Coal mine gas and coal dust explosion sound recognition method and system

Abstract

The invention provides a coal mine gas and coal dust explosion sound recognition method and system, and the method comprises the steps: carrying out the discrete wavelet transformation of a noisy mine sound signal, obtaining different sub-bands, estimating a denoising threshold value through the estimation relation of a noise coefficient, carrying out the calculation of the sub-bands obtained through decomposition, carrying out the goodness-of-fit test statistics, obtaining the statistical magnitude, and carrying out the recognition of the coal mine gas and coal dust explosion sound. And comparing the statistical magnitude with a threshold value so as to realize de-noising. Preprocessing and feature vector extraction are carried out on the denoised signal training sample, then the feature vector is input into an SVM model and an LSSVM model at the same time, and a recognition model is obtained through training; after a signal test sample is subjected to preprocessing and feature vector extraction, a feature value matrix is input into an SVM model and an LSSVM model, result labels of the SVM model and the LSSVM model are obtained, judgment fusion is carried out on identification result label results, and finally an explosion sound identification result is obtained. According to the method, the accuracy of sound signal recognition can be improved, and the method has relatively high compatibility and portability.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of signal processing technology, and in particular to a method and system for identifying coal mine gas and coal dust explosion sounds. Background Art

[0002] Mine explosion is a serious disaster in coal mines, mainly including coal mine gas and coal dust explosion. In coal mine production, mine explosion alarms should be timely and accurate. At present, explosion alarms are mainly based on sensors and precursor phenomena. Sensor monitoring can prevent underground explosion accidents, but due to the possible poor on-site working environment, poor management and mental paralysis of personnel, sensors cannot completely prevent the occurrence of explosions, let alone alarm for sudden mine explosions; precursor phenomena are mainly based on human experience, and there are relatively large objective factors, which are prone to false alarms and missed alarms, so that emergency measures cannot be taken in time, which is easy to cause explosions and casualties.

[0003] Therefore, in order to effectively reduce the property losses and casualties caused by explosion disasters, a new mine explosion disaster alarm identification method is needed to accurately sense underground explosion disasters and gain precious time for disaster relief and escape. Summary of the invention

[0004] The embodiment of the present application provides a method and system for identifying the sound of coal mine gas and coal dust explosions. The method can improve the accuracy of sound signal recognition and has strong compatibility and portability.

[0005] In a first aspect, an embodiment of the present application provides a method for identifying coal mine gas and coal dust explosion sounds, comprising:

[0006] Collecting sound signals in the target area through sound collection equipment;

[0007] The sound signal is transformed by discrete wavelet transform to obtain different sub-bands, and the different sub-bands are subjected to fit test statistics to obtain test statistics; a denoising threshold is obtained by using an estimated relationship of a noise coefficient; the test statistic is compared with the denoising threshold, and an expected signal coefficient is retained, where the expected signal coefficient corresponds to a target sound signal;

[0008] Preprocessing the training samples of the target sound signal and extracting training feature vectors, wherein the training feature vectors are respectively input into the SVM model and the LSSVM model for training to obtain a first recognition model corresponding to the SVM and a second recognition model corresponding to the LSSVM;

[0009] Preprocessing the test sample of the target sound signal and extracting a test feature vector, wherein the test feature vector is used to input the first recognition model and the second recognition model respectively to obtain a result label;

[0010] According to the result label, the category of the target sound is determined.

[0011] In a possible implementation, the sound signal is transformed through a discrete wavelet transform to obtain different sub-bands, and the different sub-bands are subjected to a goodness-of-fit test to obtain a test statistic; a denoising threshold is obtained using an estimated relationship of a noise coefficient; the test statistic is compared with the denoising threshold, and an expected signal coefficient is retained, where the expected signal coefficient corresponds to a target sound signal, including:

[0012] Normalizing the discrete wavelet coefficients to determine the discrete wavelet coefficients, and determining the threshold of the false alarm probability through the fit test;

[0013] Performing a fit test on the discrete wavelet coefficient and the threshold of the false alarm probability to obtain a threshold coefficient;

[0014] The statistical distance between the empirical distribution function of discrete wavelet coefficients and the cumulative distribution function of simulated WGN coefficients is calculated to obtain the test statistic;

[0015] Comparing the test statistic with the threshold coefficient, retaining the coefficient of the expected signal, and discarding the coefficient of the noise signal;

[0016] Through inverse discrete wavelet transform, the expected signal coefficients are outputted to correspond to the target sound signal.

[0017] In a possible implementation, determining the category of the target sound according to the result label includes:

[0018] The result label includes a first result label corresponding to the first recognition model and a second result label corresponding to the second recognition model; when the results of the identification of the first result label and the second result label are both explosion sound signals, the target sound is determined to be an explosion sound.

