Investigation of repeated blasts at Aitik mine using waveform cross correlation

Tensor Subspace Projection for Dimensionality Reduction. For typical image and video tensor objects although
the corresponding tensor space is of high dimensionality, tensor objects typically are embedded in a lower
dimensional tensor subspace (or manifold), in analogy to the vectorized face image embedding problem where
vector image inputs reside in a low-dimensional subspace of the original input space (M. Turk, et al 1991). Thus,
it is possible to find a tensor subspace that captures most of the variation in the input tensor objects and it can be
used to extract features for recognition and classification applications. To achieve this objective, 𝑃𝑛 < 𝐼 𝑛
orthonormal basis vectors (principle axes) of the 𝑛 - mode linear space ℝ𝐼 𝑛 are sought for each mode 𝑛 and a
tensor subspace ℝ𝐼1 ⊗ ℝ𝐼2 … ⊗ ℝ𝐼 𝑁 is formed by these linear subspaces.
The MPCA objective is to define a multilinear transformation
𝐔(𝑛)
∈ 𝑅 𝐼 𝑛 × 𝑃 𝑛 , 𝑛 = 1, … , 𝑁 that maps the original tensor space 𝑅 𝐼1 ⊗ 𝑅 𝐼2 … ⊗ 𝑅 𝐼 𝑁 into a tensor subspace
𝑅 𝑃1⊗ 𝑅 𝑃2 … ⊗ 𝑅 𝑃 𝑁(with 𝑃𝑛 < 𝐼 𝑛 for 𝑛 = 1, … 𝑁):
𝑌 𝑀 = 𝑋 𝑀 ×1 𝐔 1 𝑇
×2 𝐔 2 𝑇
… × 𝑁 𝐔 𝑁 𝑇
, 𝑚 = 1, … 𝑀 , such that
𝑌 𝑀 ∈ 𝑅 𝑃1⊗ 𝑅 𝑃2 … ⊗ 𝑅 𝑃 𝑁 , 𝑚 = 1, … 𝑀 captures most of the variations observed in the original tensor objects,
assuming that these variations are measured by the total tensor scatter.
In other words, the MPCA objective is the determination of the 𝑁 projection matrices
𝐔(𝑛) ∈ ℝ𝐼 𝑛 × 𝑃 𝑛 , 𝑛 = 1, … , 𝑁 that maximize the total tensor scatter 𝛹 𝑌:
{𝐔(𝑛)
, 𝑛 = 1, … , 𝑁} = 𝑎𝑟𝑔 𝐔 1 ,𝐔 2 ,…𝐔 𝑁 max 𝛹 𝑌
Here, the dimensionality 𝑃𝑛 for each mode is assumed to be known or predetermined (Lu, Plataniotis, and
Venetsanopoulos, 2006).
Implementation. Our processing tool uses TensorReg Toolbox as a core (see below) providing the tensor PCA or
high order SVD (HO SVD). As other tensor processing software (see Tensorlab, for example) it creates the
output structures similar to the classical PCA structures. For instance, traditional PCA notation can be used to
describe algorithm:
[PC, SCORE, LATENT] = PCA(A)
where A is a matrix with column vectors representing observations. In tensor notation it can be written as:
[MPROJ, PC, LATENT] = TPCA(𝒜,d) following the Zhou, 2013 notation,
where 𝒜 is a tensor with dimensions [p_1, p_2, …,p_D,n], d is a target dimensions [d_1,…d_D],
MPROJ is a tensor with dimensions [d_1, …d_D,n] after change of basis (analogous to SCORE for regular PCA)
and PC is a D-by-1 cell array containing principal components along each mode. PC{i} has dimension p_i-by-
d_i, and LATENT is D-by-1 cell array containing the d_i eigenvalues along each mode, ordered from largest to
smallest.
