Multi-channel micro-seismic event automatic classification method based on SNR signal selection

Abstract

The invention relates to the technical field of micro-seismic monitoring, in particular to a multi-channel micro-seismic event automatic classification method based on SNR signal selection, which comprises the following steps of: segmenting a main body part and a noise part of each channel signal of a micro-seismic event by adopting a long-short time window ratio algorithm; calculating the signal-to-noise ratio of each channel signal; the first N channels are selected according to the signal-to-noise ratio sequence; re-sampling the selected first N channels to a fixed length; carrying out learnable band-pass filtering on the N channel signals through a SincNet layer; performing multi-scale deep feature extraction on shallow features of the N channel signals through an IPPA attention module; performing cross-channel fusion on the multi-scale deep features of the N channel signals output by the IPPA attention module through a feature fusion module to obtain multi-channel features; and classifying the multi-channel features based on a classifier to obtain a classification result of the micro-seismic event. According to the invention, efficient and accurate classification of multi-channel micro-seismic signals is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of microseismic monitoring, and more particularly to a multi-channel microseismic event automatic classification method based on SNR signal selection. Background Art

[0002] In the field of mine safety monitoring and geological disaster early warning, microseismic event classification is a key link in the microseismic monitoring system, and its accuracy directly affects the reliability of subsequent event positioning.

[0003] Traditional methods, such as the short-time average / long-time average (STA / LTA) algorithm, while computationally simple, suffer from high misclassification rates in low signal-to-noise ratio environments. Machine learning methods based on manual feature extraction, such as support vector machines (SVMs), can partially alleviate this problem, but their reliance on manually designed features hinders generalization. In recent years, the introduction of deep learning techniques has driven the development of single-channel signal classification, such as using CNNs or LSTMs for end-to-end classification. However, these methods independently process each channel's signal and then fuse the results through voting, which can easily lead to systematic errors due to noise-dominated channels and ignore the spatiotemporal correlations between multiple sensors.

[0004] To address this issue, researchers have attempted to convert multi-channel signals into images and process them using image classification models (such as DCNN-SPP). Although these methods have achieved high accuracy, the signal imaging process destroys the continuity of the time and frequency domains and is computationally expensive, making it difficult to meet the needs of real-time monitoring. Furthermore, existing methods generally fail to effectively address issues such as the uneven number of channels, variable lengths, and low signal-to-noise ratio interference in the original signal, limiting the model's performance in complex real-world environments. Therefore, there is an urgent need for a microseismic event classification method that can balance multi-channel signal collaborative processing, efficient computation, and high classification accuracy to improve the reliability and real-time performance of mine safety monitoring. Summary of the Invention

[0005] In view of this, the present invention provides a multi-channel microseismic event automatic classification method based on SNR signal selection, which realizes efficient and accurate classification of multi-channel microseismic signals.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A multi-channel microseismic event automatic classification method based on SNR signal selection, characterized by comprising the following steps:

[0008] S1. Use the long-short time window ratio algorithm to segment the main part and noise part of each channel signal of the microseismic event;

[0009] S2. Calculate the signal-to-noise ratio of each channel signal;

[0010] S3, sort by signal-to-noise ratio and select the first N channels;

[0011] S4, resampling the selected first N channels to unify the lengths of the N channel signals to a fixed length;

[0012] S5. Build a deep learning model, which includes a SincNet layer, a convolutional layer, an IPPA attention module, a feature fusion module, and a classifier; after performing learnable bandpass filtering on N channel signals through the SincNet layer, perform shallow feature extraction through the convolutional layer;

[0013] S6, perform multi-scale deep feature extraction on the shallow features of N channel signals through the IPPA attention module to capture the local-global features of each channel signal;

[0014] S7, cross-channel fusion of the multi-scale deep features of the N channel signals output by the IPPA attention module is performed through the feature fusion module to obtain multi-channel features;

[0015] S8. Classify the multi-channel features based on the classifier to obtain the classification results of the microseismic events.