[0019] In a possible implementation manner, preprocessing the target sound signal includes signal anomaly detection, and the signal anomaly detection includes:

[0020] When the adjacent signal is greater than the set threshold, it is determined that the local peak value change of the regional signal is abnormal.

[0021] In a possible implementation, processing the test sample by using the SVM model includes:

[0022] Obtaining a feature vector representing the target sound signal through wavelet transform, and inputting the feature vector into a SVM classifier for training;

[0023] Establish a linear SVM classifier according to the SVM model to find the optimal hyperplane;

[0024] The target sound signal is divided into explosion sound and non-explosion sound according to the SVM classifier.

[0025] In one possible implementation, the training of the classifier of the SVM model is obtained through a first hyperparameter, wherein the first hyperparameter includes a regularization coefficient and a gamma kernel parameter, wherein the regularization coefficient is used to characterize the tolerance of the SVM model to errors, and the gamma kernel parameter is used to characterize the Gaussian range of action corresponding to the support vector.

[0026] In a possible implementation, training the test sample by using the LSSVM model includes:

[0027] Obtain a feature vector representing the target sound signal through wavelet transform, and input the feature vector into the LSSVM classifier for training;

[0028] Establishing a classifier of the model through the decision function of the LSSVM model;

[0029] The target sound signal is classified into explosion sound and non-explosion sound according to the classifier.

[0030] In one possible implementation, the training of the LSSVM classifier is obtained by optimizing the second hyperparameter by the whale optimization algorithm, wherein the second hyperparameter includes a regularization coefficient and a gamma kernel parameter, the regularization coefficient is used to characterize the tolerance of the LSSVM model to errors, and the gamma kernel parameter is used to characterize the Gaussian range of action corresponding to the support vector.

[0031] In a second aspect, an embodiment of the present application provides a coal mine gas and coal dust explosion sound recognition system, comprising:

[0032] A collection module, used to collect sound signals in a target area through sound collection equipment;

[0033] A denoising module is used to obtain different sub-bands by discrete wavelet transform of the sound signal, perform fit test statistics on the different sub-bands to obtain a test statistic; obtain a denoising threshold by using an estimated relationship of a noise coefficient; compare the test statistic with the denoising threshold, and retain an expected signal coefficient, wherein the expected signal coefficient corresponds to a target sound signal;

[0034] A training module, used for preprocessing the training samples of the target sound signal and extracting training feature vectors, wherein the training feature vectors are respectively input into the SVM model and the LSSVM model for training to obtain a first recognition model corresponding to the SVM and a second recognition model corresponding to the LSSVM;

[0035] A testing module, used for preprocessing the test sample of the target sound signal and extracting a test feature vector, wherein the test feature vector is used for inputting the first recognition model and the second recognition model respectively to obtain a result label;

[0036] A classification module is used to determine the category of the target sound according to the result label.

[0037] In a possible implementation, the system is used to implement the method as described in any one of the embodiments of the first aspect.

[0038] In summary, the coal mine gas and coal dust explosion sound recognition method and system provided by the embodiment of the present application can effectively remove noise from the mine sound signal and maintain the detail information and edge retention capability of the sound signal. When recognizing the denoised sound signal, the accuracy of sound signal recognition can be improved, and it has strong compatibility and portability. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0041] Figure 1 A schematic diagram of a method for identifying the sound of coal mine gas and coal dust explosions provided in an embodiment of the present application;

[0042] Figure 2 Figures ad show the output results obtained by different methods of adding noise with a signal-to-noise ratio of -10dB;

[0043] Figure 3 A schematic diagram of output result classification obtained according to an embodiment of the present application;

[0044] Figure 4 A schematic diagram of a coal mine gas and coal dust explosion sound recognition system provided in an embodiment of the present application;

[0045] Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0046] When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments are only examples of devices and methods consistent with some aspects of the embodiments of the present application as detailed in the appended claims.

[0047] The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments. The following will describe the embodiments of the present application in conjunction with the accompanying drawings.

[0048] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0049] The following first describes in detail the relevant terms involved in the embodiments of the present application.

[0050] (1) Discrete Wavelet Transform

[0051] Discrete Wavelet Transformation (DWT) discretizes the scale and translation of the basic wavelet. It is a signal decomposition method based on wavelet analysis. It decomposes the signal into wavelet components of different frequencies and extracts the local time-frequency information of the signal. Therefore, it can accurately represent the frequency domain characteristics of the signal, and also has good local characteristics, and can well preserve the detailed characteristics of the signal. Its calculation formula is:

[0052]

[0053] In the formula, m is the frequency, n is the actual value, and a=2 m is the scale factor, b = 2 m n is the displacement factor.

