An automatic detection method for microseismic signals of rolling stones on slopes

Abstract

The present invention belongs to the technical field of slope microseismic signal analysis and identification, and relates to an automatic detection method for slope rockfall microseismic signals, including collecting signals using nodal seismographs and establishing a database; obtaining time-frequency diagrams by using continuous wavelet transform; building a convolutional neural network introducing a dual attention mechanism, reducing the loss value function through training, and obtaining network weight parameters with a sufficiently small loss value; picking up abnormal signals from the real-time detected data using the denoised MER-AIC method; taking the signals picked up after continuous wavelet transform as the input of the convolutional neural network, and giving the recognition result by the trained model to achieve automatic detection. This method can achieve the full-automatic and accurate picking up of abnormal signals and the detection of slope rockfall microseismic signals from the data mixed with various noises collected in real time in rocky slopes, realize the intelligence and automation of the real-time detection work of rockfall events, reduce the dependence on manual labor, and improve the recognition efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of slope microseismic signal analysis and identification, and particularly relates to an automatic detection method for slope rockfall microseismic signals based on abnormal event picking and introducing a dual-attention mechanism convolutional neural network. Background Art

[0002] Rock landslides are geological disasters caused by the overall sliding of slope rock and soil masses along the slope under the action of gravity. They are also one of the most frequently occurring and most harmful geological disasters globally, and are characterized by wide distribution, strong concealment, and serious harm. At present, there are few available monitoring devices under complex geomorphic conditions, and it is necessary to develop intelligent monitoring devices to improve the detection ability; the occurrence of rolling stones and falling rocks in high-steep terrains is the most obvious precursor of geological disasters. The movement of rolling stones can be used as a warning for large-scale rock landslides, and timely and accurate detection of rock slope disasters can be carried out using this disaster precursor information, greatly reducing human life and property losses. Therefore, it is particularly important to develop an automatic detection method for rolling stone microseismic events with high precision, high efficiency, and not easily affected by signal-to-noise ratio.

[0003] Generally, in the signals collected by cableless node seismographs on rock slopes, in addition to rolling stone microseismic signals, there are also natural noise signals, irrelevant signals such as seismic signals and large mechanical microseismic signals; these irrelevant signals interfere with the identification of rolling stone microseismic signals. To identify rolling stone microseismic signals from the vast amount of detected data, the manual identification method has a large identification difficulty and low work efficiency. And currently, the identification accuracy of some time-domain-based detection methods is still lacking. Therefore, it is necessary to analyze the collected signals in an automated manner to achieve the automatic detection of rolling stone microseismic signals and thus improve the detection efficiency. Summary of the Invention

[0004] The purpose of the present invention is to provide an automatic detection method for slope rockfall microseismic signals based on abnormal event picking and introducing a dual-attention mechanism convolutional neural network. By applying deep learning technology, features are extracted from rolling stone microseismic signals and other types of microseismic signals, a convolutional neural network microseismic signal recognition model introducing a dual-attention mechanism is constructed and trained, and abnormal events are picked from the slope microseismic signals detected in real time, and the convolutional neural network model is used to identify the slope abnormal microseismic signals, realizing the automation and intelligence of slope rolling stone microseismic signal detection to solve the problems of large manual identification difficulty and low work efficiency.

[0005] The purpose of the present invention is achieved by the following technical solutions:

[0006] An automatic detection method for slope rockfall microseismic signals, comprising the following steps:

[0007] Step A: Use a cable-free node seismograph to collect signals, including rolling stone microseismic signals, natural noise signals, seismic signals, and large machinery microseismic signals, and preprocess the collected signal data to establish a database.

[0008] Step B: Use continuous wavelet transform to obtain time-frequency diagrams for creating training sets and validation sets for classifying slope rolling stone microseismic signals, and perform data augmentation on the training set data.

[0009] Step C: Construct a convolutional neural network incorporating a dual attention mechanism, reduce the loss value function through training, obtain the network weight parameters with a loss value less than 0.1 during 50 rounds of training, and use the smallest of these network weight parameters as the parameters of the convolutional neural network detection model incorporating the dual attention mechanism.

[0010] Step D: Pick up signals from the real-time detected seismic data, take every 3s of data as a data segment, use the denoised MER-AIC method to pick up abnormal signals within each data segment, and obtain the abnormal signal position index within each data segment.