[0016] Furthermore, S1 includes:

[0017] Generate STA / LTA characteristic function, the expression is:

[0018]

[0019] Where STA represents the average energy in a short period of time, LTA represents the average energy in a long period of time, t is the current sample index, nsta is the STA window length, nlta is the LTA window length, a is the input signal, and i is the index in the window;

[0020] After generating the STA / LTA characteristic function, the open threshold and closed threshold are set, and the segment where the STA / LTA characteristic function value is continuously higher than the open threshold and not lower than the closed threshold is determined as the main part of the valid signal, and the rest is classified as the noise part.

[0021] Furthermore, in S2, the signal-to-noise ratio calculation formula is:

[0022]

[0023] Among them, α is the main part or noise part of the signal, N signal is the length of the main part of the signal, N noise is the length of the signal-noise portion.

[0024] Furthermore, in S5, a SincNet layer, two convolutional layers, and an IPPA attention module constitute a feature extraction unit; the deep learning model contains a total of N feature extraction units, and the N channel signals are input into the N feature extraction units one by one for feature extraction.

[0025] Furthermore, in S5, the calculation formula of the SincNet layer is:

[0026] y i [n]=x[n]*g i [n,f (i,1) ,f (i,2) ]

[0027] g i [n,f (i,1) ,f (i,2) ]=2f (i,2) sinc(2πf (i,2) n)-2f (i,1) sinc(2πf (i,1) n)

[0028]

[0029] Among them, i is the output channel, that is, the i-th Sinc filter; n is the input signal index, f (i,1) 、f (i,2) represents the learnable lower cutoff frequency and upper cutoff frequency of the i-th filter; x[n] represents the one-dimensional original input signal, y i [n] represents the output signal of the i-th channel, g i [·] represents the filter function constructed based on the i-th Sinc function.

[0030] Furthermore, in S5, the channel signal passes through the SincNet layer and then undergoes maximum pooling with a kernel size of 3×1, layer normalization, and Leaky ReLU activation processing, and then passes through two standard convolutional layers, followed by maximum pooling, layer normalization, and Leaky ReLU activation processing to obtain shallow features.

[0031] Furthermore, in S6, the IPPA attention module consists of an adaptive average pooling submodule and an adaptive maximum pooling submodule connected in series;

[0032] The adaptive average pooling submodule focuses on extracting the global statistical features of the input features and depicts the overall trend and stability of the signal through the adaptive average pooling operation;

[0033] The adaptive max pooling submodule focuses on local high response areas and uses adaptive max pooling operations to capture peaks and sudden changes in input features.

[0034] Furthermore, both the adaptive average pooling submodule and the adaptive maximum pooling submodule use a parallel multi-scale pyramid structure to extract hierarchical features at three scales from the shallow features of the input.

[0035] Among them, the first path uses 1×1 adaptive pooling to capture global features, the second path uses 2×1 adaptive pooling to extract medium-scale features, and the third path uses 4×1 adaptive pooling to focus on local detail features;

[0036] The feature extraction results of each path are fused through the splicing operation and input into an MLP network containing two fully connected layers, and then the channel attention weights are generated through the sigmoid activation function;

[0037] The generated channel attention weights are multiplied element-wise with the input shallow features to output deep features that fused the attention information.

[0038] Furthermore, the process of IPPA attention module calculating input features is expressed as:

[0039] V APP =σ(MLP(Concat[AAP1(F1),AAP2(F1),AAP4(F1)]))

[0040]

[0041] V AMP =σ(MLP(Concat[AAP1(F AAP ),AAP2(F AAP ),AAP4(F AAP )]))

[0042]

[0043] Among them, F1 is the shallow feature output by S6, AAP n represents the adaptive average pooling layer with kernel n in the adaptive average pooling submodule, F AAP represents the intermediate features output by the adaptive average pooling submodule, V AAP Represents the channel attention weight generated by the adaptive average pooling submodule; AMP n represents the adaptive maximum pooling layer with kernel n in the adaptive maximum pooling submodule, V AMP represents the channel attention weight generated by the adaptive maximum pooling submodule; Concat[·] represents the concatenation operation of adaptive average pooling layers or adaptive maximum pooling layers of different scales; MLP represents two fully connected layers, σ represents the sigmoid activation function, and F2 represents the deep features output by the adaptive maximum pooling submodule.