[0054] Select appropriate wavelet transform and then analyze the sound signal through multi-resolution to get the high-frequency part. and the low frequency part ψ m,n (t), that is

[0055]

[0056] (2) Mahalanobis distance

[0057] Mahalanobis distance (MD) is used to represent the distance between a sample and a distribution. It is defined as: Let sample x i ∈R(i=1,2,…,m), if X represents an m×p input matrix, then the sample x i and x j The Mahalanobis distance D(x i ,x j ) is calculated as follows:

[0058] D(x i ,x j )=(x i -x j )Σ -1 (x i -x j ) T

[0059] Where Σ is a real symmetric matrix.

[0060] As can be seen from the above formula, the Mahalanobis distance is a weighted distance. Based on the sample data, different attributes of the data are weighted. The Mahalanobis distance has nothing to do with the correlation of the sample, and the unit of the sample does not affect its calculation. Compared with the Euclidean distance calculation method, it is closer to the characteristics of the sample.

[0061] (3) Goodness of fit test

[0062] Goodness of fit (GOF) is a method used to test whether a sample fits a given distribution or model. It measures the difference between the observed value and the theoretical distribution by calculating the test statistic and comparing it with the critical value to determine whether to reject the null hypothesis.

[0063] (4) Anderson-Darling test

[0064] The AD test is a goodness-of-fit test method based on weighted gaps, which is used to test whether the observed data conforms to a given theoretical distribution. It differs from other goodness-of-fit test methods in that the AD test is more sensitive to the fitting of tail data. The basic idea of ​​the AD test is to arrange the sample data by size, and calculate their proportion in the sample and the value of the cumulative distribution function corresponding to them in the theoretical distribution. Then, the gap between the two is weighted to obtain the test statistic. The weighting method depends on the location of the data, and the data close to the tail will be given a larger weight.

[0065] The Anderson-Darling (AD) test statistic is defined as follows:

[0066]

[0067] Where F(t) represents the empirical distribution function of the input sample z with support t, F r (t) represents the assumed cumulative distribution function corresponding to the probability density function, and ω(t) is the distribution function F r The weight function of (t) is calculated as follows:

[0068] ω(t)=(F r (t)(1-F r (t))) -1

[0069] In actual calculation process, it is better to abbreviate it as:

[0070] τ AD =-LQ

[0071] Where L is the size of the given observation value, and the calculation formula of Q is as follows:

[0072]

[0073] In the formula, F r (z t ) is the value of z t , which corresponds to the assumed cumulative distribution function of the probability density function.

[0074] In the goodness-of-fit test framework, the probability of incorrectly rejecting a candidate distribution is called the false alarm probability P. fa , which is specifically defined as:

[0075]

[0076] Where H0 represents the null hypothesis corresponding to the noise detection.

[0077] Mine explosions mainly include coal mine gas and coal dust explosions, which are one of the causes of major disasters in coal mines. Therefore, in order to effectively reduce the property losses and casualties caused by explosion disasters, the embodiment of the present application provides a coal mine gas and coal dust explosion sound recognition method and system, which can improve the accuracy of sound signal recognition, has strong compatibility and portability, and can accurately perceive underground explosion disasters to gain precious time for disaster relief and escape.

[0078] Figure 1 A method for identifying the sound of coal mine gas and coal dust explosions is provided for the embodiment of the present application. Figure 1The method comprises steps S101 to S105, wherein:

[0079] Step S101: Collect sound signals in a target area through sound collection equipment.

[0080] In the embodiment of the present application, the sound collection equipment (module) is installed in the important monitoring area underground, and is used to collect the sound signals of the important monitoring area underground in the coal mine in real time. That is, by setting sound sensors in places such as tunnels, coal mining faces, excavation faces, central water pump rooms and power distribution rooms, the sounds of places such as tunnels, coal mining faces, excavation faces, central water pump rooms and power distribution rooms are collected and monitored. When the collected sound characteristics meet the set conditions, the alarm state is entered.

[0081] Step S102, the sound signal is transformed through discrete wavelet transform to obtain different sub-bands, and the different sub-bands are subjected to fit test statistics to obtain a test statistic; a denoising threshold is obtained using an estimated relationship of a noise coefficient; the test statistic is compared with the denoising threshold, and an expected signal coefficient is retained, where the expected signal coefficient corresponds to a target sound signal.