[0011] Step E: According to the abnormal signal position index results obtained in Step D, calculate the continuous wavelet transform of a total of 3000 sampling points before and after each signal index, form a two-dimensional time-frequency diagram and automatically save it, use it as the input of the convolutional neural network detection model incorporating the dual attention mechanism in Step C, and obtain the recognition result to achieve automatic detection.

[0012] Further, in Step A, the preprocessing process includes removing the DC signal from the signal and using a filter for denoising.

[0013] Further, in Step B, the training set data augmentation operations include random horizontal flipping and random cropping.

[0014] Further, in Step C, the convolutional neural network model incorporating the dual attention mechanism is trained according to the specified parameters, including using cross-entropy as the loss function, SGD as the network optimizer, a batch size of 12, a learning momentum of 0.9, an initial learning rate set to 0.001, and then the learning rate becomes one-tenth of the original every 10 rounds of training.

[0015] Further, in Step C, constructing the convolutional neural network model incorporating the dual attention mechanism includes three parts: a convolutional module, an attention module, and a classification module;

[0016] The convolutional module part is used to extract local features after the time-frequency transformation of the signal, and it consists of three convolutional blocks, each convolutional block containing three CNN layers, three batch normalization layers (BatchNorm, BN), and three RELU activation functions;

[0017] The attention module part is composed of three parallel multi-head self-attention layers, an SE channel attention layer, a max pooling layer, and a feature fusion layer;

[0018] Among them, the multi-head self-attention layer consists of three parts, namely layer normalization, multi-head self-attention, and MLP; MLP consists of a first fully connected layer, an activation function, and a second fully connected layer; the process of feature extraction by multi-head self-attention is as follows: the input feature X is respectively passed through three groups of linear layers to generate Q, K, and V: , ; Calculate the dot product similarity between Q and K to measure the correlation of each patch to other patches:

[0019] ,

[0020] The multi-head attention splits the input into multiple subspaces through separate calculations and captures features from different perspectives:

[0021] ,

[0022] In the above formulas, Q, K, and V are the query, key, and value respectively; is the weight matrix of the linear transformation; is the feature dimension of each attention head; h is the number of heads, and each head calculates attention independently; Concat is to concatenate the outputs of h heads; is the output of the i-th head; is the trainable linear projection matrix; Softmax normalizes the similarity to obtain the attention weights;

[0023] The SE channel attention layer extracts the global features of each channel by combining max pooling and average pooling, learns the importance distribution of each channel through dimensionality reduction and dimensionality increase operations, and finally converts the importance of the feature channels into weight coefficients between 0 and 1 through the sigmoid function, and redistributes the weights of each channel to the input feature map through dot product and addition operations;

[0024] The classification module part is composed of an FC module and a softmax function. Among them, the FC module is composed of a global average pooling layer, two Dropout layers, and three fully connected layers.

[0025] Furthermore, in step D, the steps of picking up abnormal signals using the denoised MER-AIC method within each data segment are as follows:

[0026] Step D1: Denoise the fragment signal using the Curvelet method. The denoising process specifically includes preprocessing the event sequence signal, Curvelet transform, threshold processing, and inverse Curvelet transform to obtain the denoised signal;

[0027] Step D2: Normalize the fragment so that the signal amplitude is within the range of [-1, 1];

[0028] Step D3: Use the MER algorithm to identify abnormal signals and the first arrival positions. Scan the signal with two time windows of length 200 before and after, and calculate the energy ratio of the signals within the two time windows; when the scanning time window passes through the first arrival position of the abnormal signal, the energy ratio will increase; when the energy ratio threshold is greater than 10, take the peak position of the energy ratio as the first arrival time. Open a time window of length 300 before and after this time point and use the AIC algorithm to pick up the accurate first arrival of the signal. Distinguish the effective signal from the noise based on the differences in statistical characteristics. At the first arrival position, the fitting degree of the effective signal and the noise is the lowest. Find the accurate first arrival time by calculating the minimum value of the cost function. Among them, the energy ratio discrimination formula of the MER algorithm and the cost function formula of the AIC algorithm are as follows:

[0029] Where, R represents the degree of energy change of the signal x within the window before and after the j-th moment, e represents the time window length, k is the index of the peak position of the energy ratio, is the variance of the data before the peak position of the energy ratio, is the variance of the data after the peak position of the energy ratio, N is the total length of the time window, and C is a constant.

