Microseismic signal segmentation method based on hidden semi-Markov model

Abstract

The application discloses a microseismic signal segmentation method based on a hidden semi-Markov model, relates to the technical field of microseismic monitoring in coal mining, and solves the technical problem that the whole P wave of a microseismic signal is not picked up accurately enough. The technical solution is that a hidden semi-Markov model is used to model the microseismic signal, whether a microseismic event occurs in the signal is regarded as a state conversion process in the HSMM, and the state residence time is considered. The HSMM is trained using the characteristic sequence of the microseismic signal. The trained HSMM can automatically learn the amplitude variation characteristics of the signal when a microseismic event occurs or not, so that the initial time and the duration of the microseismic event can be accurately captured, and thus the segmentation of the microseismic signal is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal mining microseismic monitoring, and particularly relates to a microseismic signal segmentation method based on a hidden semi-Markov model. BACKGROUND

[0002] In the process of coal mining, many possible disasters such as mine microseismic, rock burst, roof disaster and gas outburst often occur, which seriously affect the safe mining of coal mines. In order to reduce the occurrence of disasters and give timely warnings, microseismic monitoring technology is widely used in coal mining. By collecting and analyzing the elastic waves emitted when the surrounding rock deforms or is damaged, i.e. microseismic signals, it plays a key role in preventing and warning coal and rock dynamic disasters. Microseismic positioning is a key content in microseismic monitoring technology, but due to the complexity of the mining environment in coal mines, there are many factors affecting the accuracy of microseismic positioning, one of which is the picking of the P-wave first arrival of microseismic signals. The accuracy of the P-wave first arrival picking is the key to microseismic positioning and even the prediction and warning of coal and rock dynamic disasters.

[0003] Although microseismic signal picking has been widely concerned and researched, most of the researches are only for the picking of the first arrival time, and there are few researches on the picking of the start and end times, and the overall picking is the segmentation of the microseismic signal. In a complex working environment, the data collected by microseismic monitoring often contains not only rock mass fracture signals, but also blasting signals, electromagnetic interference signals and drill rig drilling signals. In order to effectively realize the subsequent work such as identification, classification and analysis of microseismic signals, the segmentation of microseismic signals is the first prerequisite. SUMMARY

[0004] The present application provides a microseismic signal segmentation method based on a hidden semi-Markov model, and the technical purpose is to pick the end time of the microseismic signal while ensuring the accuracy of the P-wave first arrival picking of the microseismic signal, so as to realize the segmentation of the microseismic signal.

[0005] The above technical purpose of the present application is realized by the following technical scheme:

[0006] A microseismic signal segmentation method based on a hidden semi-Markov model, comprising:

[0007] S1: collecting an original microseismic signal X and pre-processing it to obtain a pre-processed microseismic signal X'; wherein X={x1, x2,..., x T};

[0008] S2: extracting features of the microseismic signal X' to obtain a feature sequence E A ' of the microseismic signal X';

[0009] S3: inputting the feature sequence E AThe feature sequence set is divided into a training set and a test set;

[0010] S4: modeling the microseismic signal X' using a hidden semi-Markov model, including: each time in the microseismic signal X' corresponds to a state, thereby forming a hidden state sequence I = {i1, i2,..., i r} in the hidden semi-Markov model, the feature sequence E A of the microseismic signal X' as an observation sequence O = {o1, o2,..., o T} in the hidden semi-Markov model, then the hidden semi-Markov model of the microseismic signal X' is λ = (π, A, B, P), expressed as:

[0011]

[0012] wherein π represents an initial state probability vector, π i,d represents the probability of the initial state being q i and the duration being d time units; A represents a state transition probability matrix, a (i,d')(j,d) represents the probability of the state being q i in the current d time units under the condition that the state is q j in the previous d' time units; B represents an observation probability matrix, b j,d (O t+1:t+d ) represents the probability of the observation sequence being O [t+1:t+d] generated by i j under the condition that the state of i t+1:t+d is q i ; P represents a state residence time probability matrix, p i (d) represents the probability of the state q

[0013] S5: initializing the parameters of λ, including: performing statistics on the initial parameters of λ through the feature sequence E A in the training set;

[0014] S6: estimating the optimal parameters and obtaining the model represented by the optimal parameters as λ * = (π, A, B, P);

[0015] S7: taking the feature sequence in the test set as the observation sequence of the model λ * , estimating the corresponding hidden state sequence in the observation sequence based on the Viterbi algorithm, and obtaining the estimated hidden state sequence I' = {i1, i2,..., i T};

[0016] S8: according to the estimated hidden state sequence I' = {i1, i2,..., i TThe arrival and end times of microseismic events are obtained, ultimately enabling the segmentation of microseismic signals.