[0082] In a possible implementation, the sound signal is transformed through a discrete wavelet transform to obtain different sub-bands, and the different sub-bands are subjected to a goodness-of-fit test to obtain a test statistic; a denoising threshold is obtained using an estimated relationship of a noise coefficient; the test statistic is compared with the denoising threshold, and an expected signal coefficient is retained, where the expected signal coefficient corresponds to a target sound signal, including:

[0083] Normalizing the discrete wavelet coefficients to determine the discrete wavelet coefficients, and determining the threshold of the false alarm probability through the fit test;

[0084] Performing a fit test on the discrete wavelet coefficient and the threshold of the false alarm probability to obtain a threshold coefficient;

[0085] The statistical distance between the empirical distribution function of discrete wavelet coefficients and the cumulative distribution function of simulated WGN coefficients is calculated to obtain the test statistic;

[0086] Comparing the test statistic with the threshold coefficient, retaining the coefficient of the expected signal, and discarding the coefficient of the noise signal;

[0087] Through inverse discrete wavelet transform, the target sound signal corresponding to the expected signal coefficient is output.

[0088] In an embodiment of the present application, after processing the downhole sound signal using discrete wavelet transform, the obtained wavelet coefficients are subjected to a goodness-of-fit test, and a discrete wavelet goodness-of-fit test (DWT-GOF) denoising algorithm is proposed to separate the signal and noise at multiple scales into separate coefficients.

[0089] Here, let T GOF Represents the threshold operator, then the threshold T at scale j j for:

[0090]

[0091] In the formula, is the discrete wavelet coefficient of white Gaussian noise (WGN) with input standard deviation of б.

[0092] Find the threshold T of false alarm probability through goodness of fit test j . Normalize the discrete wavelet coefficients and normalize the discrete wavelet coefficients Perform a goodness-of-fit test to produce threshold coefficients The relationship between the three is:

[0093]

[0094] The denoising assumption of this algorithm is expressed as follows:

[0095]

[0096] In the formula, H0 and H1 represent the null hypothesis and alternative hypothesis under multi-scale, respectively. The specific process is: calculate the statistical distance between the empirical distribution function of the discrete wavelet coefficients and the cumulative distribution function of the simulated WGN coefficients to obtain the test statistic τ AD ; Compare it with the threshold, the coefficient of the noise (i.e. H0) sample is identified in the transformation matrix, then the coefficient originates from the noise; the remaining coefficient is retained as the expected signal (i.e. H1), then the coefficient originates from the signal and retains the original value. Finally, the coefficient matrix is ​​inverse normalized and the denoised audio signal is restored by inverse discrete wavelet transform.

[0097] The main process of denoising is as follows:

[0098] (1) Input mine sound signal;

[0099] (2) Perform discrete wavelet transform on the sound signal to obtain its coefficient matrix;

[0100] (3) Simulate the input WGN and estimate T based on the algorithm theory proposed in the embodiment of the present application j , the present invention fa The value is: 0.005;

[0101] (4) Calculate the test statistic τ by combining the algorithm theory and fit test proposed in the application embodiment AD

[0102] (5) Determine whether the coefficient belongs to the signal or the noise, retain the coefficient that belongs to the signal, and discard the coefficient that belongs to the noise;

[0103] (6) Through inverse discrete wavelet transform, the denoised sound signal is output.

[0104] In the embodiments of the present application, the denoising algorithm is verified in detail, such as Figure 2 a is a timing diagram of a sound signal with a noise of -10dB added. Figure 2 b It can be seen that the wavelet threshold denoising algorithm has a better denoising effect on the sound signal with -10dB noise added, and there is a situation of locally amplifying the noise. Figure 2 c It can be seen that the EEMD+wavelet denoising algorithm has a poor denoising effect on the sound signal and there is a phenomenon of increasing noise. Figure 2 It can be seen that the algorithm proposed in this paper can effectively overcome the shortcomings of the above methods, and is the best among several algorithms overall.

[0105] Step S103, preprocessing the training samples of the target sound signal and extracting training feature vectors, wherein the training feature vectors are respectively input into the SVM model and the LSSVM model for training to obtain a first recognition model corresponding to the SVM and a second recognition model corresponding to the LSSVM.

[0106] In a possible implementation manner, preprocessing the target sound signal includes signal anomaly detection, and the signal anomaly detection includes:

[0107] When the adjacent signal is greater than the set threshold, it is determined that the local peak value change of the regional signal is abnormal.

[0108] In the embodiment of the present application, in order to accurately identify the explosion sound signal, it is necessary to accurately distinguish the abnormal situation from the non-abnormal situation in the sound signal. The abnormality of the sound signal in the time domain waveform is reflected in the obvious change of the waveform steepness of the signal at the abnormal point. Based on this, the method intends to use the conditional local peak rate to detect the signal abnormality.

[0109] The local peak is determined by calculating the first-order derivative of the signal. If the signal at the mth point satisfies the following formula, it is considered to be a local peak.