[0030] Furthermore, in step E, adjust the shape of the time-frequency diagram saved in step D to 256*256*3, and input it into the convolutional neural network detection model with a dual attention mechanism for further judgment, and save the obtained results, including loading the network weight parameters with the smallest loss value in the 50 rounds of training obtained in step C into the convolutional neural network model, and then using the bicubic interpolation method to adjust the size of the two-dimensional time-frequency diagram of the abnormal microseismic signal to 256*256*3 to meet the network input requirements. Then input it into the convolutional neural network. After calculation, two outputs will be generated, respectively representing the probability that it is a rolling stone signal and the probability of other noises; when the probability of the rolling stone signal is greater than the other probability, finally determine that it is a rolling stone signal; then, calculate the time when it appears; repeat the above steps until all time-frequency diagrams are judged.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] The present invention uses an appropriate method to reduce the interference of background noise in complex slope environments, thereby highlighting abnormal signals and achieving more accurate detection of possible rolling stone microseismic signals. The convolutional neural network introducing a dual attention mechanism extracts local features of signals through convolutional layers, picks up long-distance global features through three parallel multi-head self-attention layers, assigns weights to the results of the three channels through the SE attention layer and then fuses the features. This method can effectively pick up local and global features of signals. After picking up abnormal signals in the time domain, combined with the convolutional neural network introducing a dual attention mechanism for comprehensive judgment, the detection accuracy is improved. Moreover, the intelligent and automated real-time detection of slope rolling stone events is realized. The present invention can automatically and accurately pick up abnormal signals and detect slope rolling stone microseismic signals from the data mixed with various noises collected in real time in rocky slopes, realize the intelligent and automated real-time detection of rolling stone events, reduce the dependence on manual work, and improve the recognition efficiency. Description of the Drawings

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other relevant drawings can also be obtained based on these drawings.

[0034] Figure 1 is the overall flowchart of the automatic detection method for slope rolling stone microseismic signals of the present invention;

[0035] Figure 2 are the original and time-frequency diagrams of actual data of a section of rolling stone microseismic signal and three sections of non-rolling stone microseismic signals provided by an embodiment of the present invention; among them, (a) is the rolling stone microseismic signal and its time-frequency diagram, (b) is the natural noise signal and its time-frequency diagram, (c) is the large mechanical microseismic signal and its time-frequency diagram, and (d) is the seismic signal and its time-frequency diagram;

[0036] Figure 3 is the flowchart of noise reduction in abnormal signal picking used in the embodiment of the present invention;

[0037] Figure 4 is the neural network structure provided by the embodiment of the present invention;

[0038] Figure 5 is the convolutional module structure diagram in the neural network provided by the embodiment of the present invention;

[0039] Figure 6 is the attention layer structure diagram in the neural network provided by the embodiment of the present invention;

[0040] Figure 7 It is the structural diagram of the SE self-attention layer in the attention layer provided by the embodiment of the present invention;

[0041] Figure 8 It is the structural diagram of the FC module in the neural network provided by the embodiment of the present invention. Specific embodiments

[0042] The present invention will be further described below in conjunction with embodiments:

[0043] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present invention, rather than limiting the present invention. In addition, it should be noted that for the sake of description, only the parts related to the present invention rather than all the structures are shown in the drawings.

[0044] It should be noted that: similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used for distinguishing descriptions, and cannot be understood as indicating or implying relative importance.

[0045] The automatic detection method for slope rolling stone microseismic signals of the present invention includes the following steps:

[0046] Step A, using a cableless node seismograph to collect signals and establish a database; B. Obtaining a time-frequency diagram by continuous wavelet transform, using 80% of the signals as the training set and 20% as the validation set; C. Constructing a convolutional neural network introducing a dual attention mechanism, reducing the loss value function through training, obtaining the network weight parameters with a loss value less than 0.1 in 50 rounds of training, and using the smallest of these network weight parameters as the parameters of the convolutional neural network detection model introducing the dual attention mechanism; D. Using the denoised MER-AIC method to pick up abnormal signals from the real-time detected data; E. Using the signals after continuous wavelet transform as the input of the convolutional neural network detection model, and giving the recognition result to achieve automatic detection.

[0047] Specifically, it includes the following steps: Step A, using a cableless node seismograph to collect signals, including rolling stone microseismic signals, natural noise signals, seismic signals, and large mechanical microseismic signals, and preprocessing the collected signal data to establish a database.

[0048] Specifically, the preprocessing process includes removing the DC signal from the signal and using a filter for denoising.

[0049] Step B, using continuous wavelet transform to obtain a time-frequency diagram for making the training set and validation set for classifying slope rolling stone microseismic signals, and performing data augmentation processing on the training set data.