[0044] Furthermore, in S7, the feature fusion module first concatenates the deep features of each channel, and then performs feature fusion through a convolutional layer. The number of convolution kernels in the convolution layer is 64×N, and the step size is 1. The deep features of different channels are fused through a 1×1 convolution kernel. The fused features are then processed by maximum pooling with a kernel of 3×1, layer normalization, and LeakyReLU activation.

[0045] It can be seen from the above technical solutions that compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. This method uses the STA / LTA algorithm to segment the signal, separating the main component and the noise component of the actual microseismic signal, thereby more accurately calculating the signal-to-noise ratio. The N channels with the highest signal-to-noise ratio are selected, which not only unifies the number of input channels but also, more importantly, significantly improves the overall quality of the input signal by prioritizing high-SNR channels. Considering that the original signal length may vary significantly depending on the event duration, the selected channel signals are resampled to a fixed length. This process preserves the key signal characteristics while ensuring the consistency of the input size of the subsequent deep learning model.

[0047] 2. The first layer of the deep learning model of the present invention adopts an innovative SincNet structure and uses a learnable bandpass filter instead of the traditional convolution kernel to more effectively extract shallow features that are highly correlated with the original signal. At the same time, the IPPA attention module uses a two-stage attention mechanism of "global pilot + local refinement". While maintaining information integrity, it effectively enhances the model's modeling ability for weak heterogeneous features and improves its adaptability and discrimination performance for complex microseismic signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0049] Figure 1 Flowchart of the multi-channel microseismic event automatic classification method based on SNR signal selection provided by the present invention;

[0050] Figure 2 A schematic diagram of the structure of the deep learning model provided by the present invention;

[0051] Figure 3This is a schematic diagram of the structure of the IPPA attention module provided by the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] like Figure 1 As shown, the embodiment of the present invention discloses a multi-channel microseismic event automatic classification method based on SNR signal selection, comprising the following steps:

[0054] S1. Use the long-short time window ratio algorithm to segment the main part and noise part of each channel signal of the microseismic event;

[0055] S2. Calculate the signal-to-noise ratio of each channel signal;

[0056] S3, sort by signal-to-noise ratio and select the first N channels;

[0057] S4, resampling the selected first N channels to unify the lengths of the N channel signals to a fixed length;

[0058] S5. Build a deep learning model, which includes a SincNet layer, a convolutional layer, an IPPA attention module, a feature fusion module, and a classifier; after performing learnable bandpass filtering on N channel signals through the SincNet layer, perform shallow feature extraction through the convolutional layer;

[0059] S6, perform multi-scale deep feature extraction on the shallow features of N channel signals through the IPPA attention module to capture the local-global features of each channel signal;

[0060] S7, cross-channel fusion of the multi-scale deep features of the N channel signals output by the IPPA attention module is performed through the feature fusion module to obtain multi-channel features;

[0061] S8. Classify the multi-channel features based on the classifier to obtain the classification results of the microseismic events.

[0062] The above steps are further explained below.

[0063] S1. In actual engineering monitoring environments, the collected microseismic signals are inevitably mixed with various background noises. In order to more accurately calculate the signal-to-noise ratio of the signal, a long-short time window ratio algorithm is used to separate the main part and noise part of each channel signal of the microseismic event. Specifically, it includes:

[0064] Generate STA / LTA characteristic function, the expression is:

[0065]

[0066] Among them, STA represents the average energy in a short time, LTA represents the average energy in a long time, t is the current sample index, nsta is the STA window length, nlta is the LTA window length, a is the input signal, and i is the index in the window.

[0067] During the calculation process, the short-time average (STA) window length nsta and the long-time average (LTA) window length nlta need to be optimized according to the actual sampling frequency. In the present invention, nsta is taken as 0.1 times the signal length and nlsa is taken as 0.2 times the signal length.