[0110] x(m)-x(m-1)>Δ

[0111] x(m+1)-x(m)≤Δ

[0112] Where x(m) is the peak value of the signal at the mth point, and Δ is the minimum height difference between adjacent peaks.

[0113] On this basis, in order to be closer to engineering applications, the difference between adjacent local peaks is used as another inspection condition, that is, when the difference between adjacent signals is greater than the set threshold, it is determined that the local peak here changes rapidly, and the signal is considered abnormal. Therefore, there are two key input items in this method: (1) the minimum height difference between adjacent peaks, which is selected as 0.1 in this paper; (2) the difference between adjacent signals is set to 1.5 times the average value of the local peak calculated by the signal.

[0114] In the embodiment of the present application, it is necessary to extract the feature vector of the sound signal. When extracting features, wavelet packet decomposition is used. Wavelet packet decomposition is to add the decomposition of the high-frequency part of the signal on the basis of wavelet decomposition, which can overcome the disadvantage of poor frequency resolution of wavelet decomposition in the high-frequency band, and can evenly divide the frequency band of the sound signal into multiple levels. In view of the fact that the sound signal collected by the sound signal collector is usually a non-stationary signal, the collected sound signal can be decomposed into a signal of any independent frequency band through wavelet packet decomposition, and the explosion sound signal and the non-explosion sound signal have obvious differences in time domain and frequency domain. Therefore, wavelet packet decomposition has more advantages in processing sound signals and can distinguish the energy distribution of sound signals in different frequency bands in coal mines. That is:

[0115] The energy of the sound signal in different sub-bands after decomposition by wavelet packets is:

[0116]

[0117] Where d i,j (k) is the amplitude of the kth (k=1, 2, ..., N, N is the signal length) sampling point corresponding to the jth layer of the i-th sub-band after signal decomposition.

[0118] The total signal energy is

[0119]

[0120] The energy proportion of wavelet packet is

[0121]

[0122] The energy proportion of wavelet packets constitutes the characteristic vector λ = [λ1,λ2,λ3,…,λ t ]. In this paper, the number of wavelet packet decomposition layers is selected as 3, so the eigenvector containing 8 eigenvalues ​​can be obtained.

[0123] Step S104: preprocess the test sample of the target sound signal and extract a test feature vector, wherein the test feature vector is used to input the first recognition model and the second recognition model respectively to obtain a result label.

[0124] In a possible implementation, processing the test sample by using the SVM model includes:

[0125] Obtaining a feature vector representing the target sound signal through wavelet transform, and inputting the feature vector into a SVM classifier for training;

[0126] Establish a linear SVM classifier according to the SVM model to find the optimal hyperplane;

[0127] The target sound signal is divided into explosion sound and non-explosion sound according to the SVM classifier.

[0128] In one possible implementation, the training of the classifier of the SVM model is obtained through a first hyperparameter, wherein the first hyperparameter includes a regularization coefficient and a gamma kernel parameter, wherein the regularization coefficient is used to characterize the tolerance of the SVM model to errors, and the gamma kernel parameter is used to characterize the Gaussian range of action corresponding to the support vector.

[0129] In a possible implementation, training the test sample by using the LSSVM model includes:

[0130] Obtain a feature vector representing the target sound signal through wavelet transform, and input the feature vector into the LSSVM classifier for training;

[0131] Establishing a classifier of the model through the decision function of the LSSVM model;

[0132] The target sound signal is classified into explosion sound and non-explosion sound according to the classifier.

[0133] In one possible implementation, the training of the LSSVM classifier is obtained by optimizing the second hyperparameter by the whale optimization algorithm, wherein the second hyperparameter includes a regularization coefficient and a gamma kernel parameter, the regularization coefficient is used to characterize the tolerance of the LSSVM model to errors, and the gamma kernel parameter is used to characterize the Gaussian range of action corresponding to the support vector.

[0134] It can be understood that in the embodiment of the present application, the SVM model uses a support vector machine, which is a binary classification model that maps the feature vector of an instance to some points in space. The purpose of the SVM is to draw a line to distinguish the two types of points, so that even if there are new points, this line can also make a good classification. SVM is suitable for small and medium-sized data samples, nonlinear, and high-dimensional classification problems.

[0135] According to the principle of SVM classifier, the specific steps of establishing SVM model in this paper are:

[0136] (1) The 8-dimensional energy ratio obtained by wavelet packet transform is used as the feature vector to characterize the characteristics of the sound signal and is input into the SVM classifier for training.

[0137] (2) Based on the established recognition model as the main classification basis, a linear SVM is established to seek the optimal hyperplane, which can be converted into obtaining:

[0138]

[0139] Where ω is the normal vector of the hyperplane and b is the displacement term, which determines the distance between the hyperplane and the origin.