[0050] Specifically, the training set data augmentation operations include random horizontal flipping and random cropping.

[0051] Step C: Construct a convolutional neural network introducing a dual attention mechanism, reduce the loss value function through training, obtain the network weight parameters with a loss value less than 0.1 in 50 rounds of training, and use the smallest network weight parameter as the parameter of the convolutional neural network detection model introducing the dual attention mechanism.

[0052] Specifically, the convolutional neural network model introducing the dual attention mechanism is trained according to the specified parameters, including using cross-entropy as the loss function, SGD as the network optimizer, a batch size of 12, a learning momentum of 0.9, an initial learning rate set to 0.001, and then the learning rate becomes one-tenth of the original every 10 rounds of training.

[0053] Among them, constructing the convolutional neural network model introducing the dual attention mechanism includes three parts: a convolutional module, an attention module, and a classification module.

[0054] The convolutional module part is used to extract the local features after the time-frequency transformation of the signal. It consists of three convolutional blocks, and each convolutional block contains three CNN layers, three batch normalization layers (BatchNorm, BN), and three RELU activation functions.

[0055] The attention module part is composed of three parallel multi-head self-attention layers, an SE channel attention layer, a max pooling layer, and a feature fusion layer.

[0056] Among them, the multi-head self-attention layer consists of three parts, namely layer normalization, multi-head self-attention, and MLP. MLP consists of a first fully connected layer, an activation function, and a second fully connected layer. The process of multi-head self-attention for feature extraction is as follows: The input feature X is respectively passed through three groups of linear layers to generate Q, K, and V: , calculate the dot product similarity of Q and K, which is used to measure the correlation of each patch to other patches:

[0057] ,

[0058] Multi-head attention calculates separately, splits the input into multiple subspaces, and captures features from different angles:

[0059] ,

[0060] In the above formulas, Q, K, and V are the query, key, and value respectively; is the weight matrix of the linear transformation; d is the feature dimension for each attention head; h is the number of heads, and each head calculates attention independently; Concat is to concatenate the outputs of h heads; is the output of the i-th head; W_i is the trainable linear projection matrix; Softmax normalizes the similarity to obtain the attention weights.

[0061] The SE channel attention layer extracts the global features of each channel by combining max pooling and average pooling. Through dimensionality reduction and upsampling operations, it learns the importance distribution of each channel. Finally, the importance of the feature channels is transformed into a weight coefficient between 0 and 1 through the sigmoid function, and the weights of each channel are redistributed to the input feature map through dot product and addition operations.

[0062] The classification module is partly composed of the FC module and the softmax function. Among them, the FC module is composed of a global average pooling layer, two Dropout layers, and three fully connected layers.

[0063] Step D: Pick up signals from the real-time detected seismic data. Take every 3s of data as a data segment, and use the denoised MER-AIC method to pick up abnormal signals within each data segment, and obtain the abnormal signal position index within each data segment.

[0064] Specifically, the steps of using the denoised MER-AIC method to pick up abnormal signals within each data segment are as follows:

[0065] Step D1: Denoise the segment signal using the Curvelet method. The denoising process is as Figure 3 shown, specifically including preprocessing of the event sequence signal, Curvelet transform, threshold processing, and inverse Curvelet transform to obtain the denoised signal;

[0066] Step D2: Normalize the segment so that the signal amplitude is within the range of [-1, 1];

[0067] Step D3: Use the MER algorithm to identify abnormal signals and first arrival positions. Scan the signal with two time windows of length 200 before and after, and calculate the energy ratio of the signals in the two time windows. When the scanning time window passes through the first arrival position of the abnormal signal, the energy ratio will increase. When the energy ratio threshold is greater than 10, take the peak position of the energy ratio as the first arrival time. Open a time window of length 300 before and after this time point and use the AIC algorithm to pick up the accurate first arrival of the signal. Distinguish effective signals from noise based on the differences in statistical characteristics. At the first arrival position, the fitting degree of the effective signal and noise is the lowest. Find the accurate first arrival time by calculating the minimum value of the cost function. The energy ratio discrimination formula of the MER algorithm and the cost function formula of the AIC algorithm are as follows:

[0068] Among them, R represents the degree of energy change of the signal x within the window before and after the j-th moment, e represents the time window length, k is the index of the peak position of the energy ratio, is the variance of the data before the peak position of the energy ratio, is the variance of the data after the peak position of the energy ratio, N is the total length of the time window, and C is a constant.

