A coal rock fracture damage mode classification method and system

Abstract

The present application provides a coal rock fracture damage mode classification method and system, comprising: preprocessing the collected acoustic emission signals and labeling the damage categories of various signal parameter data; inputting and training the overall coal rock fracture damage mode classification model of various signal parameter data and labeled data, the model comprising an adaptive diversity wavelet kernel module, a CWT-based convolution feature extraction layer, a BiLSTM feature fusion layer, and a classification layer; the adaptive diversity wavelet kernel module extracts the time domain features, frequency domain features and CWT features of various signal parameter data, and adaptively adjusts the wavelet kernel parameters to generate a diversified wavelet kernel set inputting the CWT-based convolution feature extraction layer, which comprises a global branch and a local branch for parallel computing, respectively extracting global features and local features; the global features and local features are fused through the BiLSTM feature fusion layer; and the fused features are inputted into the classification layer to classify the coal rock fracture damage mode. The present application can classify the coal rock fracture damage mode.

Description

Technical Field

[0001] The present invention relates to the technical field of coal rock fracture damage pattern classification, and in particular to a coal rock fracture damage pattern classification method and system. Background Art

[0002] With the continuous increase in the depth of coal mining in my country, the efficient identification of coal rock fracture damage has become a technical challenge that urgently needs to be solved in intelligent mining. In deep mining environments, the risk of coal rock fracture damage is increasing, becoming one of the major hazards restricting safe and efficient mine production. The timely and accurate detection and classification of coal rock fracture damage modes (such as tension and shear) are of great significance for preventing geological disasters and reducing safety hazards. Traditional coal rock fracture damage mode classification methods mostly rely on time domain or frequency domain analysis, such as Fourier transform and short-time Fourier transform. However, these methods lack the time-frequency resolution of non-stationary signals and cannot effectively capture low-frequency trends and high-frequency transient characteristics. Summary of the Invention

[0003] In order to solve the technical problems existing in the above-mentioned prior art, the present invention provides a method and system for classifying coal rock fracture damage patterns. The technical solution is as follows:

[0004] In one aspect, a method for classifying coal rock fracture damage patterns is provided, the method comprising:

[0005] S1. Collect acoustic emission signals of coal rock fracture damage through acoustic emission sensors and uniaxial compression tests;

[0006] S2. Preprocess the collected acoustic emission signals to construct a standardized data set, and label the damage categories of various signal parameter data in the standardized data set;

[0007] S3. Inputting various signal parameter data and labeled data in the standardized data set into a training model for the overall coal-rock fracture damage pattern classification model, wherein the overall fracture damage pattern classification model includes an adaptive diversity wavelet kernel module, a CWT-based convolution feature extraction layer, a BiLSTM feature fusion layer, and a classification layer;

[0008] The adaptive diversity wavelet kernel module extracts the time domain features, frequency domain features and CWT features of various signal parameter data, and adaptively adjusts the wavelet kernel parameters to generate a diverse wavelet kernel set;

[0009] Inputting the generated diverse wavelet kernel set into the CWT-based convolution feature extraction layer, the CWT-based convolution feature extraction layer includes a global branch and a local branch of parallel calculation, respectively extracting global features and local features;

[0010] The extracted global features and local features are fused through the BiLSTM feature fusion layer to obtain fused features;

[0011] Inputting the fused features into a classification layer to classify coal rock fracture damage patterns;

[0012] S4. Use the trained overall coal-rock fracture damage pattern classification model to classify the coal-rock fracture patterns to be classified.

[0013] In another aspect, a coal rock fracture damage pattern classification system is provided, the system comprising:

[0014] The acquisition module is used to collect acoustic emission signals of coal rock fracture damage through acoustic emission sensors and uniaxial compression tests;

[0015] A preprocessing module is used to preprocess the collected acoustic emission signals, construct a standardized data set, and label the damage categories of various signal parameter data in the standardized data set;

[0016] A training module, configured to input various signal parameter data and annotated data in the standardized data set and train an overall coal-rock fracture damage pattern classification model, wherein the overall fracture damage pattern classification model includes an adaptive diversity wavelet kernel module, a CWT-based convolution feature extraction layer, a BiLSTM feature fusion layer, and a classification layer;

[0017] The adaptive diversity wavelet kernel module extracts the time domain features, frequency domain features and CWT features of various signal parameter data, and adaptively adjusts the wavelet kernel parameters to generate a diverse wavelet kernel set;

[0018] Inputting the generated diverse wavelet kernel set into the CWT-based convolution feature extraction layer, the CWT-based convolution feature extraction layer includes a global branch and a local branch of parallel calculation, respectively extracting global features and local features;

[0019] The extracted global features and local features are fused through the BiLSTM feature fusion layer to obtain fused features;

[0020] Inputting the fused features into a classification layer to classify coal rock fracture damage patterns;

[0021] The classification module is used to classify the coal rock fracture patterns to be classified using the trained overall coal rock fracture damage pattern classification model.

[0022] On the other hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the above-mentioned coal rock fracture damage mode classification method.

[0023] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned coal rock fracture damage mode classification method.

[0024] The beneficial effects brought about by the technical solution provided by the present invention include at least:

[0025] 1) Comprehensive and in-depth feature mining:

[0026] The adaptive diversity wavelet kernel module can adaptively adjust the wavelet kernel parameters and generate a diverse wavelet kernel set.