[0140] Among them, the constraint condition of formula (4) is formula (5):

[0141] y i (ω T x i +b)≥1,i=1,2,…,p#(5)

[0142] In the formula, y i is the sound signal category label, and p is the number of sound signals to be tested.

[0143] (3) The sound signal to be tested is divided into explosion sound signal and non-explosion sound signal through the classifier, and the result is output.

[0144] In order to select appropriate hyperparameters to complete the training of the SVM classifier and improve the recognition accuracy of SVM classification, this method chooses practical Bayesian optimization to solve this problem. Bayesian optimization finds the value that minimizes the objective function by establishing a probability model based on the evaluation results of the objective function. The hyperparameters of Bayesian optimization used in this method model are the regularization coefficient and the gamma kernel parameter. The numerical value of the regularization coefficient represents the tolerance of the model to errors, and the gamma kernel parameter determines the range of action of the Gaussian corresponding to the support vector.

[0145] The goal of the hyperparameter optimization algorithm is to find the global optimal solution in the fastest way. To achieve this goal, the Bayesian optimization steps in this paper are as follows: First, determine the input quantity, and use the 8 parameters obtained by wavelet packet transform of the sound signal as the calculation quantity, and the sound signal type label as the verification value. Secondly, determine the probability model, and use the probability model to replace the complex objective function of the original evaluation. This method selects Gaussian process regression. Finally, train and verify the network, and use the set of hyperparameters with the smallest error in the optimization process as the optimized hyperparameters. The verification error meets the requirements and the network is terminated.

[0146] In the embodiment of the present application, the specific steps of establishing the LSSVM model recognition model are:

[0147] (1) The 8-dimensional energy ratio obtained by wavelet packet transform is used as the feature vector to characterize the characteristics of the sound signal and is input into the LSSVM classifier for training.

[0148] (2) Based on the established recognition model as the main classification basis, the decision function is as follows:

[0149]

[0150] In the formula, x is the input feature vector, K(x,x i ) is the kernel function, α i is the Lagrange multiplier.

[0151] According to the characteristics of the sound signal, the sum function of this method selects the RBF radial basis function, and its formula is expressed as follows:

[0152]

[0153] Where σ is the kernel parameter, ||x i -x j || is the Euclidean distance between two points.

[0154] (3) The sound signal to be tested is divided into explosion sound signal and non-explosion sound signal through the classifier, and the result is output.

[0155] In order to select appropriate hyperparameters, the hyperparameters optimized by the whale optimization algorithm in this method model are the regularization coefficient and the gamma kernel parameter. The specific process is:

[0156] (1) Initialization parameters: the size of the whale population SN, which is 5 in this method; the maximum number of iterations T max , this paper takes 10;

[0157] (2) The algorithm initializes the location of the whale population;

[0158] (3) Calculate the fitness value of each whale, sort them according to the size of the fitness value, and select SN as the initial population;

[0159] (4) Calculate the fitness values ​​of SN individuals and find the individual position with the smallest fitness value as the optimal position;

[0160] (5) Update the position of the next generation;

[0161] (6) If the termination condition is met, the optimal individual is output, that is, the optimal solution found by the algorithm; otherwise, return to step (4).

[0162] Step S105: Determine the category of the target sound according to the result label.

[0163] In a possible implementation, the category of the target sound is determined according to the result label, including: the result label includes a first result label corresponding to the first recognition model and a second result label corresponding to the second recognition model; when the results of identification of the first result label and the second result label are both explosion sound signals, the target sound is determined to be an explosion sound.

[0164] In the embodiment of the present application, assuming that 0 represents a non-explosion sound signal and 1 represents an explosion sound signal, the output result can be obtained according to the method as follows: Figure 3 This method can achieve a higher recognition accuracy by fusing the SVM and LSSVM models, and can solve the problems of missed detection and false detection caused by the insufficiency of the recognition algorithm in the process of explosion sound signal recognition.

[0165] Figure 4 The present application embodiment provides a coal mine gas and coal dust explosion sound recognition system, including:

[0166] The sound collection module 408, namely the collection module, is used to collect sound signals of the target area through sound collection equipment, is installed in the underground monitored area, and is used to collect sound signals of the underground monitored area of ​​the coal mine in real time.

[0167] The sound collection module 408 of this embodiment adopts a mining explosion-proof sound collection module, which is fixedly installed in the tunnel, coal mining face and excavation face, and is connected to the substation 406 through a communication cable for wired communication. The substation 406 is responsible for the access and data exchange of underground communication equipment including sound sensors, which meets the explosion-proof requirements of underground coal mines.