[0069] Step E: According to the abnormal signal position index result calculated in Step D, calculate the continuous wavelet transform of 3000 sampling points before and after each signal index, form a two-dimensional time-frequency diagram and save it automatically. Use it as the input of the convolutional neural network detection model with a dual attention mechanism introduced in Step C to obtain the recognition result and achieve automatic detection.

[0070] Specifically, adjust the shape of the time-frequency diagram saved in Step D to 256*256*3 and input it into the convolutional neural network detection model with a dual attention mechanism for further judgment, and save the obtained results. This includes loading the network weight parameters with the minimum loss value in the 50 rounds of training obtained in Step C into the convolutional neural network model. Then, use the bicubic interpolation method to adjust the size of the two-dimensional time-frequency diagram of the abnormal microseismic signal to 256*256*3 to meet the network input requirements. Then, input it into the convolutional neural network. After calculation, two outputs will be generated, respectively representing the probability that it is a rolling stone signal and the probability of other noises. When the probability of the rolling stone signal is greater than the other probability, finally determine that it is a rolling stone signal. Then, calculate the time when it appears. Repeat the above steps until all time-frequency diagrams are judged.

[0071] If in the output of the neural network, the probability of the rolling stone signal is greater than the other probability, finally determine that it is a rolling stone signal. Then, calculate the time when it appears. Repeat the above steps until all time-frequency diagrams are judged. Write the position information, time information, and probability information of all judged rolling stone signals into a csv file for recording and saving.

[0072] In the signals collected by the actual slope microseismic monitoring, the interference noise accounts for a large proportion, and the monitoring environment is relatively harsh, resulting in a low signal-to-noise ratio of the acquired data. Therefore, an abnormal signal picking method is needed to quickly and accurately detect possible rolling rock microseismic events in the continuous waveform signals.

[0073] Detect the position of abnormal microseismic signals from the slope signals collected, generate a time-frequency diagram for the abnormal microseismic signals using the continuous wavelet transform method, and call a convolutional neural network introducing a dual attention mechanism to discriminate them and retain the final detection results of the rolling rock microseismic signals.

[0074] Embodiment 1: An automatic detection method for slope rolling rock microseismic signals, as Figure 1 shown, includes the following steps:

[0075] Step A, use a cableless node seismograph to collect microseismic signals, including rolling rock microseismic signals, natural noise signals, seismic signals and large machinery microseismic signals, and preprocess the collected signal data to establish a database;

[0076] According to the frequency domain characteristics of the rolling rock microseismic signals, select an appropriate frequency band to perform band-pass filtering on the collected signals, filter out the clutter significantly lower than the frequency band of the rolling rock microseismic signals and the complex background noise of the slope, and improve the signal-to-noise ratio. As Figure 2 shown in (a)- Figure 2 shown in (d) of Figure 2 where (a) is the rolling rock microseismic signal and its time-frequency diagram, Figure 2 where (b) is the natural noise signal and its time-frequency diagram, Figure 2 where (c) is the large machinery microseismic signal and its time-frequency diagram; Figure 2 where (d) is the seismic signal and its time-frequency diagram; the band-pass filter uses a direct-form FIR band-pass filter designed with a Hanning window, with a low-frequency cut-off frequency of 50 Hz, a high-frequency cut-off frequency of 150 Hz, and an order of 128.

[0077] Step B, use the continuous wavelet transform to make the training set and validation set for the classification of slope rolling rock microseismic signals, perform data augmentation processing on the training set and validation set data, generate a signal time-frequency diagram with a size of 256*256*3 and save it as the input for training the model;

[0078] Step C, construct a convolutional neural network introducing a dual attention mechanism, reduce the loss value function through training, obtain the network weight parameters with a loss value less than 0.1 in 50 rounds of training, and use the smallest network weight parameter among them as the parameter of the convolutional neural network detection model introducing the dual attention mechanism.

[0079] Based on the pytorch framework, construct a convolutional neural network, and its overall network structure and the output size of each level of feature maps are as Figure 4As shown. In this method, the optimized convolutional neural network consists of a convolutional module, an attention module, and a classification module. Among them, as Figure 5 shown, the architecture of the deep convolutional module consists of three convolutional blocks, and each convolutional block contains three CNN layers, three batch normalization layers (BatchNorm, BN), and three RELU activation functions. As Figure 6 shown, the attention module is composed of three parallel multi-head self-attention layers, an SE channel attention layer, a max pooling layer, and a feature fusion layer. As Figure 7 shown, the SE channel attention layer uses max pooling and average pooling to extract global features, generates channel weights through dimensionality reduction, dimensionality increase, and the sigmoid function, and redistributes the channel importance of the input feature map through channel-wise weighting. As Figure 8 shown, the classification module is composed of an FC module and a softmax function, where the FC module is composed of a global average pooling layer, two Dropout layers, and three fully connected layers.