[0027] Feature extraction is performed through the CWT-based convolution feature extraction layer, which includes global branches and local branches of parallel computing. The global branch grasps the overall trend and macroscopic characteristics of coal rock fracture damage, and the local branch captures local subtle changes. It fully mines data features from multiple scales and angles. Compared with the traditional single feature extraction method, it can obtain richer and more comprehensive coal rock fracture feature information, providing strong support for accurate classification. The global branch is more complex: it is suitable for processing complex time-frequency distributions of low-frequency trends and extracting more comprehensive and robust features. The adaptive CWT in the global branch is suitable for low-frequency non-stationary characteristics. The FAN model adaptively screens key features and optimizes attention fusion. The contribution of wavelet kernel and subsequent processing improve stability. Multi-scale convolution pooling captures multi-scale information. Full connection and Dropout integrate features and prevent overfitting. Local branches are simpler: high-frequency transient features are usually more local and direct, and can be efficiently extracted with a simple structure, avoiding unnecessary computational overhead. Wavelet kernel convolution in local branches efficiently extracts high-frequency transients. Feature splicing integrates multi-wavelet kernel information. The KAN model (B-spline basis function) processes high-frequency nonlinear patterns. Full connection dimensionality reduction (to 16 dimensions) aligns with global features and prevents overfitting. The overall design ensures the complementarity of low-frequency and high-frequency features, improving classification performance.

[0028] 2) Significant advantages of feature fusion:

[0029] The Bidirectional Long Short-Term Memory (BiLSTM) network is used for feature fusion, combining the bidirectional extraction of global branches (low-frequency macro features) and local branches (high-frequency detailed features). This fully captures the time-frequency characteristics of coal and rock fractures and reduces feature omissions. BiLSTM effectively processes sequence data and learns features from both positive and negative directions, making the fused feature vector more representative and comprehensively reflecting both global and local information about coal and rock fractures. This improves the model's ability to understand and classify complex coal and rock fracture patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0031] Figure 1 This is a flow chart of a coal rock fracture damage pattern classification method provided by an embodiment of the present invention;

[0032] Figure 2 This is a schematic diagram of classification and discrimination of coal rock fracture damage patterns provided by an embodiment of the present invention;

[0033] Figure 3 Schematic diagram of the overall fracture damage mode classification model structure provided by an embodiment of the present invention;

[0034] Figure 4 Schematic diagram of the structure of the adaptive diversity wavelet kernel module provided by an embodiment of the present invention;

[0035] Figure 5 This is a block diagram of a coal rock fracture damage pattern classification system provided by an embodiment of the present invention;

[0036] Figure 6 It is a structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0038] An embodiment of the present invention provides a method for classifying coal rock fracture damage patterns. The method can be implemented by an electronic device, which can be a terminal or a server. Figure 1 The flowchart of the method is shown, and the processing flow may include the following steps:

[0039] S1. Collect acoustic emission signals of coal rock fracture damage through acoustic emission sensors and uniaxial compression tests;

[0040] Because the acoustic emission signal of coal rock fracture damage is the instantaneous elastic wave generated by the monitored micro-fracture, it has various parameters, including energy, count, etc., which can well analyze the internal fracture mechanism and determine the degree and type of fracture. Therefore, the embodiment of the present invention collects the acoustic emission signal of coal rock fracture damage for coal rock fracture damage mode classification.

[0041] S2. Preprocess the collected acoustic emission signals to construct a standardized data set, and label the damage categories of various signal parameter data in the standardized data set;

[0042] Optionally, the S2 specifically includes:

[0043] S201, using a Gaussian low-pass filter to remove baseline drift from the acoustic emission signal, using the following filtering formula:

[0044]

[0045] in, is the impulse response function of the Gaussian low-pass filter, is the acoustic emission signal, is the signal after high-pass filtering, is the time offset variable in the convolution integral;

[0046] S202, apply DWT soft threshold denoising, select db4 wavelet basis, decompose Signal to L level, , soft threshold processing is applied to the high-frequency detail coefficients, and the processed high-frequency detail coefficients as follows:

[0047]

[0048] in, is the Nyquist efficiency, is the lowest effective frequency 1kHz, d is the detail coefficient, is the universal threshold, is the noise standard deviation, is the signal length;

[0049] S203, reconstruct the signal, the reconstructed signal as follows:

[0050]

[0051] in, is the low-frequency approximation coefficient, IDWT is the inverse DWT;

[0052] S204. Various parameters of the reconstructed signal are stored as time series data, including the amplitude, rise time, duration, energy, average frequency, count, and peak frequency of the acoustic emission signal. The time series corresponding to each parameter is segmented and normalized. The signal is divided into windows (10 ms, 10,000 points), and each window is normalized separately.

[0053] S205, fixed window segmentation, output standardized data set , range [0,1], , segmented signal parameter set , the segmentation formula is as follows:

[0054]

[0055] in, is the signal parameter value in the kth segment window, is the normalized signal parameter value, is the sampling rate, 5000 is the segment step size, and 10000 is the window length expressed in the number of sampling points. is the starting time of the kth window, is the end time of the kth window, is the time range of the k-th window;

[0056] S206, calculate the rise time / amplitude RA value and count / duration AF value of each fixed segment window, and perform classification and discrimination by comparing the changes in the distribution of RA and AF values ​​during the injury process: set the threshold slope to , when calculating When the rupture damage mode is determined to be tensile rupture, When , the fracture damage mode is determined to be shear fracture, such as Figure 2 As shown;

[0057] S207, marking the coal rock fracture damage pattern of each fixed segment window as a corresponding classification, generating a classification label, and dividing the signal parameters of the fixed segment window into two categories based on the comparison between the AF value and the threshold slope k: , it is marked as tensile rupture. , it is marked as shear fracture.