Multimodal dimensionality reduction with other methods. The multidimensional discrete (MD) cosine transform
(DCT-II and DCT-III for inverse) is popular compression structures for MPEG-4, H.264, and HEVC (high
efficiency video coding), and is accepted as the best suboptimal transformation since its performance is very
close to that of the statistically optimal Karhunen-Loeve transform
𝑋(𝑘1, 𝑘2, … , 𝑘 𝑟) = …
𝑁2−1
𝑛2=0
𝑁1−1
𝑛1=0
𝑥(𝑛1, 𝑛2, … , 𝑛 𝑟) ∙ cos
𝜋 2𝑛1 + 1 𝑘1
2𝑁1
𝑁 𝑟−1
𝑛 𝑟=0
… cos
𝜋 2𝑛 𝑟 + 1 𝑘 𝑟
2𝑁𝑟
,
where 𝑘𝑖 = 0,1, … , 𝑁𝑖 -1 and 𝑖 = 1,2, … , 𝑟. Inverse truncated MD DCT returns the reduced tensor array with the
required number of components used for cross correlation template construction.
MD variants of the various DCT types follow from the one-dimensional definitions: they are a separable product
(equivalently, a composition) of DCTs along each dimension, and computing a multidimensional DCT by
sequences of one-dimensional DCTs along each dimension is known as a row-column algorithm. The inverse of a
multi-dimensional DCT is just a separable product of the inverse(s) of the corresponding one-dimensional
DCT(s), e.g. the one-dimensional inverses applied along one dimension at a time in a row-column algorithm.
There are different methods to compute DCT (Bi, Zeng, 2004). We follow the tensor reduction procedure
implemented by Hua Zhou, 2013, in his tensor regression software (TensorReg_toolbox). We use a 4D DCT
applied to 4-order tensor composing the 3-component seismic array data.
The dimensionality reduction based on tensor interpolation is also based on the tensor regression software
by Zhou, 2013.
A TensorToolbox software by T. Kolda (Sandia National Lab) is also required for this processing.
2DPCA discussed in (Lu et al, 2006) is also tested in this work through the TensorReg as well. We are not
discussing the results in this presentation since we could not obtain good enough performance with
this method.
Testing the Reviewed Event Bulletin of the International Data Centre Using Waveform Cross Correlation:
Repeat Events at Aitik Copper Mine, Sweden
Nikita Rozhkov1, Dmitry Bobrov2, Ivan Kitov2, Mikhail Rozhkov2, Matthew Yedlin3
1SYNAPSE Science Center, 2Preparatory Commission for the Comprehensive Nuclear Test-Ban Treaty Organization, 3University of British Columbia
Abstract. The quality of the Reviewed Event Bulletin (REB) issued by the
International Data Centre (IDC) of the Comprehensive Nuclear-Test-Ban Treaty
Organization (CTBTO) is crucial for the Member States as well as for the
seismological community. One of the most efficient methods to test the REB quality
is using repeat events having very accurate absolute locations. Hundreds of quarry
blasts detonated at Aitik copper mine (the central point of active mining - 67.08N,
20.95E) were recorded by several seismic arrays of the International Monitoring
System (IMS), found by IDC automatic processing and then confirmed by analysts
as REB events. The size of the quarry is approximately 1 km and one can consider
that the uncertainty in absolute coordinates of the studied events is less than 0.5 km
as measured from the central point. In the REB, the corresponding epicenters are
almost uniformly scattered over the territory 67.0N to 67.3N, and 20.7E to 21.5E.
These REB locations are based on the measured arrival times as well as azimuth and
slowness estimates at several IMS stations with the main input from ARCES, NOA,
FINES, and HFS. The higher scattering of REB locations is caused by the
uncertainty in measurements and velocity model. Seismological methods based on
waveform cross correlation allow very accurate relative location of repeat events.
Here we test the level of similarity between signals from these events. It was found
that IMS primary array station ARCES demonstrates the highest similarity as
expressed by cross correlation coefficient (CC) and signal-to-noise ratio (SNR)
calculated at the CC traces. The small-aperture array FINES is the second best and
large-aperture array NOA demonstrating mediocre performance likely due to its size
and the loss of coherency between high-frequency and relatively low-velocity
signals from the mine. During the past five years station ARCES has been upgraded
from a vertical array to a 3-C one. This transformation has improved the
performance of CC-technique as applied to the Aitik mine events. We have also
applied a Principal Component Analysis to estimate the level of variability in the
signals as well as to build the best waveform template for effective detection and
identification of all blasts conducted at Aitik mine.