[0068] After generating the STA / LTA characteristic function, a dynamic threshold is set (open threshold 1.5, closed threshold 0.5). The segment where the STA / LTA characteristic function value is continuously higher than the open threshold and not lower than the closed threshold is determined to be the main part of the valid signal, and the rest is classified as the noise part.

[0069] In particular, when the signal waveform presents multi-peak characteristics, it is also possible to accurately identify multiple independent valid signal segments [[on1,off1],[on2,off2],...,[on n ,off n ]].

[0070] S2. Calculate the signal-to-noise ratio (SNR) of each channel signal. The signal-to-noise ratio calculation formula is:

[0071]

[0072] Among them, α is the main part or noise part of the signal, N signal is the length of the main part of the signal, N noise is the length of the signal-noise portion.

[0073] S3. Select the top N channels by signal-to-noise ratio: Arrange the SNR values ​​calculated for each channel in the microseismic event in descending order and select the N channels with the highest SNR. In this embodiment, N is set to 4. This step not only standardizes the number of input channels but, more importantly, significantly improves the overall quality of the input signal by prioritizing channels with high signal-to-noise ratios. In special cases where fewer than N channels are available, zero padding is automatically performed to ensure uniformity of the input dimensionality.

[0074] S4. Resample the selected first N channels to unify the lengths of the N channel signals to a fixed length.

[0075] Considering that the original signal length may vary significantly due to different event durations, a resampling technique based on polyphase sinc interpolation was used to uniformly adjust the signal length of all selected channels to 10,000 sampling points (corresponding to 1 second at a sampling frequency of 10 kHz).

[0076] The selected channel signals are then resampled to a fixed length (10,000 in this paper, corresponding to a sensor sampling frequency of 10 kHz and a sampling period of 1 second). This ensures consistent signal length for the input to the deep learning model. This process preserves the key signal characteristics while ensuring consistent input size for subsequent deep learning models.

[0077] Through the above complete preprocessing process, the original non-uniform multi-channel microseismic signals are converted into standardized N-channel fixed-length signals, laying a solid foundation for subsequent deep learning classification.

[0078] The above steps S1 to S4 are the preprocessing process of the signals of each channel of the microseismic event. Next, the prediction steps S5 to S8 of the deep learning model are introduced.

[0079] Deep Learning Model MPA-SincNet Structure Reference Figure 2 The deep learning model includes a SincNet layer, a convolutional layer, an IPPA attention module, a feature fusion module and a classifier; a SincNet layer, two convolutional layers and an IPPA attention module constitute a feature extraction unit; the deep learning model contains a total of N feature extraction units, and the N channel signals are input into the N feature extraction units one by one for feature extraction.

[0080] After N channel signals are input into the deep learning model, each channel signal is processed individually by the feature extraction unit, which sequentially performs learnable bandpass filtering, shallow feature extraction, and deep feature extraction using the IPPA attention module. The deep features of each channel are then fused across channels, and the classification result is finally output by the classifier. The detailed prediction steps are as follows:

[0081] The first layer of the S5 deep learning model uses an innovative SincNet structure, using a learnable bandpass filter instead of a traditional convolution kernel to more effectively extract shallow features that are highly correlated with the original signal. The calculation formula of the SincNet layer is:

[0082] y i [n]=x[n]*g i [n,f (i,1) ,f (i,2) ]

[0083] g i [n,f (i,1) ,f(i,2) ]=2f (i,2) sinc(2πf (i,2) n)-2f (i,1) sinc(2πf (i,1) n)

[0084]

[0085] Among them, i is the output channel, that is, the i-th Sinc filter; n is the input signal index, f (i,1) 、f (i,2) represents the learnable lower cutoff frequency and upper cutoff frequency of the i-th filter; x[n] represents the one-dimensional original input signal, y i [n] represents the output signal of the i-th channel, g i [·] represents the filter function constructed based on the i-th Sinc function.