[0058] S3, input various signal parameter data and annotation data in the standardized data set and train the overall coal-rock fracture damage pattern classification model, the overall fracture damage pattern classification model includes an adaptive diversity wavelet kernel module, a CWT-based convolution feature extraction layer, a BiLSTM feature fusion layer, and a classification layer, such as Figure 3 As shown;

[0059] The adaptive diversity wavelet kernel module extracts the time domain features, frequency domain features and CWT features of various signal parameter data, and adaptively adjusts the wavelet kernel parameters to generate a diverse wavelet kernel set;

[0060] Inputting the generated diverse wavelet kernel set into the CWT-based convolution feature extraction layer, the CWT-based convolution feature extraction layer includes a global branch and a local branch of parallel calculation, respectively extracting global features and local features;

[0061] The extracted global features and local features are fused through the BiLSTM feature fusion layer to obtain fused features;

[0062] Inputting the fused features into a classification layer to classify coal rock fracture damage patterns;

[0063] Alternatively, as Figure 4 As shown in FIG, the processing process of the adaptive diversity wavelet kernel module specifically includes:

[0064] S311, segmented signal , perform feature extraction to obtain time domain features, including mean, variance, peak and energy: , where N=10000, , , ;

[0065] Output: time domain feature vector , dimension is 4;

[0066] S312, segmented signal , use the Fast Fourier Transform FFT to calculate the power spectrum:

[0067]

[0068] Compute frequency-domain statistics:

[0069] Main frequency: , represents the maximum frequency component of the power spectrum;

[0070] Spectral entropy: ,in, , a measure of the complexity of the spectrum;

[0071] Spectrum Center: , represents the weighted center frequency of the spectrum;

[0072] Output: frequency domain feature vector , dimension is 3;

[0073] S313. Use the default Morlet wavelet for preliminary CWT feature extraction:

[0074]

[0075] Scale range: , corresponding to frequencies from 1 kHz to 500 kHz;

[0076] Compute the average energy at each scale: ,in, , uniformly sample 10 scales, , forming a feature vector;

[0077] Output: CWT feature vector , dimension is 10;

[0078] S314. Concatenate the time domain feature vector, the frequency domain feature vector, and the CWT feature vector:

[0079] , the dimension is 17;

[0080] S315, the concatenated feature vector Input a multi-branch wavelet kernel generation network, wherein the multi-branch wavelet kernel generation network includes:

[0081] Shared layer: fully connected layer FC, input dimension 17, output dimension 64, activation function ReLU;

[0082] Branch 1: Morlet branch

[0083] FC(64→32)→FC(32→2), output wavelet parameters :

[0084] , output , ranges [2,10] and [0.5,5];

[0085] Branch 2: Mexican Hat Branch

[0086] FC(64→32)→FC(32→1), output wavelet parameters :

[0087] , output , range [0.5,5];

[0088] Branch 3: Paul branch

[0089] FC(64→32)→FC(32→1), output wavelet parameters :

[0090] , output , range [2,8];

[0091] S316. Perform network training and define the loss function:

[0092]

[0093] in, is used Calculated CWT coefficients, using Adam optimizer, learning rate , control parameter update speed, is the total loss function, balancing the reconstruction error and signal-to-noise ratio, To reconstruct the error weight, we prioritize ensuring that the wavelet kernel can effectively decompose the signal. is the weight of the signal-to-noise ratio, improving the decomposition quality, is the reconstruction error, which measures the ability of the wavelet kernel to decompose the signal. is the signal-to-noise ratio, which measures the quality of the decomposed features;

[0094] S317, output parameter set ,in is the Morlet wavelet parameter , is the Mexican Hat wavelet parameter , is the Paul wavelet parameter ;

[0095] S318, according to the parameters , generate a wavelet kernel set :

[0096] Generate a Morlet wavelet kernel:

[0097] in, is the center frequency, which controls the frequency resolution of the wavelet. is the width parameter, which controls the time resolution of the wavelet. is a time variable, discretized , It is Morlet wavelet, which is suitable for extracting periodic features;

[0098] Generate the Mexican Hat wavelet kernel:

[0099] in, is the scale factor, controlling the wavelet width, To normalize time and adjust the wavelet shape, It is the Mexican Hat wavelet, which is suitable for extracting transient features;

[0100] Generate Paul wavelet kernel:

[0101] Among them, m is the order, which controls the shape of the wavelet. The larger the value, the more complex it is. Paul wavelet, suitable for extracting high-frequency complex features;

[0102] S319, Discretization and Normalization, at time points Discretize the wavelet kernel, discrete form:

[0103]

[0104] Normalized form:

[0105]

[0106] Where n is the time index, ranging from [-5000, 4999], corresponding to a 10ms window. The discrete wavelet kernel value is normalized to ensure that the wavelet kernel energy is 1, which meets the mathematical requirements of CWT;

[0107] S3110, output discrete and normalized wavelet kernel set:

[0108]

[0109] in, It is an adaptive diversity wavelet kernel, including 3 types of wavelet kernels. They correspond to Morlet, Mexican Hat, and Paul wavelets respectively, adapting to different characteristics of the signal.

[0110] Alternatively, as Figure 3 As shown in FIG, the processing process of the global branch of the CWT-based convolution feature extraction layer specifically includes:

[0111] S321, perform initial CWT calculation:

[0112] The adaptive diversity wavelet kernel set , define the initial scale range, focusing on low frequencies from 1 kHz to 50 kHz:

[0113]

[0114] Use three parallel branches to perform each wavelet kernel Execute CWT:

[0115]

[0116] Discrete implementation, using FFT to accelerate convolution:

[0117]

[0118] Output: Initial global time-frequency feature map set , each feature map dimension is ;

[0119] in, is the initial scale, is the time shift, the range is [0,10ms], and it is discretized into 10000 points. is the initial CWT coefficient, which represents the time-frequency distribution of the signal at the low-frequency scale;

[0120] S322. Use the FAN model to dynamically adjust the scale and wavelet kernel weight, including:

[0121] Input: Initial global time-frequency feature map set ;

[0122] Among them, the FAN model architecture includes:

[0123] Feature extraction layer:

[0124] For each feature map Perform global average pooling:

[0125]

[0126] in, , the output dimension is 1;

[0127] Concatenate feature vectors: , the output dimension is 3;

[0128] Fully connected layer FC1: input dimension 3, output dimension 32, activation function ReLU;

[0129] Fully connected layer FC2: input dimension 32, output dimension 49, corresponding to 49 scales;

[0130] Sigmoid activation:

[0131]

[0132] in, ;

[0133] Output: scale weight vector , dimension 49;

[0134] Fully connected layer FC3: input dimension 3, output dimension 16, activation function ReLU;

[0135] Fully connected layer FC4: input dimension 16, output dimension 3, corresponding to 3 wavelet kernels;

[0136] Softmax activation:

[0137]

[0138] in, ;

[0139] Output: wavelet kernel weight vector ;

[0140] Scale screening:

[0141] according to Important screening criteria:

[0142] choose The scale of ;

[0143] like , then add the selection The highest scale, until ;

[0144] Output: The filtered scale set, , size is 24, wavelet kernel weight ;

[0145] in, is the global pooling feature of the 𝑖th wavelet kernel, dimension 1, reflecting the overall energy of the feature map, which is used to evaluate the contribution of each wavelet kernel. To concatenate feature vectors, dimension 3, integrate the features of three wavelet kernels, and input them into the FAN model. is the scale weight, ranging from [0,1]. The larger the value, the more important the scale is for global feature extraction. is the wavelet kernel weight, ranging from [0,1], which dynamically adjusts the contribution of each wavelet kernel to the global features. The scale set after filtering is 24 in size, focusing on the important low-frequency range and reducing redundant calculations;

[0146] S323, performing adaptive CWT calculation, including:

[0147] Input: Segmented signal , adaptive diversity wavelet kernel set , the filtered scale set ;

[0148] S323-1 performs CWT recalculation:

[0149] For each wavelet kernel , in the screening scale Re-execute CWT:

[0150]

[0151] Use FFT to speed up:

[0152]

[0153] Output: adaptive global time-frequency feature map set , each feature map dimension is ;

[0154] in, The low-frequency scale after screening is 24 in size, focusing on important frequencies and reducing the amount of calculation;

[0155] S323-2. Perform attention fusion, including:

[0156] Input: adaptive global time-frequency feature map set , wavelet kernel weight ;

[0157] Perform weighted fusion of wavelet kernel weights:

[0158]

[0159] Output: Comprehensive global time-frequency feature map , the dimension is ;

[0160] in, Wavelet kernel weight, range [0,1], dynamically adjust the contribution of each wavelet kernel, To fuse the feature maps, the features of three wavelet kernels are integrated to highlight the important low-frequency patterns;

[0161] S323-3, As input, it is processed by Conv-BN-ReLU for the first time and outputs the convolution feature map , the dimension is ;

[0162] S323-4. Convolution feature map Perform CWT processing:

[0163] Perform CWT on each channel (16 channels in total) using the default Morlet wavelet, :

[0164]

[0165] Discrete implementation, using FFT to accelerate convolution:

[0166]

[0167] Output: time-frequency feature map , the dimension is ;

[0168] in, is the default Morlet wavelet, is the center frequency of the Morlet wavelet, which controls the balance of time-frequency resolution. is the width parameter of the Morlet wavelet, which controls the time window. is the CWT scale, For time translation;

[0169] S323-5, the time-frequency feature map Perform the second Conv-BN-ReLU process and output the convolution feature map , the dimension is ;

[0170] S323-6. Convolution feature map Perform ICWT treatment:

[0171] Perform inverse CWT on each channel (16 channels in total) to reconstruct the time domain signal:

[0172]

[0173] Output: reconstructed signal , the dimension is ;

[0174] S323-7, output the first Conv-BN-ReLU With ICWT output Form a residual connection and perform addition operations, including:

[0175] Will Along the scale dimension Perform global average pooling to compress the scale dimension:

[0176]

[0177] in, , the output dimension is ,and Dimension alignment;

[0178] After aligning and Form a residual connection and add channel by channel:

[0179]

[0180] Output: residual features , the dimension is ;

[0181] S323-8, the residual features ,application Convolution, output convolution feature map , the dimension is ;

[0182] S324. Perform multi-scale wavelet convolution pooling, including convolution and pooling operations, wherein the multi-scale wavelet convolution includes:

[0183] Input: is the input;

[0184] Multi-scale settings: define 3 scale levels, , corresponding to the original scale, magnification by 2 times, and reduction by 2 times;

[0185] Wavelet convolution: For each channel (32 channels in total) and each wavelet kernel (3 types in total), convolution is performed at 3 scale levels:

[0186]

[0187] in, , indicating the Wavelet kernels at scale The transformation below;

[0188] Discrete implementation, using FFT to accelerate convolution:

[0189]

[0190] Output: Multi-scale convolution feature set ,common feature maps, each feature map has a dimension of ;

[0191] Multi-scale pooling includes:

[0192] Input: Multi-scale convolution feature set , a total of 288 feature maps, each dimension is ;

[0193] Pooling operation: Apply maximum pooling to each feature map with a window size of 100 points and a stride of 50 points:

[0194]

[0195] in, The output time dimension is:

[0196]

[0197] Output: Pooled feature set , a total of 288 feature maps, each dimension is ;

[0198] Feature splicing and convolution fusion include:

[0199] Feature splicing: Splice all pooled features along the channel dimension. , the dimension after splicing is ;

[0200] Convolution fusion: First, apply 1×1 convolution to reduce the dimension, output , the dimension is , and then further fused by applying 3×3 convolution , the dimension is ;

[0201] Output: multi-scale feature map , the dimension is ;

[0202] S325, multi-scale feature map The Flatten layer flattens it into a vector: , the dimension after flattening is ;

[0203] S326. Apply the fully connected layer FC to the flattened vector and use the ReLU activation function for calculation: ,in is the weight matrix, which can be used to learn parameters. is the bias vector, which can be learned as a parameter;

[0204] S327, apply Dropout, discard probability , to prevent overfitting, , output global feature vector , the dimension is 512;

[0205] S328, yes Applying the fully connected layer for dimensionality reduction, the output is , with a dimension of 16.