Disclaimer: The views expressed on this poster are those of the authors
and do not necessary reflect the views of the PTS CTBTO
2. Tensor Approach to Seismic Array Data Processing
1. Aitik and Kiruna mines seismicity
3. Tensor Templates Test
Preparatory Commission for the Comprehensive Nuclear-Test Ban Treaty Organization,
Provisional Technical Secretariat, Vienna International Centre, P.O. Box 1200, A-1400 Vienna,
Austria. E-mail: mikhail.rozhkov@ctbto.org
www.ctbto.org
S31A-2705
Conclusion. We have tested a design of multidimensional (high order, HO)
master event template for further seismic event location based on cross
correlation. The reason for the HO template design emerges from the ability to
represent a multidimensional seismic 3-component array as a multidimensional
array in a computer semantic sense and to apply the dimensionality reduction
technique in a manner we used to do it with the single stations (single or 3-
component). The array seismogram was considered as a 3-order tensor so the
training data set turned to 4-order tensor. Then various tensor reduction methods
were applied in order to produce the most efficient template subset for cross
correlation methods. The following methods were tested: (1) high-order SVD
(HOSVD), (2) multidimensional discreet cosine transform (MD DCT), (3) two
dimensional SVD (2D SVD), and (4) tensor interpolation. The methods were
applied to the data from mining explosions conducted at regional distances at
Sweden mines Aitik and Kiruna and recorded at ARCES 3-component seismic
array of International Monitoring System. The designed master event templates
were tested with the training set as well as with the extended set comprising 50%
of new events recorded by the same array. Different time windows and filter
bands were evaluated. Almost all methods showed very good performance most
of all in terms of detection rate showing excellent SNR for the detected signal
comparing to the preceding background noise.
This work is considered by us as a first test of applying the high order tensor
methods for seismic array data processing, and more work is needed to tune the
methods carefully and test thoroughly the performance of these methods. Also
we are working on implementation of the tensor detector instead of using a 1D
detector working with the vectorized templates. In this case a 3D cross
correlation detector must be applied (or a 2D flattening as a special case), which
requires specific correlation measures to be determined.
Introduction. In this poster, we are discussing an approach to construct master event templates for
further cross-correlation-based location with data recorded at multichannel seismic installations, such
as 3-C seismic arrays of the International Monitoring System (IMS) of the CTBTO. This approach can
be extended to other so called “vector sensor” applications such as time lapse seismics, hydraulic
fracturing monitoring, or reservoir characterization. As discussed in (Spilker1, Rozhkov2, 1984-1987),
a multicomponent seismic sample can be represented, in general case, as a hypercomplex number, a
quaternion as a special case. A pure quaternion of order 4 and other hypercomplex systems are the
natural cases of representing a 3-component digital seismogram sample requiring however special
attention to the non-commutativity or non-transitivity of the corresponding algebra. Dealing with more
complex observations within the approach of assigning the number’s dimension to the observed
physical dimension (more precisely, to the measurement linked with the dimension) demands higher
than 4 dimensional algebras, or some specific grouping of them. Dealing with the tensor
representation of seismic wavefields can simplify in certain sense the multidimensional approach to
data processing, in particular, to the data set which was the same as used for the single component
processing. Multichannel data corresponding to a seismic event from the Aitik and Kiruna quarries in
Sweden can be rearranged as a 3-mode tensor, where first mode is time, or sample number, the second
mode is station, or sensor number, and the third mode is the direction of ground motion (Z, N and E).
Then, a complete test data set would consist of a 4-mode tensor with the event number corresponding
to the 4th dimension. Considering a 3-component seismic array as a multitude of observations with a
tensor description (not the tensor field in general sense), the corresponding data tensor, formally, can
be regarded as a tensor product of 3 vector spaces, each with its own coordinate system. Then we
could apply tensor operations to the data recorded by such arrays gaining certain benefits from
utilizing joint volumetric (sensor) and spatial (array) information. Further dimensionality reduction of
tensor data produces low order principal components, a basis for the multidimensional waveform
templates. Note that a first-mode, or first-order tensor is a vector, a second-order tensor is a matrix,
and tensors of higher orders are higher-order tensors.