[0017] The beneficial effects of this application are as follows: It uses a Hidden Semi-Markov Model (HSMM) to model microseismic signals, treating the presence or absence of microseismic events in the signal as a state transition process within the HSMM, and considering the state dwell time. The HSMM is trained using the feature sequences of the microseismic signal. The trained HSMM can automatically learn the amplitude variation characteristics of the signal when microseismic events occur, thus accurately capturing the initial moment of microseismic events and their duration, thereby achieving microseismic signal segmentation. Attached Figure Description

[0018] Figure 1 The flowchart shows the microseismic signal segmentation method based on the hidden semi-Markov model.

[0019] Figure 2 This is a schematic diagram comparing the microseismic signals before and after preprocessing.

[0020] Figure 3 A schematic diagram of the normalized Shannon energy envelope of the microseismic signal;

[0021] Figure 4 A schematic diagram of the hidden state sequence estimated using the Viterbi algorithm for a given characteristic sequence of a microseismic signal;

[0022] Figure 5 A schematic diagram comparing the P-wave first arrival pickup results of different methods with the method described in this application;

[0023] Figure 6 This is a schematic diagram of the segmentation results for a single microseismic event.

[0024] Figure 7 This is a schematic diagram showing the segmentation results of multiple microseismic events. Detailed Implementation

[0025] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0026] like Figure 1 As shown, the microseismic signal segmentation method based on the hidden semi-Markov model described in this application includes:

[0027] S1: Acquire the raw microseismic signal X and preprocess it to obtain the preprocessed microseismic signal X'; where X = {x1, x2, ..., x...} T}

[0028] Specifically, a real microseismic signal from a mine is collected, i.e., the original microseismic signal X. Variational mode decomposition is used to decompose X into K modal components. High-frequency noise components in the K modal components are removed by spectral analysis. The remaining modal components are added together to reconstruct the denoised microseismic signal X. denoise For the denoised micro-vibration signal X denoise After amplitude normalization, the preprocessed microseismic signal X' is obtained, expressed as: X' = X denoise / max(|X denoise ). Figure 2 The comparison of microseismic signals before and after preprocessing is shown.

[0029] When using variational mode decomposition to denoise the original microseismic signal, a genetic algorithm is used, with the number of decomposed modes K and the penalty coefficient α as chromosomes, and the sample entropy of the decomposed modal components as the fitness function, to iteratively find the optimal parameters K and α. Based on the characteristics of the microseismic signal, modal components with frequencies greater than 500 Hz (i.e., high-frequency noise components) are removed, and the remaining modal components are added together to reconstruct the denoised microseismic signal X. denoise .

[0030] S2: Perform feature extraction on the microseismic signal X' to obtain the feature sequence E of the microseismic signal X'. A '.

[0031] Specifically, step S2 includes:

[0032] S21: The Shannon energy of the microseismic signal X' is calculated and expressed as: E = -X' 2 log(X' 2 );

[0033] S22: Shannon's energy envelope E A The calculation is performed and expressed as follows: Where N represents the length of the sliding window, N = 20, and k represents the index of the sample point within the sliding window;

[0034] S23: Shannon's energy envelope E A Perform normalization, such as Figure 3 As shown, the normalized Shannon energy envelope E is obtained. A ', E A 'This is the characteristic sequence of the microseismic signal X'; where E A '=E A / max(|E A |).

[0035] S3: E represents the characteristic sequence of multiple microseismic signals. AThe feature sequence set is divided into a training set and a test set, the training set is used to estimate the parameters of the hidden semi-Markov model (HSMM), and the test set is used to verify the performance of the model.

[0036] S4: modeling the microseismic signal X' using a hidden semi-Markov model, including: each time point in the microseismic signal X' corresponds to a state, thereby forming a hidden state sequence I = {i1, i2,..., i r} in the hidden semi-Markov model, the feature sequence E A ' of the microseismic signal X' is taken as an observation sequence O = {o1, o2,..., o T} in the hidden semi-Markov model, and the hidden semi-Markov model of the microseismic signal X' is λ = (π, A, B, P), which is expressed as:

[0037]

[0038] wherein π represents an initial state probability vector, π i,d represents the probability that the initial state is q i and the duration is d time units; A represents a state transition probability matrix, a (i,d')(j,d) represents the probability that the state is q i in the current d time units, given that the state is q j in the previous d' time units; B represents an observation probability matrix, b j,d (O t+1:t+d ) represents the probability that the observation sequence is O [t+1:t+d] , given that the state sequence i j is q t+1:t+d ; P represents a state residence time probability matrix, p i (d) represents the probability that the state q i lasts for d time units.