[0168] The sounds collected by the sound collection module 408 include the normal operating sounds of the coal mining face and the tunneling face equipment, the driving sounds of cars, the working sounds of water pumps, the working sounds of fans, the operating sounds of low-voltage power feeding equipment, the operating sounds of high-voltage power distribution equipment, the operating sounds of primary power distribution equipment, the sounds of belt motors, the operating sounds of belts, and the sounds of people talking.

[0169] The explosion disaster monitoring server 402 is installed on the well and is in communication with the sound collection module to receive the real-time sound signal of the sound collection module. The explosion disaster monitoring server 402 includes a denoising module 4021, a training module 4022, a testing module 4023, and a classification module 4024.

[0170] The denoising module 4021 is used to obtain different sub-bands by discrete wavelet transform of the sound signal, perform fit test statistics on the different sub-bands to obtain test statistics; obtain a denoising threshold by using an estimated relationship of a noise coefficient; compare the test statistic with the denoising threshold, and retain a desired signal coefficient, wherein the desired signal coefficient corresponds to a target sound signal;

[0171] A training module 4022 is used to preprocess the training samples of the target sound signal and extract training feature vectors, wherein the training feature vectors are respectively input into the SVM model and the LSSVM model for training to obtain a first recognition model corresponding to the SVM and a second recognition model corresponding to the LSSVM;

[0172] A testing module 4023 is used to preprocess the test sample of the target sound signal and extract a test feature vector, wherein the test feature vector is used to input the first recognition model and the second recognition model respectively to obtain a result label;

[0173] The classification module 4024 is used to determine the category of the target sound according to the result label.

[0174] The monitoring terminal 404 is installed on the well and is connected to the explosion disaster monitoring server for communication, and is used for receiving the alarm signal from the explosion disaster monitoring server and generating an alarm.

[0175] The monitoring terminal of this embodiment has the function of sound and light alarm, receives the alarm signal of the explosion disaster monitoring server and generates sound and light alarm; it has the function of real-time sound monitoring and historical sound retrieval, and the production management personnel can view the on-site sound through the monitoring host.

[0176] Optionally, this embodiment further includes a storage server 403, which is communicatively connected with the sound collection module 408 and is used to store the sound data collected by the sound collection module 408;

[0177] The monitoring terminal 404 is connected to the storage server and is used to retrieve and view the on-site sound of the area to be monitored or retrieve historical monitoring sound data through the storage server.

[0178] In addition, this embodiment is also provided with:

[0179] The network switch 404 is responsible for the management and data exchange of all devices connected to the mining Ethernet.

[0180] The underground switch 405 is responsible for the direct data exchange between the substation and the network switch. It has a flameproof casing and meets the explosion-proof requirements of underground coal mines.

[0181] The AC / DC conversion device 407 converts the underground AC power into DC power, and supplies power to underground switches, substations, sound collection modules and other equipment through power supply cables.

[0182] In a possible implementation, the system is used to implement Figures 1 to 3 The method of any one of the embodiments.

[0183] In summary, the coal mine gas and coal dust explosion sound recognition method and system provided by the embodiment of the present application can effectively remove noise from the mine sound signal and maintain the detail information and edge retention capability of the sound signal. When recognizing the denoised sound signal, the accuracy of sound signal recognition can be improved, and it has strong compatibility and portability.

[0184] Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application is shown in FIG. Figure 5 As shown, the electronic device 500 includes: a memory 510 and a processor 520 .

[0185] The memory 510 stores a computer program that can be executed by at least one processor 520. The computer program is executed by at least one processor 520 to enable the electronic device to implement the method provided in any of the above embodiments.

[0186] The memory 510 and the processor 520 may be connected via a bus 530 .

[0187] The relevant instructions can be understood by referring to the relevant descriptions and effects corresponding to the method embodiments, which will not be repeated here.

[0188] An embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the following Figures 1 to 3 Corresponding to any embodiment method.

[0189] Among them, the computer-readable storage medium may be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like.

[0190] An embodiment of the present application provides a computer program product, which includes computer-executable instructions, which are used to implement the following when the computer-executable instructions are executed by a processor: Figures 1 to 3 The method of any corresponding embodiment.

[0191] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0192] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the disclosure disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only, and the true scope of the present application is indicated by the claims.