General approach. High-dimensional tensors or so called multi-way data are becoming more and
more popular in areas such as biomedical imaging, chemometrics and networking. Traditional
approaches to finding lower dimensional representations of tensor data include flattening the data and
applying matrix factorizations such as principal components analysis (PCA) or employing tensor
decompositions such as the CANDECOMP/PARAFAC (canonical polyadic decomposition with
parallel factor analysis) and Tucker decompositions, which may be regarded as a more
flexible PARAFAC model. Tucker decomposition, which we use in this work, decomposes a tensor
into a set of matrices and one small core tensor. Then the eigen-images can be extracted for resizing
the input tensor to lower dimensions. There are more approaches to the multimodal dimensionality
reduction such as the multidimensional Discrete Fourier Transform (DCT) mostly used in image
processing (JPEG, for instance), 2D SVD (based on low rank approximation of the matrix), and tensor
interpolation (for example, Hotz, et al, 2010, Tensor Field Reconstruction Based on Eigenvector and
Eigenvalue Interpolation). In this poster, we provide results for all the above mentioned methods with
the accent on the Tucker tensor decomposition made with the alternating least squares (ALS) method.
Math formalism. Following standard multilinear algebra, any tensor can be expressed as the product
𝒜 = 𝒮 ×1 𝑈(1) ×2 𝑈(2) … × 𝑁 𝑈(𝑁)
where 𝑈(𝑛)
= (𝑈1
𝑛
𝑈2
𝑛
… 𝑈𝐼 𝑛
(𝑛)
) is an orthogonal (𝐼 𝑛 × 𝐼 𝑛) matrix (H. Lu, K.N. Plataniotis, and
A.N. Venetsanopoulos (2006), Multilinear principal component analysis for tensor objects for
classification). A visual representation of this decomposition in the third-order case is shown on next
figure:
A matrix representation of this decomposition can be obtained by unfolding 𝒜 and 𝒮 as:
𝐀(𝑛) = 𝐔(𝑛) ∙ 𝐒 𝑛 ∙ (𝐔 𝑛+1 ⊗ 𝐔 𝑛+2 … ⊗ 𝐔 𝑁 ⊗ 𝐔 1 ⊗ 𝐔 2 ⊗ ⋯ ⊗ 𝐔 𝑛−1 ) 𝑇
where ⊗ denotes the Kronecker product and 𝒮 is a core tensor of size 𝑅1 × 𝑅2 × ・・・ × 𝑅 𝑁 .The
decomposition can also be written as
𝒜 = …
𝐼2
𝑖2=1
𝐼1
𝑖1=1
𝒮(𝑖1 𝑖2, … , 𝑖 𝑁 ) ×
𝐼 𝑁
𝑖 𝑁=1
𝐮𝑖1
1
° 𝐮𝑖2
2
° … °𝐮𝑖 𝑁
𝑁
i.e., any tensor can be written as a linear combination of 𝐼1 × 𝐼2 × ⋯ × 𝐼 𝑁 rank-1 tensors.
This decomposition is used in the following to formulate a multilinear projection for
dimensionality reduction.
______________________________________________________________________________
1. A suite of papers in Doklady Academii Nauk SSSR, including (1) On commutative
hypercomplex system of order four, 1984, and (2) Some spectral properties of
hypercomplex wavefields representations, 1987.
2. Homomorphic processing in geophysical distributed systems, 1986. Dissertation Thesis.
We have conducted a study on template selection for further location
of seismic events by the waveform cross-correlation method using
regional quarry blast data recorded at 3-component IMS seismic array
ARCES. One of reasons for this study was the difference in location
conducted by the IDC and presented in the REB, and the location
results presented by the International Seismological Center, ISC,
based on data from the seismographic network of the University of
Helsinki (UH, inlay on a Google Earth map below). The blast clusters
by the IDC are white balls for both mines; the UH results are blue
clouds.