[0206] Alternatively, as Figure 3 As shown in FIG, the processing process of the local branch of the CWT-based convolution feature extraction layer specifically includes:

[0207] S331, perform wavelet kernel convolution:

[0208] Adaptive diversity wavelet kernel ensemble and segmented signals As input, convolution is performed on each wavelet kernel to provide high-frequency features:

[0209]

[0210] Discrete implementation, using FFT to accelerate convolution:

[0211]

[0212] S332, use the maximum pooling layer for pooling dimensionality reduction:

[0213] Apply max pooling to each convolution result with a window size of 100 points and a stride of 50 points:

[0214]

[0215] in, The output time dimension is:

[0216]

[0217] Output: wavelet kernel feature set , each feature dimension is ;

[0218] in, is the wavelet kernel convolution feature, is the feature after pooling;

[0219] S333, perform feature splicing:

[0220] Wavelet kernel feature set Splicing along the channel dimension:

[0221]

[0222] S334, will Flattened to a vector by the flatten layer: , the dimension after flattening is ;

[0223] S335. The flattened vector is subjected to a KAN model for feature extraction. The KAN model includes:

[0224] First layer:

[0225] Input dimension: 600;

[0226] Hidden units: 128;

[0227] Each input dimension is mapped to a hidden unit using a learnable B-spline function:

[0228]

[0229] in, is the B-spline basis function, K=5 is the spline order, is a learnable parameter;

[0230] Output: , the dimension is 128, and is calculated as, ;

[0231] Second layer:

[0232] Input dimension: 128;

[0233] Output dimension: 64;

[0234] Also using the B-spline function:

[0235]

[0236] Output: , the dimension is 64, and is calculated as, ;

[0237] Output: KAN model features , dimension is 64;

[0238] S336, KAN model features After the fully connected layer, the dimension is reduced and the local branch features are output , with a dimension of 16.

[0239] Optionally, the processing of the BiLSTM feature fusion layer specifically includes:

[0240] Forward LSTM: ;

[0241] Backward LSTM: ;

[0242] Splicing: ;

[0243] Take the hidden state of the last time step as the fusion feature: ;

[0244] Output: fused features , dimension is 64;

[0245] in, The dimension is 64. For the parameter setting, the number of hidden units is 64, 32 each for the forward and backward directions, the input dimension is 32, and the output dimension is 64, 32 for the forward direction and 32 for the backward direction. are the forward and backward hidden states, each with a dimension of 32, is the hidden state after splicing, dimension 64, To fuse features, the dimension is 64 to capture temporal dependencies.

[0246] Optionally, the processing of the classification layer specifically includes:

[0247] The fusion features After full connection, the output of the dimension is calculated as: ,in, The dimension is The learnable parameters of Output of dimension 2, classification score , the dimension is 2, corresponding to the two categories of tension and shear;

[0248] Output the category probability through Softmax and map the features to different categories:

[0249]

[0250] Where c=1 means tension, c=0 means shear, is the category probability, ranging from [0,1], and is 1, is the probability of the tension category, is the probability of the cut category.

[0251] The overall fracture damage pattern classification model of the embodiment of the present invention uses classification evaluation indicators for evaluation and a confusion matrix for classification analysis, including:

[0252] Evaluation analysis using classification evaluation metrics:

[0253]

[0254]

[0255]

[0256]

[0257] Among them, when the detected fixed segmentation window and the true label are both cut, it is defined as , when the fixed segment window of the inspected sample is shear but the label is tension, it is defined as , when the detected fixed segmentation window and the true label are both tensed, it is defined as , when the detected fixed segmentation window is tensioned but the true label is sheared, it is defined as ;

[0258] Use confusion matrix for visualization analysis, the matrix form is:

[0259]

[0260] The confusion matrix shows the specific types of errors, including and ,By analyzing these errors, we can find the weak points of the model in these two categories.

[0261] S4. Use the trained overall coal-rock fracture damage pattern classification model to classify the coal-rock fracture patterns to be classified.

[0262] like Figure 5 As shown, an embodiment of the present invention further provides a coal rock fracture damage mode classification system, the system comprising:

[0263] The acquisition module 510 is used to collect acoustic emission signals of coal rock fracture damage through acoustic emission sensors and uniaxial compression tests;

[0264] A preprocessing module 520 is used to preprocess the collected acoustic emission signals, construct a standardized data set, and label damage categories for various signal parameter data in the standardized data set;

[0265] A training module 530 is configured to input various signal parameter data and annotated data in the standardized data set and train an overall coal-rock fracture damage pattern classification model, wherein the overall fracture damage pattern classification model includes an adaptive diversity wavelet kernel module, a CWT-based convolution feature extraction layer, a BiLSTM feature fusion layer, and a classification layer;

[0266] The adaptive diversity wavelet kernel module extracts the time domain features, frequency domain features and CWT features of various signal parameter data, and adaptively adjusts the wavelet kernel parameters to generate a diverse wavelet kernel set;

[0267] Inputting the generated diverse wavelet kernel set into the CWT-based convolution feature extraction layer, the CWT-based convolution feature extraction layer includes a global branch and a local branch of parallel calculation, respectively extracting global features and local features;

[0268] The extracted global features and local features are fused through the BiLSTM feature fusion layer to obtain fused features;

[0269] Inputting the fused features into a classification layer to classify coal rock fracture damage patterns;

[0270] The classification module 540 is used to classify the coal-rock fracture patterns to be classified using the trained overall coal-rock fracture damage pattern classification model.

[0271] The functional structure of a coal rock fracture damage pattern classification system provided by an embodiment of the present invention corresponds to a coal rock fracture damage pattern classification method provided by an embodiment of the present invention, and will not be described in detail here.

[0272] Figure 6 1 is a schematic structural diagram of an electronic device 600 provided in an embodiment of the present invention. The electronic device 600 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 601 and one or more memories 602, wherein the memories 602 store at least one instruction, and the at least one instruction is loaded and executed by the processor 601 to implement the steps of the above-mentioned coal rock fracture damage mode classification method.