[0086] The number of convolution kernels in the SincNet layer is set to 40, the convolution kernel size is set to 251×1, and the step size is set to 1. The upper and lower cutoff frequencies f(i,1) and f(i,2) of each filter are automatically optimized through training. The input and output length relationship of the SincNet convolution layer is:

[0087] L out =L in -Filt_dim+1

[0088] Among them, L in 、L out are the input and output lengths, respectively, and Filt_dim is the size of the convolution kernel. This method takes the signal length as 10000, and the output shape after the SincNet convolution layer is (40, 9750).

[0089] To enhance the feature expression capability, the SincNet layer is followed by the maximum pooling with a kernel size of 3×1, layer normalization, and Leaky ReLU activation (negative slope coefficient α=0.2). Then, two standard convolutional layers are passed, each with 64 5×1 convolution kernels and a step size of 1. After that, the maximum pooling, layer normalization, and Leaky ReLU activation are also performed to obtain the shallow feature F1. The shallow feature shape is (64, 360).

[0090] S6. Use the IPPA attention module to perform multi-scale deep feature extraction on the shallow features of N channel signals to capture the local-global features of each channel signal.

[0091] The structure of the IPPA attention module is as follows Figure 3As shown in the figure, it consists of two sub-modules connected in series, namely the adaptive average pooling (AAP) sub-module and the adaptive maximum pooling (AMP) sub-module. Both adopt a parallel multi-scale pyramid design in structure, aiming to extract signal features from different scales and achieve feature enhancement and complementary modeling. Specifically, the adaptive average pooling sub-module focuses on extracting the global statistical features of the input features, and characterizes the overall trend and stability of the signal through the adaptive average pooling operation; the adaptive maximum pooling sub-module focuses on local high-response areas, and uses the adaptive maximum pooling operation to capture peaks and sudden changes in the input features. Through the design of a two-layer feature modeling strategy of "module-level attention direction + internal multi-scale receptive field", multi-level feature extraction and expression from the global statistical characteristics of the signal to local significant features are achieved.

[0092] Both the adaptive average pooling submodule and the adaptive maximum pooling submodule use a parallel multi-scale pyramid structure to simultaneously extract hierarchical features at three scales from the input shallow features. Global statistical features and local salient features are reflected through the multi-scale pyramid structure. The first path uses 1×1 adaptive pooling to capture global features, the second path uses 2×1 adaptive pooling to extract medium-scale features, and the third path uses 4×1 adaptive pooling to focus on local detail features. The feature extraction results of each path are fused through a splicing operation and input into an MLP network consisting of two fully connected layers. The sigmoid activation function then generates channel attention weights. The generated channel attention weights are element-wise multiplied with the input shallow features to achieve channel weighting, thereby improving the model's ability to focus on key signal areas and ultimately outputting deep features that incorporate the attention information.

[0093] During the calculation process, the shallow feature F1 from S5 is first input into the AAP submodule, and the intermediate feature F is obtained through global feature modeling. AAP ; Subsequently, the intermediate feature is input into the AMP submodule to further extract the local high response area features, and finally outputs the deep feature F2 that integrates multi-scale attention information. The calculation process is expressed as:

[0094] V APP =σ(MLP(Concat[AAP1(F1),AAP2(F1),AAP4(F1)]))

[0095]

[0096] V AMP =σ(MLP(Concat[AAP1(F AAP ),AAP2(F AAP ),AAP4(F AAP )]))

[0097]

[0098] Among them, F1 is the shallow feature output by S6, AAP n represents the adaptive average pooling layer with kernel n in the adaptive average pooling submodule, F AAP represents the intermediate features output by the adaptive average pooling submodule, V AAP Represents the channel attention weight generated by the adaptive average pooling submodule; AMP n represents the adaptive maximum pooling layer with kernel n in the adaptive maximum pooling submodule, V AMP represents the channel attention weight generated by the adaptive maximum pooling submodule; Concat[·] represents the concatenation operation of adaptive average pooling layers or adaptive maximum pooling layers of different scales; MLP represents two fully connected layers, σ represents the sigmoid activation function, and F2 represents the deep features output by the adaptive maximum pooling submodule.