In total, 122 seismic events from the Aitik (copper) and Kiruna (iron)
mines were studied at 19 3-component stations of ARCES.
Aitik mine, aerial view
Seismic event (blast at
Aitik mine) recorded
at ARCES. Stations
with only vertical
components were not
used in the analysis.
W
𝒮U
V
≈
𝒜
We have conducted a number of tests with the obtained reduced tensor components in order to evaluate
performance of the cross correlation detector with these components used as templates. Complete sets of
components for the 3C array were produced with 4 methods: tensor interpolation, DCT, 2DSVD and
HOSVD. In this study, the 3D templates were reduced to 1D vectorized case ([Z,NS,WE]) and we applied a
well established system of tests. Selected results are presented on figures below. The first test was based on
cross correlation (CC) of the developed templates with continuous waveforms measured from the set of 122
events and determining the detection rate based on SNR threshold (see poster S51A-2758 for details, also
Bobrov, et al. 2012): the percentage of detections having SNRCC>3.5. Then, we tested 46 events not
included into the training set of 122. In this test, the ten first reduced tensor components were used for
detection, i.e. 40 components were tested altogether as presented on figures below: (1) 2D SVD, (2) MD
DCT, (3) HOSVD, and (4) tensor interpolation. Different template lengths were used from 10 to 30 seconds.
The best results were produced by the MD DCT and interpolation templates for average SNR; the minimum
SNRCC over all tested signals for a given template, which has to be above 3.5; and the average CC. The
overall difference in detection rates is not large. Similar tests were also carried out with the eigenimages
(instead of reduced back-projections) produced for the HOSVD (PC cell array in TPCA algorithm). It was
found that the algorithm destroys the proper channel alignment in sensor triads and move-outs related to
different stations of the array and the test results were not impressive. In case of single 3-C station, it does
not make any difference since all the channels in a training set are aligned by default and there no is need to
keep the move-outs so the regular SVD/PCA case works fine.
In short, the algorithm works like this: given n tensor
observations (number of seismic events in our case) the
program flattens each along mode I to obtain a matrix,
take the outer product of the matrix (p_i-by-p_i). Then do
the classical PCA on the sum of outer products, retrieve
the first d_i principal components (p_i-by-d_i) and then
scale the original tensor in the PC basis along each mode.
Following this procedure we were able to produce 3C
array projected components of the reduced tensor
composed of 57 seismic channel (19 stations x [N, EW,
NS]), corresponding to the tested reduction to the
dimension of 1 to n. A figure on a left is a snapshot of
one out of ten 3D components of the reduced tensor
seismogram. A reduced tensor set (3D for single
component and 4D for more components) inherits the
properties of the input training data set.
After all methods were tested with different template lengths and filters, we have taken the window (10 seconds) and the filter (3-6 Hz), which work the best. Then we returned to the high order
tensor decomposition and tried our own truncated reconstruction of the reduced tensor set. The detection rate was the same as for the best methods above, and the smallest SNRCC was even larger
(4.2 against 3.5). More work has to be conducted to find the optimal multidimensional template design.
Left figure:
4 methods, 10
templates in each
method, variable
time window length
Left figure:
2 methods, 10
templates in each
method, 5 filter bands:
2-4 Hz, 4-8 Hz, 3-6
Hz, 6-12 Hz, 8-16 Hz.
Left figure:
HOSVD tests non-
standard truncated
reconstruction of
the reduced tensor
set
3-component records at a single ARCES station. Bandpass filter is 3-6 Hz.
160 (upper figure) and 27 (lower) second time windows presented.

Investigation of repeated blasts at Aitik mine using waveform cross correlation

More Related Content

What's hot (19)

Similar to Investigation of repeated blasts at Aitik mine using waveform cross correlation (20)

More from Ivan Kitov (20)

Recently uploaded (20)

Investigation of repeated blasts at Aitik mine using waveform cross correlation