[0273] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory device including instructions. The instructions are executable by a processor in a terminal to implement the above-described coal-rock fracture damage pattern classification method. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device.

[0274] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0275] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A coal rock fracture damage pattern classification method, characterized in that: The method comprises: S1. Collect acoustic emission signals of coal rock fracture damage through acoustic emission sensors and uniaxial compression tests; S2. Preprocess the collected acoustic emission signals to construct a standardized data set, and label the damage categories of various signal parameter data in the standardized data set; S3. Inputting various signal parameter data and labeled data in the standardized data set into a training model for the overall coal-rock fracture damage pattern classification model, wherein the overall fracture damage pattern classification model includes an adaptive diversity wavelet kernel module, a CWT-based convolution feature extraction layer, a BiLSTM feature fusion layer, and a classification layer; The adaptive diversity wavelet kernel module extracts the time domain features, frequency domain features and CWT features of various signal parameter data, and adaptively adjusts the wavelet kernel parameters to generate a diverse wavelet kernel set; Inputting the generated diverse wavelet kernel set into the CWT-based convolution feature extraction layer, the CWT-based convolution feature extraction layer includes a global branch and a local branch of parallel calculation, respectively extracting global features and local features; The extracted global features and local features are fused through the BiLSTM feature fusion layer to obtain fused features; Inputting the fused features into a classification layer to classify coal rock fracture damage patterns; S4. Use the trained overall coal-rock fracture damage pattern classification model to classify the coal-rock fracture patterns to be classified.

2. The method according to claim 1, characterized in that Said S2 specifically includes: S201, using a Gaussian low-pass filter to remove baseline drift from the acoustic emission signal, using the following filtering formula: ； in, is the impulse response function of the Gaussian low-pass filter, is the acoustic emission signal, is the signal after high-pass filtering, is the time offset variable in the convolution integral; S202, apply DWT soft threshold denoising, select db4 wavelet basis, decompose Signal to L level, , apply soft threshold processing to the high-frequency detail © coefficient, and the processed high-frequency detail coefficient as follows: ； in, is the Nyquist efficiency, is the lowest effective frequency 1kHz, d is the detail coefficient, is the universal threshold, is the noise standard deviation, is the signal length; S203, reconstruct the signal, the reconstructed signal as follows: ； in, is the low-frequency approximation coefficient, IDWT is the inverse DWT; S204. Storing various parameters of the reconstructed signal as time series data, including the amplitude, rise time, duration, energy, average frequency, count, and peak frequency of the acoustic emission signal, performing segmented normalization on the time series corresponding to each parameter, dividing the signal into windows, and normalizing each window separately; S205, fixed window segmentation, output standardized data set , range [0,1], , segmented signal parameter set , the segmentation formula is as follows: ； in, is the signal parameter value in the kth segment window, is the normalized signal parameter value, is the sampling rate, 5000 is the segment step size, and 10000 is the window length expressed in the number of sampling points. is the starting time of the kth window, is the end time of the kth window, is the time range of the k-th window; S206, calculate the rise time / amplitude RA value and count / duration AF value of each fixed segment window, and perform classification and discrimination by comparing the changes in the distribution of RA and AF values ​​during the injury process: set the threshold slope to , when calculating When the rupture damage mode is determined to be tensile rupture, When , the fracture damage mode is determined to be shear fracture; S207, marking the coal rock fracture damage pattern of each fixed segment window as a corresponding classification, generating a classification label, and dividing the signal parameters of the fixed segment window into two categories based on the comparison between the AF value and the threshold slope k: , it is marked as tensile rupture. , it is marked as shear fracture.

3. The method according to claim 2, characterized in that The processing process of the adaptive diversity wavelet kernel module specifically includes: S311, segmented signal , perform feature extraction to obtain time domain features, including mean, variance, peak and energy: , where N=10000, , , ; Output: time domain feature vector , dimension is 4; S312, segmented signal , use the Fast Fourier Transform FFT to calculate the power spectrum: ； Compute frequency-domain statistics: Main frequency: , represents the maximum frequency component of the power spectrum; Spectral entropy: ,in, , a measure of the complexity of the spectrum; Spectrum Center: , represents the weighted center frequency of the spectrum; Output: frequency domain feature vector , dimension is 3; S313. Use the default Morlet wavelet for preliminary CWT feature extraction: ； Scale range: , corresponding to frequencies from 1 kHz to 500 kHz; Compute the average energy at each scale: ,in, , uniformly sample 10 scales, , forming a feature vector; Output: CWT feature vector , dimension is 10; S314. Concatenate the time domain feature vector, the frequency domain feature vector, and the CWT feature vector: , the dimension is 17; S315, the concatenated feature vector Input a multi-branch wavelet kernel generation network, wherein the multi-branch wavelet kernel generation network includes: Shared layer: fully connected layer FC, input dimension 17, output dimension 64, activation function ReLU; Branch 1: Morlet branch FC(64→32)→FC(32→2), output wavelet parameters : , output , ranges [2,10] and [0.5,5]; Branch 2: Mexican Hat Branch FC(64→32)→FC(32→1), output wavelet parameters : , output , range [0.5,5]; Branch 3: Paul branch FC(64→32)→FC(32→1), output wavelet parameters : , output , range [2,8]; S316. Perform network training and define the loss function: ； in, is used Calculated CWT coefficients, using Adam optimizer, learning rate , control parameter update speed, is the total loss function, balancing the reconstruction error and signal-to-noise ratio, To reconstruct the error weight, we prioritize ensuring that the wavelet kernel can effectively decompose the signal. is the weight of the signal-to-noise ratio, improving the decomposition quality, is the reconstruction error, which measures the ability of the wavelet kernel to decompose the signal. is the signal-to-noise ratio, which measures the quality of the decomposed features; S317, output parameter set ,in is the Morlet wavelet parameter , is the Mexican Hat wavelet parameter , is the Paul wavelet parameter ; S318, according to the parameters , generate a wavelet kernel set : Generate a Morlet wavelet kernel: ; in, is the center frequency, which controls the frequency resolution of the wavelet. is the width parameter, which controls the time resolution of the wavelet. is a time variable, discretized , It is Morlet wavelet, which is suitable for extracting periodic features; Generate the Mexican Hat wavelet kernel: in, is the scale factor, controlling the wavelet width, To normalize time and adjust the wavelet shape, It is the Mexican Hat wavelet, which is suitable for extracting transient features; Generate Paul wavelet kernel: ; Among them, m is the order, which controls the shape of the wavelet. The larger the value, the more complex it is. Paul wavelet, suitable for extracting high-frequency complex features; S319, Discretization and Normalization, at time points Discretize the wavelet kernel, discrete form: ； Normalized form: ； Where n is the time index, ranging from [-5000, 4999], corresponding to a 10ms window. The discrete wavelet kernel value is normalized to ensure that the wavelet kernel energy is 1, which meets the mathematical requirements of CWT; S3110, output discrete and normalized wavelet kernel set: ； in, It is an adaptive diversity wavelet kernel, including 3 types of wavelet kernels. They correspond to Morlet, Mexican Hat, and Paul wavelets respectively, adapting to different characteristics of the signal.