[0099] The IPPA attention module utilizes a two-stage attention mechanism of "global prioritization followed by local refinement" to effectively enhance the model's ability to model weak heterogeneous features while maintaining information integrity. This improves both adaptability and discrimination performance for complex microseismic signals. The IPPA attention module is a plug-and-play module. The feature shapes output by both submodules are consistent with the shallow features of the input, resulting in the deep feature shape output by the IPPA attention module remaining (64, 360).

[0100] S7. The local-global features of the N channel signals output by the IPPA attention module are cross-channel fused through the feature fusion module to obtain multi-channel features.

[0101] Steps S5 to S6 extract features from shallow to deep for each channel of the microseismic event, and then fuse the features of each channel to generate a feature representation of the entire microseismic event. The structure of the cross-channel feature fusion module is as follows: Figure 2 shown.

[0102] Specifically, the feature fusion module first splices the deep features of each channel, and then performs feature fusion through a convolutional layer. The number of convolution kernels in the convolution layer is 64×N (256 in the present invention), and the step size is 1. The deep features of different channels are fused through a 1×1 convolution kernel, which can effectively explore the spatiotemporal correlation characteristics between multi-channel signals. The fused features are then processed by maximum pooling with a kernel of 3×1, layer normalization, and Leaky ReLU activation.

[0103] S8. Classify the multi-channel features based on the classifier to obtain the microseismic event classification results. Output the probability distribution of the microseismic event type through a fully connected layer and a softmax activation function, and take the maximum probability in the probability distribution as the final microseismic event classification result.

[0104] The present invention realizes the whole process of high-precision automatic classification of microseismic events with uneven channel numbers and non-fixed channel signal lengths through S1-S8.

[0105] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0106] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-channel microseismic event automatic classification method based on SNR signal selection, characterized in that: The following steps are involved: S1. Use the long-short time window ratio algorithm to segment the main part and noise part of each channel signal of the microseismic event; S2. Calculate the signal-to-noise ratio of each channel signal; S3, sort by signal-to-noise ratio and select the first N channels; S4, resampling the selected first N channels to unify the lengths of the N channel signals to a fixed length; S5. Build a deep learning model, which includes a SincNet layer, a convolutional layer, an IPPA attention module, a feature fusion module, and a classifier; after performing learnable bandpass filtering on N channel signals through the SincNet layer, perform shallow feature extraction through the convolutional layer; S6, perform multi-scale deep feature extraction on the shallow features of N channel signals through the IPPA attention module to capture the local-global features of each channel signal; S7, cross-channel fusion of the multi-scale deep features of the N channel signals output by the IPPA attention module is performed through the feature fusion module to obtain multi-channel features; S8. Classify the multi-channel features based on the classifier to obtain the classification results of the microseismic events.

2. The multi-channel microseismic event automatic classification method based on SNR signal selection according to claim 1 is characterized in that: S1 includes: Generate STA / LTA characteristic function, the expression is: Where STA represents the average energy in a short period of time, LTA represents the average energy in a long period of time, t is the current sample index, nsta is the STA window length, nlta is the LTA window length, a is the input signal, and i is the index in the window; After generating the STA / LTA characteristic function, the open threshold and closed threshold are set, and the segment where the STA / LTA characteristic function value is continuously higher than the open threshold and not lower than the closed threshold is determined as the main part of the valid signal, and the rest is classified as the noise part.

3. The multi-channel microseismic event automatic classification method based on SNR signal selection according to claim 1 is characterized in that: In S2, the signal-to-noise ratio calculation formula is: Among them, α is the main part or noise part of the signal, N signal is the length of the main part of the signal, N noise is the length of the signal-noise portion.

4. The multi-channel microseismic event automatic classification method based on SNR signal selection according to claim 1 is characterized in that: In S5, a SincNet layer, two convolutional layers, and an IPPA attention module constitute a feature extraction unit; The deep learning model contains N feature extraction units in total, and N channel signals are input into the N feature extraction units one by one for feature extraction.