[0039] Specifically, when the microseismic signal X' is modeled using the hidden semi-Markov model, each time point in the microseismic signal X' corresponds to a state, wherein the sample points when no microseismic event occurs correspond to state 0, and the sample points when a microseismic event occurs correspond to state 1, thereby forming a hidden state sequence I = {i1, i2,..., i r} in the hidden semi-Markov model, let Q = {q1, q2,..., q N} be the set of all possible states, then q1 = 0, q2 = 1, and N = 2; the feature sequence E A ' corresponding to the microseismic signal X' is taken as an observation state sequence, let V = {v1, v2,..., v M} be the set of all possible observation values, and M represents the number of observation values, then v M∈[0,1], that is, the observed value is the normalized Shannon energy envelope E. A The amplitude of '.

[0040] S5: Initialize the parameters of λ, including: using the feature sequence E from the training set. A 'Statistics are performed on the initial parameters of λ.'

[0041] Specifically, let the initial state probability vector be... That is, the initial occurrence probability of states 0 and 1 is equal, indicating that at the beginning of the microseismic signal, there may be a microseismic event or no microseismic event.

[0042] Let the state transition probability matrix be... That is, state self-transition is not considered.

[0043] Suppose that the observation probability matrix B consists of two Gaussian distributions with different parameters: the observation probability corresponding to state 0 follows a Gaussian distribution with mean μ0 and standard deviation σ0, and the observation probability corresponding to state 1 follows a Gaussian distribution with mean μ1 and standard deviation σ1. The initial parameters of the observation probability matrix are obtained by statistically analyzing the mean and variance of different states in the feature sequence of the training set: μ0 = 0.012, σ0 = 0.006, μ1 = 0.276, σ1 = 0.316.

[0044] Let the state dwell time probability matrix P follow a multinomial distribution, i.e., p i (d)~PN(p i1 ,p i2 ,...,p id ,...,p iL ), Where, p id This represents the probability that the duration of the i-th state is d time units.

[0045] S6: Transfer the feature sequences E from the training set A As the observation sequence of model λ, the model parameters that maximize the probability of the current observation sequence are estimated based on the EM algorithm. After iterative estimation, until convergence or the maximum number of iterations is reached, the optimal parameters of the model are obtained. The model corresponding to the optimal parameters is represented as λ. * = (π, A, B, P).

[0046] S7: Use the feature sequences in the test set as the model λ * Given the observed sequence, the hidden state sequence corresponding to the observed sequence is estimated using the Viterbi algorithm, resulting in the estimated hidden state sequence I'={i1,i2,...,i...}. T The estimation results are as follows: Figure 4 As shown.

[0047] S8: obtaining the start time and end time of the microseismic event according to the estimated hidden state sequence I' = {i1, i2,..., i T event, and finally achieving the segmentation of the microseismic signal.

[0048] Specifically, in the estimated hidden state sequence I' = {i1, i2,..., i T , state 0 represents that no microseismic event occurs at this time in the microseismic signal, and state 1 represents that a microseismic event occurs. The first sample point with state 1 is the start time of the microseismic event, and the last sample point with state 1 is the end time of the microseismic event, thereby achieving the segmentation of the microseismic signal.

[0049] To show the accuracy of the method in picking up P waves of the microseismic signal, Figure 5 the comparison results between manual picking and STA / LTA method and the method described in the present application are shown in the schematic diagram, from which Figure 5 it can be seen that the picking accuracy of HSMM is closer to the manual picking result. Figure 6 、 Figure 7 The ability of the method to segment single microseismic event and multiple microseismic events is shown.

[0050] The above is the exemplary embodiment of the present application, and the protection scope of the present application is defined by the claims and their equivalents.

Claims

1. A microseismic signal segmentation method based on a hidden semi-Markov model, characterized in that, Comprising: S1: collect the original microseismic signal X and pre-process it to obtain the pre-processed microseismic signal X'; wherein X = {x1, x2,..., xN}; X' = {x1', x2',..., xN'}; T} S2: feature extraction is performed on the microseismic signal X' to obtain a feature sequence E of the microseismic signal X' A '; S3: E represents the characteristic sequence of multiple microseismic signals. A The feature sequence set is divided into a training set and a test set; S4: modeling the microseismic signal X' using a hidden semi-Markov model, including: each time in the microseismic signal X' corresponds to a state, thereby constituting a hidden state sequence I = {i1, i2,..., i r} in the hidden semi-Markov model, the feature sequence E A ' of the microseismic signal X' as an observation sequence O = {o1, o2,..., o T} in the hidden semi-Markov model, then the hidden semi-Markov model of the microseismic signal X' is λ = (π, A, B, P), expressed as: where π denotes the initial state probability vector, π i,d denotes the probability of the initial state being q i and the duration being d time units; A denotes the state transition probability matrix, a (i,d')(j,d) denotes the probability of the state being q i in the current d time units given that the state was q j in the previous d' time units; B denotes the observation probability matrix, b j,d (O t+1:t+d ) denotes the probability of the observation sequence O [t+1:t+d] being generated given that the state sequence i j was q t+1:t+d ; P denotes the state duration probability matrix, p i (d) denotes the probability of the state q i lasting d time units. S5: initialize the parameters of λ, including: through the feature sequence E in the training set A statistics on the initial parameters of λ; S6: Estimate the optimal parameters and obtain the model representation corresponding to the optimal parameters as λ * = (π, A, B, P); S7: take the feature sequence in the test set as the observation sequence of the model λ * , estimate the corresponding hidden state sequence in the observation sequence based on the Viterbi algorithm, and obtain the estimated hidden state sequence I' = {i1, i2,..., i T} S8: According to the estimated hidden state sequence I' = {i1, i2,..., i T}obtain the start time and end time of the microseismic event, and finally realize the segmentation of the microseismic signal.