4. The method according to claim 3, characterized in that The processing process of the global branch of the CWT-based convolution feature extraction layer specifically includes: S321, perform initial CWT calculation: The adaptive diversity wavelet kernel set , define the initial scale range, focusing on low frequencies from 1 kHz to 50 kHz: ； Use three parallel branches to perform each wavelet kernel Execute CWT: ； Discrete implementation, using FFT to accelerate convolution: ； Output: Initial global time-frequency feature map set , each feature map dimension is ; in, is the initial scale, is the time shift, the range is [0,10ms], and it is discretized into 10000 points. is the initial CWT coefficient, which represents the time-frequency distribution of the signal at the low-frequency scale; S322. Use the FAN model to dynamically adjust the scale and wavelet kernel weight, including: Input: Initial global time-frequency feature map set ; Among them, the FAN model architecture includes: Feature extraction layer: For each feature map Perform global average pooling: ； in, , the output dimension is 1; Concatenate feature vectors: , the output dimension is 3; Fully connected layer FC1: input dimension 3, output dimension 32, activation function ReLU; Fully connected layer FC2: input dimension 32, output dimension 49, corresponding to 49 scales; Sigmoid activation: ； in, ; Output: scale weight vector , dimension 49; Fully connected layer FC3: input dimension 3, output dimension 16, activation function ReLU; Fully connected layer FC4: input dimension 16, output dimension 3, corresponding to 3 wavelet kernels; Softmax activation: ； in, ; Output: wavelet kernel weight vector ; Scale screening: according to Important screening criteria: choose The scale of ; like , then add the selection The highest scale, until ; Output: The filtered scale set, , size is 24, wavelet kernel weight ; in, is the global pooling feature of the 𝑖th wavelet kernel, dimension 1, reflecting the overall energy of the feature map, which is used to evaluate the contribution of each wavelet kernel. To concatenate feature vectors, dimension 3, integrate the features of three wavelet kernels, and input them into the FAN model. is the scale weight, ranging from [0,1]. The larger the value, the more important the scale is for global feature extraction. is the wavelet kernel weight, ranging from [0,1], which dynamically adjusts the contribution of each wavelet kernel to the global features. The scale set after filtering is 24 in size, focusing on the important low-frequency range and reducing redundant calculations; S323, performing adaptive CWT calculation, including: Input: Segmented signal , adaptive diversity wavelet kernel set , the filtered scale set ; S323-1 performs CWT recalculation: For each wavelet kernel , in the screening scale Re-execute CWT: ； Use FFT to speed up: ； Output: adaptive global time-frequency feature map set , each feature map dimension is ; in, The low-frequency scale after screening is 24 in size, focusing on important frequencies and reducing the amount of calculation; S323-2. Perform attention fusion, including: Input: adaptive global time-frequency feature map set , wavelet kernel weight ; Perform weighted fusion of wavelet kernel weights: ； Output: Comprehensive global time-frequency feature map , the dimension is ; in, Wavelet kernel weight, range [0,1], dynamically adjust the contribution of each wavelet kernel, To fuse the feature maps, the features of three wavelet kernels are integrated to highlight the important low-frequency patterns; S323-3, As input, it is processed by Conv-BN-ReLU for the first time and outputs the convolution feature map , the dimension is ; S323-4. Convolution feature map Perform CWT processing: Perform CWT on each channel, using the default Morlet wavelet, : ； Discrete implementation, using FFT to accelerate convolution: ； Output: time-frequency feature map , the dimension is ; in, is the default Morlet wavelet, is the center frequency of the Morlet wavelet, which controls the balance of time-frequency resolution. is the width parameter of the Morlet wavelet, which controls the time window. is the CWT scale, For time translation; S323-5, the time-frequency feature map Perform the second Conv-BN-ReLU process and output the convolution feature map , the dimension is ; S323-6. Convolution feature map Perform ICWT treatment: Perform the inverse CWT on each channel to reconstruct the time domain signal: ； Output: reconstructed signal , the dimension is ; S323-7, output the first Conv-BN-ReLU With ICWT output Form a residual connection and perform addition operations, including: Will Along the scale dimension Perform global average pooling to compress the scale dimension: ； in, , the output dimension is ,and Dimension alignment; After aligning and Form a residual connection and add channel by channel: ； Output: residual features , the dimension is ; S323-8, the residual features ,application Convolution, output convolution feature map , the dimension is ; S324. Perform multi-scale wavelet convolution pooling, including convolution and pooling operations, wherein the multi-scale wavelet convolution includes: Input: is the input; Multi-scale settings: define 3 scale levels, , corresponding to the original scale, magnification by 2 times, and reduction by 2 times; Wavelet convolution: For each channel and each wavelet kernel, convolution is performed at three scale levels: ； in, , indicating the Wavelet kernels at scale The transformation below; Discrete implementation, using FFT to accelerate convolution: ； Output: Multi-scale convolution feature set ,common feature maps, each feature map has a dimension of ; Multi-scale pooling includes: Input: Multi-scale convolution feature set , a total of 288 feature maps, each dimension is ; Pooling operation: Apply maximum pooling to each feature map with a window size of 100 points and a stride of 50 points: ； in, The output time dimension is: ； Output: Pooled feature set , a total of 288 feature maps, each dimension is ; Feature splicing and convolution fusion include: Feature splicing: Splice all pooled features along the channel dimension. , the dimension after splicing is ; Convolution fusion: First, apply 1×1 convolution to reduce the dimension, output , the dimension is , and then further fused by applying 3×3 convolution , the dimension is ; Output: multi-scale feature map , the dimension is ; S325, multi-scale feature map The Flatten layer flattens it into a vector: , the dimension after flattening is ; S326. Apply the fully connected layer FC to the flattened vector and use the ReLU activation function for calculation: ,in is the weight matrix, which can be used to learn parameters. is the bias vector, which can be learned as a parameter; S327, apply Dropout, discard probability , to prevent overfitting, , output global feature vector , the dimension is 512; S328, yes Applying the fully connected layer for dimensionality reduction, the output is , with a dimension of 16.