5. The multi-channel microseismic event automatic classification method based on SNR signal selection according to claim 1 is characterized in that: In S5, the calculation formula of the SincNet layer is: y i [n]=x[n]*g i [n,f (i,1) ,f (i,2) ] g i [n,f (i,1) ,f (i,2) ]=2f (i,2) sinc(2πf (i,2) n)-2f (i,1) sinc(2πf (i,1) n) Among them, i is the output channel, that is, the i-th Sinc filter; n is the input signal index, f (i,1) 、f (i,2) represents the learnable lower cutoff frequency and upper cutoff frequency of the i-th filter; x[n] represents the one-dimensional original input signal, y i [n] represents the output signal of the i-th channel, g i [·] represents the filter function constructed based on the i-th Sinc function.

6. The multi-channel microseismic event automatic classification method based on SNR signal selection according to claim 1 is characterized in that: In S5, the channel signal passes through the SincNet layer and then performs maximum pooling with a kernel size of 3×1, layer normalization, and LeakyReLU activation. It then passes through two standard convolutional layers, followed by maximum pooling, layer normalization, and LeakyReLU activation to obtain shallow features.

7. The multi-channel microseismic event automatic classification method based on SNR signal selection according to claim 1 is characterized in that: In S6, the IPPA attention module consists of an adaptive average pooling submodule and an adaptive maximum pooling submodule connected in series; The adaptive average pooling submodule focuses on extracting the global statistical features of the input features and depicts the overall trend and stability of the signal through the adaptive average pooling operation; The adaptive max pooling submodule focuses on local high response areas and uses adaptive max pooling operations to capture peaks and sudden changes in input features.

8. The multi-channel microseismic event automatic classification method based on SNR signal selection according to claim 7 is characterized in that: Both the adaptive average pooling submodule and the adaptive maximum pooling submodule use a parallel multi-scale pyramid structure to extract hierarchical features of three scales from the shallow features of the input simultaneously; Among them, the first path uses 1×1 adaptive pooling to capture global features, the second path uses 2×1 adaptive pooling to extract medium-scale features, and the third path uses 4×1 adaptive pooling to focus on local detail features; The feature extraction results of each path are fused through the splicing operation and input into an MLP network containing two fully connected layers, and then the channel attention weights are generated through the sigmoid activation function; The generated channel attention weights are multiplied element-wise with the input shallow features to output deep features that fused the attention information.

9. The multi-channel microseismic event automatic classification method based on SNR signal selection according to claim 8, characterized in that: The process of IPPA attention module calculating input features is expressed as: V APP =σ(MLP(Concat[AAP1(F1),AAP2(F1),AAP4(F1)])) V AMP =σ(MLP(Concat[AAP1(F AAP ),AAP2(F AAP ),AAP4(F AAP )])) Among them, F1 is the shallow feature output by S6, AAP n represents the adaptive average pooling layer with kernel n in the adaptive average pooling submodule, F AAP represents the intermediate features output by the adaptive average pooling submodule, V AAP Represents the channel attention weight generated by the adaptive average pooling submodule; AMP n represents the adaptive maximum pooling layer with kernel n in the adaptive maximum pooling submodule, V AMP represents the channel attention weight generated by the adaptive maximum pooling submodule; Concat[·] represents the concatenation operation of adaptive average pooling layers or adaptive maximum pooling layers of different scales; MLP represents two fully connected layers, σ represents the sigmoid activation function, and F2 represents the deep features output by the adaptive maximum pooling submodule.

10. The multi-channel microseismic event automatic classification method based on SNR signal selection according to claim 1, characterized in that: In S7, the feature fusion module first concatenates the deep features of each channel, and then performs feature fusion through a convolutional layer. The number of convolution kernels in the convolution layer is 64×N, and the step size is 1. The deep features of different channels are fused through a 1×1 convolution kernel. The fused features are then processed by maximum pooling with a kernel of 3×1, layer normalization, and Leaky ReLU activation.

Images