfei (Mark) Zhang, “Deep Structured Energy
Based Models for Anomaly Detection”, the 33rd International Conference on Machine Learning,
New York, NY, USA, 2016. JMLR: W&CP volume 48, 2016.
[17] Görnitz, Nico, Marius Kloft, Konrad Rieck, and Ulf Brefeld. "Toward Supervised Anomaly
Detection." Journal of Artificial Intelligence Research 46 (2013): 235-262.
[18] Z Yin and J Hou, “Recent advances on SVM based fault diagnosis and process monitoring in
complicated industrial processes”, Neurocomputing, 2016.
[19] Chao Dong, Chen Change Loy, and Xiaoou Tang, “Accelerating the Super-Resolution Convolutional
Neural Network”, arXiv:1608.00367v1, Aug 2016.
[20] Yong Oh Lee, Jun Jo, Jongwoon Hwang, “Application of Deep Neural Network and Generative
Adversarial Network to Industrial Maintenance: A Case Study of Induction Motor Fault Detection”,
2017 IEEE International Conference on Big Data (BIGDATA).
[21] Dan Garber, Elad Hazan, “Fast and Simple PCA via Convex Optimization”, arXiv:1509.05647v4,
25th Nov 2015.
[22] Akshay Balsubramani, Sanjoy Dasgupta, Yoav Freund, “The Fast Convergence of Incremental PCA”,
arXiv:1501.03796v1 [cs.LG] 15 Jan 2015.
[23] Hugo Vieira Neto and Ulrich Nehmzow, “Incremental PCA: An Alternative Approach for Novelty
Detection”, - Towards Autonomous Robotic Systems, 2005.
[24] Levy, Avraham, and Michael Lindenbaum. "Sequential Karhunen-Loeve basis extraction and its
application to images." Proceedings 1998 International Conference on Image Processing. ICIP98
(Cat. No. 98CB36269). Vol. 2. IEEE, 1998.
[25] David A. Ross, Jongwoo Lim, Ruei-Sung Lin, Ming-Hsuan Yang, “Incremental Learning for Robust
Visual Tracking”. Int J Comput Vis (2008) 77: 125–141, DOI 10.1007/s11263-007-0075-7.
[26] Blake, Andrew, and Michael Isard. "The Condensation Algorithm-Conditional Density Propagation
and Applications To Visual Tracking." Advances in Neural Information Processing Systems 9 (1996).
[27] Nauman Munir, Jinhyun Park, Hak-Joon Kim, Sung-Jin Song, Sung-Sik Kang, “Performance
enhancement of convolutional neural network for ultrasonic flaw classification by adopting
autoencoder”. NDT&E International 111 (2020) 102218.
[28] Burkni Palsson, Jakob Sigurdsson, Johnannes R. Sveinsson, and Magnus O. Ulfarsson,
“Hyperspectral Unmixing Using a Neural Network Autoencoder”, Digital Object Identifier
10.1109/ACCESS.2018.2818280.
[29] Min Chen, Xiaobo Shi, Yin Zhang, Di Wu, Mohsen Guizani, “Deep Feature Learning for Medical
Image Analysis with Convolutional Autoencoder Neural Network”, DOI
10.1109/TBDATA.2017.2717439, IEEE Transactions on Big Data.
[30] Chong Zhou, Randy C. Paffenroth, “Anomaly Detection with Robust Deep Autoencoders”, KDD
2017 Research Paper.
[31] Davide Chicco, Peter Sadowski, Pierre Baldi, “Deep Autoencoder Neural Networks for Gene
Ontology Annotation Predictions”, ACM-BCB 2014.
[32] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, Dahua Lin, “Unsupervised Feature Learning via Non-
Parametric Instance Discrimination”, CVPR 2018.
[33] Mayu Sakurada, Takehisa Yairi, “Anomaly Detection Using Autoencoders with Non-linear
Dimensionality Reduction”, ACM 978-1-4503-3159-3, 2014.
[34] Tara N. Sainath, Abdel-Rahman Mohamed, Brian Kingsbury, Bhuvana Ramabhadran, “Deep
Convolutional Neural Networks For LVCSR” 2013.
[35] Gu, Jiuxiang, Zhenhua Wang, Jason Kuen, Lianyang Ma, Amir Shahroudy, Bing Shuai, Ting Liu et
al. "Recent Advances in Convolutional Neural Networks." Pattern Recognition 77 (2018): 354-377.
[36] Steve Lawrence, C. Lee Giles, Ah Chung Tsoi, Andrew D. Back, “Face Recognition: A
Convolutional Neural-Network Approach” Jan 1997.
[37] Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet Classification with Deep
Convolutional Neural Networks." Advances in neural information processing systems 25 (2012).
[38] Xiaohu Zhang, Yuexian Zou, Wei Shi, “Dilated Convolution Neural Network with Leaky ReLU for
Environment Sound Classification”, 2017 22nd International Conference on Digital Signal Processing
(“DSP”), 1-5, 2017.
[39] Diederik P. Kingma, Jimmy Lei Ba, “Adam: A Method for Stochastic Optimization”, ICLR 2015.

51
[40] Dokkyun Yi, Jaehyun Ahn and Sangmin Ji, “An Effective Optimization Method for Machine
Learning Based on ADAM”, Applied Sciences, 2020.
[41] Xiao Tan, “Genetic Algorithm Based Bidirectional Generative Adversary Network for LIBOR
Prediction”, Proceedings of the 8th International Conference on Emerging Internet, Data and Web
Technologies (EIDWT 2020): Advances in Internet, Data and Web Technologies, Kitakyushu, Japan,
24-26 February 2020, p. 440-447.
[42] Jaroslav Kopčan, Ondrej Škvarek, Martin Klimo, “Anomaly detection using Autoencoders and Deep
Convolution Generative Adversarial Networks”, 14th International Scientific Conference on
Sustainable, Modern And Safe Transport.
[43] Carlos Ruiz, Myra Spiliopoulou, and Ernestina Menasalvas, “C-DBSCAN: Density-Based Clustering
with Constraints”, A. An et al. (Eds.): RSFDGrC 2007, LNAI 4482, pp. 216–223, 2007.
[44] Junhao Gan, and Yufei Tao, “On the Hardness and Approximation of Euclidean DBSCAN”, ACM
Transactions on Database Systems, Vol. 1, No. 1, Article 1, Publication date: January 2016.

INTRODUCTION OF A NOVEL ANOMALOUS SOUND DETECTION METHODOLOGY

More Related Content

Similar to INTRODUCTION OF A NOVEL ANOMALOUS SOUND DETECTION METHODOLOGY (20)

Recently uploaded (20)

INTRODUCTION OF A NOVEL ANOMALOUS SOUND DETECTION METHODOLOGY