5. The method according to claim 4, characterized in that The processing process of the local branch of the CWT-based convolution feature extraction layer specifically includes: S331, perform wavelet kernel convolution: Adaptive diversity wavelet kernel ensemble and segmented signals As input, convolution is performed on each wavelet kernel to provide high-frequency features: ； Discrete implementation, using FFT to accelerate convolution: ； S332, use the maximum pooling layer for pooling dimensionality reduction: Apply max pooling to each convolution result with a window size of 100 points and a stride of 50 points: ； in, The output time dimension is: ； Output: wavelet kernel feature set , each feature dimension is ; in, is the wavelet kernel convolution feature, is the feature after pooling; S333, perform feature splicing: Wavelet kernel feature set Splicing along the channel dimension: ； S334, will Flattened to a vector by the flatten layer: , the dimension after flattening is ; S335. The flattened vector is subjected to a KAN model for feature extraction. The KAN model includes: First layer: Input dimension: 600; Hidden units: 128; Each input dimension is mapped to a hidden unit using a learnable B-spline function: ； in, is the B-spline basis function, K=5 is the spline order, is a learnable parameter; Output: , the dimension is 128, and is calculated as, ; Second layer: Input dimension: 128; Output dimension: 64; Also using the B-spline function: ； Output: , the dimension is 64, and is calculated as, ; Output: KAN model features , dimension is 64; S336, KAN model features After the fully connected layer, the dimension is reduced and the local branch features are output , with a dimension of 16.

6. The method according to claim 5, characterized in that The processing of the BiLSTM feature fusion layer specifically includes: Forward LSTM: ; Backward LSTM: ; Splicing: ; Take the hidden state of the last time step as the fusion feature: ; Output: fused features , dimension is 64; in, The dimension is 64. For the parameter setting, the number of hidden units is 64, 32 each for the forward and backward directions, the input dimension is 32, and the output dimension is 64, 32 for the forward direction and 32 for the backward direction. are the forward and backward hidden states, each with a dimension of 32, is the hidden state after splicing, dimension 64, To fuse features, the dimension is 64 to capture temporal dependencies.

7. The method according to claim 6, characterized in that The processing process of the classification layer specifically includes: The fusion features After full connection, the output of the dimension is calculated as: ,in, The dimension is The learnable parameters of Output of dimension 2, classification score , the dimension is 2, corresponding to the two categories of tension and shear; Output the category probability through Softmax and map the features to different categories: ； Where c=1 means tension, c=0 means shear, is the category probability, ranging from [0,1], and is 1, is the probability of the tension category, is the probability of the cut category.

8. A coal rock fracture damage pattern classification system, characterized by: The system comprises: The acquisition module is used to collect acoustic emission signals of coal rock fracture damage through acoustic emission sensors and uniaxial compression tests; A preprocessing module is used to preprocess the collected acoustic emission signals, construct a standardized data set, and label the damage categories of various signal parameter data in the standardized data set; A training module, configured to input various signal parameter data and annotated data in the standardized data set and train an overall coal-rock fracture damage pattern classification model, wherein the overall fracture damage pattern classification model includes an adaptive diversity wavelet kernel module, a CWT-based convolution feature extraction layer, a BiLSTM feature fusion layer, and a classification layer; The adaptive diversity wavelet kernel module extracts the time domain features, frequency domain features and CWT features of various signal parameter data, and adaptively adjusts the wavelet kernel parameters to generate a diverse wavelet kernel set; Inputting the generated diverse wavelet kernel set into the CWT-based convolution feature extraction layer, the CWT-based convolution feature extraction layer includes a global branch and a local branch of parallel calculation, respectively extracting global features and local features; The extracted global features and local features are fused through the BiLSTM feature fusion layer to obtain fused features; Inputting the fused features into a classification layer to classify coal rock fracture damage patterns; The classification module is used to classify the coal rock fracture patterns to be classified using the trained overall coal rock fracture damage pattern classification model.

9. An electronic device comprising a processor and a memory, wherein the memory stores at least one instruction, characterized in that: The at least one instruction is loaded and executed by the processor to implement the coal rock fracture damage mode classification method according to any one of claims 1-7.

10. A computer-readable storage medium, wherein at least one instruction is stored in the storage medium, characterized in that: The at least one instruction is loaded and executed by the processor to implement the coal rock fracture damage mode classification method according to any one of claims 1 to 7.

Images