2. The method of claim 1, wherein, In step S1, the original microseismic signal X is preprocessed, including: decomposing X into K modal components by using variational modal decomposition, removing high-frequency noise components in the K modal components through spectrum analysis, and adding the remaining modal components to reconstruct to obtain a denoised microseismic signal X denoise , which is represented as: X' = X denoise . denoise The amplitude of the denoised microseismic signal X denoise is normalized to obtain a preprocessed microseismic signal X', which is represented as: X' = X denoise / max(|X denoise |).

3. The method of claim 1, wherein, Step S2 comprises: S21: Calculate the Shannon energy of the microseismic signal X', denoted as: E = -X' 2 log(X' 2 ); S22: calculating the Shannon energy envelope E A The calculation is expressed as: where N represents the length of the sliding window, N = 20, and k represents the sample point index within the sliding window. S23: normalizing the Shannon energy envelope E A to obtain a normalized Shannon energy envelope E A ', E A ' is a characteristic sequence of the microseismic signal X' ; wherein E A ' = E A / max (|E A |).

4. The method of claim 1, wherein, In step S4, the microseismic signal X' is modeled using a hidden semi-Markov model, each time of the microseismic signal X' corresponds to a state, wherein the sample point when no microseismic event occurs corresponds to state 0, and the sample point when a microseismic event occurs corresponds to state 1, thereby forming a hidden state sequence I = {i1, i2,..., i r} in the hidden semi-Markov model. N} is the set of all possible states, then q1 = 0, q2 = 1, and N = 2; the feature sequence E A ' corresponding to the microseismic signal X' is regarded as an observed state sequence, V = {v1, v2,..., v M} is the set of all possible observation values, and M represents the number of observation values, then v M ∈ [0, 1], that is, the observation value is the amplitude of the normalized Shannon energy envelope E A '.

5. The method of claim 1, wherein, In step S5, the parameters of λ are initialized, comprising: Let the initial state probability vector be i.e. the initial occurrence probabilities of states 0 and 1 are equal; Let state transition probability matrix i.e. no self-transition is considered; Suppose the observation probability matrix B is composed of two Gaussian distributions with different parameters, the observation probability corresponding to state 0 obeys a Gaussian distribution with a mean of μ0 and a standard deviation of σ0, and the observation probability corresponding to state 1 obeys a Gaussian distribution with a mean of μ1 and a standard deviation of σ1; Let the state residence time probability matrix P follow a multinomial distribution, i.e., p i (d) ~ PN(p i1 , p i2 ,..., p id ,..., p iL ), where p id denotes the probability of the ith state duration being d time units.

6. The method of claim 5, wherein, By training the feature sequence E in the training set A The initial parameters for λ are statistically obtained as μ0= 0.012, σ0= 0.006, μ1= 0.276, σ1= 0.

316.

7. The method of claim 1, wherein, Step S6 includes: taking the feature sequence E A As the observation sequence of the model λ, the model parameters that make the probability of the current observation sequence maximum are estimated based on the EM algorithm, and through iterative estimation, the optimal parameters of the model are obtained until convergence or the maximum iteration number is reached.

8. The method of claim 1, wherein, In step S8, the hidden state sequence I' = {i1, i2,..., iN} is estimated, where state 0 indicates that no microseismic event occurs at this moment in the microseismic signal, and state 1 indicates that a microseismic event occurs. The first sample point with state 1 is the initial time of the microseismic event, and the last sample point with state 1 is the end time of the microseismic event, so as to realize segmentation of the microseismic signal. T} In step S8, the hidden state sequence I' = {i1, i2,..., iN} is estimated, where state 0 indicates that no microseismic event occurs at this moment in the microseismic signal, and state 1 indicates that a microseismic event occurs. The first sample point with state 1 is the initial time of the microseismic event, and the last sample point with state 1 is the end time of the microseismic event,

Images