[0193] It should be understood that the present application is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

Claims

1. A method for identifying the sound of coal mine gas and coal dust explosions, characterized in that: include: Collecting sound signals in the target area through sound collection equipment; The sound signal is transformed by discrete wavelet transform to obtain different sub-bands, and the different sub-bands are subjected to fit test statistics to obtain test statistics; A denoising threshold is obtained by using an estimated relationship of the noise coefficient; the test statistic is compared with the denoising threshold, and an expected signal coefficient is retained, where the expected signal coefficient corresponds to a target sound signal; Preprocessing the training samples of the target sound signal and extracting training feature vectors, wherein the training feature vectors are respectively input into the SVM model and the LSSVM model for training to obtain a first recognition model corresponding to the SVM and a second recognition model corresponding to the LSSVM; Preprocessing the test sample of the target sound signal and extracting a test feature vector, wherein the test feature vector is used to input the first recognition model and the second recognition model respectively to obtain a result label; According to the result label, the category of the target sound is determined.

2. The method according to claim 1, characterized in that The sound signal is transformed by discrete wavelet transform to obtain different sub-bands, and the different sub-bands are subjected to fit test statistics to obtain test statistics; and a denoising threshold is obtained by using an estimated relationship of a noise coefficient; The test statistic is compared with the denoising threshold, and the expected signal coefficient is retained, where the expected signal coefficient corresponds to the target sound signal, including: Normalizing the discrete wavelet coefficients to determine the discrete wavelet coefficients, and determining the threshold of the false alarm probability through the fit test; Performing a fit test on the discrete wavelet coefficient and the threshold of the false alarm probability to obtain a threshold coefficient; Calculate the statistical distance between the empirical distribution function of discrete wavelet coefficients and the cumulative distribution function of simulated WGN coefficients to obtain the test statistic; Comparing the test statistic with the threshold coefficient, retaining the coefficient of the expected signal, and discarding the coefficient of the noise signal; Through inverse discrete wavelet transform, the expected signal coefficients are outputted to correspond to the target sound signal.

3. The method according to claim 1 or 2, characterized in that: Determining the category of the target sound according to the result label includes: The result label includes a first result label corresponding to the first recognition model and a second result label corresponding to the second recognition model; when the results of identification of the first result label and the second result label are both explosion sound signals, the target sound is determined to be an explosion sound.

4. The method according to claim 1, characterized in that: Preprocessing the target sound signal includes signal anomaly detection, and the signal anomaly detection includes: When the adjacent signal is greater than the set threshold, it is determined that the local peak value change of the regional signal is abnormal.

5. The method according to any one of claims 1 to 4, characterized in that: The test sample is processed by the SVM model, including: Obtain a feature vector representing the target sound signal through wavelet transform, and input the feature vector into a SVM classifier for training; Establish a linear SVM classifier according to the SVM model to find the optimal hyperplane; The target sound signal is divided into explosion sound and non-explosion sound according to the SVM classifier.

6. The method according to claim 5, characterized in that The training of the classifier of the SVM model is obtained through the first hyperparameter, which includes a regularization coefficient and a gamma kernel parameter. The regularization coefficient is used to characterize the tolerance of the SVM model to errors, and the gamma kernel parameter is used to characterize the Gaussian action range corresponding to the support vector.

7. The method according to any one of claims 1 to 4, characterized in that: The test sample is trained by the LSSVM model, including: Obtain a feature vector representing the target sound signal through wavelet transform, and input the feature vector into the LSSVM classifier for training; Establishing a classifier of the model through the decision function of the LSSVM model; The target sound signal is classified into explosion sound and non-explosion sound according to the classifier.

8. The method according to claim 7, characterized in that The training of the LSSVM classifier is obtained by optimizing the second hyperparameter by the whale optimization algorithm, wherein the second hyperparameter includes a regularization coefficient and a gamma kernel parameter, wherein the regularization coefficient is used to characterize the tolerance of the LSSVM model to errors, and the gamma kernel parameter is used to characterize the Gaussian action range corresponding to the support vector.

9. A coal mine gas and coal dust explosion sound recognition system, characterized in that: include: A collection module, used to collect sound signals in a target area through sound collection equipment; A denoising module, used for obtaining different sub-bands by discrete wavelet transforming the sound signal, and performing fit test statistics on the different sub-bands to obtain test statistics; A denoising threshold is obtained by using an estimated relationship of the noise coefficient; the test statistic is compared with the denoising threshold, and an expected signal coefficient is retained, where the expected signal coefficient corresponds to a target sound signal; A training module, used for preprocessing the training samples of the target sound signal and extracting training feature vectors, wherein the training feature vectors are respectively input into the SVM model and the LSSVM model for training to obtain a first recognition model corresponding to the SVM and a second recognition model corresponding to the LSSVM; A testing module, used for preprocessing the test sample of the target sound signal and extracting a test feature vector, wherein the test feature vector is used for inputting the first recognition model and the second recognition model respectively to obtain a result label; A classification module is used to determine the category of the target sound according to the result label.

10. The system according to claim 9, characterized in that The system is used to implement the method according to any one of claims 2-8.

Images