[0080] The neural network constructed by this method mainly consists of two major parts. The first part is the feature extraction part, which extracts the features of the training set data through the combination of the convolutional module and the attention module.

[0081] Specifically, the input of the network is an RGB image of 256*256*3. The first convolutional layer is set as follows: input 3 channels, output 16 channels, convolutional kernel size of 3*3, stride of 1 pixel, padding of 1 pixel, and the size of the output feature map becomes 256*256*16. This output passes through the batch normalization layer and the RELU activation function and enters the next convolutional layer.

[0082] The second convolutional layer is set as follows: input 16 channels, output 16 channels, convolutional kernel size of 3*3, stride of 2 pixels, padding of 1 pixel, and the size of the output feature map becomes 128*128*32. This output passes through the batch normalization layer and the RELU activation function and enters the next convolutional layer.

[0083] The third convolutional layer is set as follows: input 32 channels, output 32 channels, convolutional kernel size of 3*3, stride of 1 pixel, padding of 1 pixel, and the size of the output feature map becomes 128*128*32; after this output passes through the batch normalization layer and the RELU activation function, it passes through the Dropout unit to improve the generalization ability of the model and prevent overfitting. At this time, the size of the feature map becomes 128*128*32.

[0084] The fourth convolutional layer is set as follows: with 32 input channels, 64 output channels, a convolutional kernel size of 3*3, a stride of 2 pixels, a padding of 1 pixel, and the size of the output feature map becomes 64*64*64. This output passes through a batch normalization layer and a RELU activation function and enters the next convolutional layer.

[0085] The fifth convolutional layer is set as follows: with 64 input channels, 64 output channels, a convolutional kernel size of 3*3, a stride of 1 pixel, a padding of 1 pixel, and the size of the output feature map becomes 64*64*64. This output passes through a batch normalization layer and a RELU activation function and enters the next convolutional layer.

[0086] The sixth convolutional layer is set as follows: with 64 input channels, 128 output channels, a convolutional kernel size of 3*3, a stride of 2 pixels, a padding of 1 pixel, and the size of the output feature map becomes 32*32*128; after this output passes through a batch normalization layer and a RELU activation function, it then passes through a Dropout unit to improve the generalization ability of the model and prevent overfitting. At this time, the size of the feature map becomes 32*32*128.

[0087] The seventh convolutional layer is set as follows: with 128 input channels, 128 output channels, a convolutional kernel size of 3*3, a stride of 1 pixel, a padding of 1 pixel, and the size of the output feature map becomes 32*32*128. This output passes through a batch normalization layer and a RELU activation function and enters the next convolutional layer.

[0088] The eighth convolutional layer is set as follows: with 128 input channels, 256 output channels, a convolutional kernel size of 3*3, a stride of 2 pixels, a padding of 1 pixel, and the size of the output feature map becomes 16*16*256. This output passes through a batch normalization layer and a RELU activation function and enters the next convolutional layer.

[0089] The ninth convolutional layer is set as follows: with 256 input channels, 256 output channels, a convolutional kernel size of 3*3, a stride of 1 pixel, a padding of 1 pixel, and the size of the output feature map becomes 16*16*256.

[0090] After this output passes through a batch normalization layer and a RELU activation function, it passes through a Dropout unit to improve the generalization ability of the model and prevent overfitting. At this time, the size of the feature map becomes 16*16*256.

[0091] Next, the feature map output by the convolutional module is flattened and used as the input of the multi-head self-attention layer to extract global features. The specific operation is as follows: First, after flattening the feature map, it is used as the output of three branches of the multi-head self-attention layer. In the first branch, path_h is 2, patch_w is 2, and num_heads is 2. In the second branch, path_h is 2, patch_w is 2, and num_heads is 3. In the first branch, path_h is 2, patch_w is 2, and num_heads is 4. Here, path_h and patch_w respectively represent the length and width of each block after the feature map is split, and num_heads represents the number of heads of the attention mechanism.

[0092] The outputs of the three multi-attention layer branches are folded to restore the feature map to a size of 16*16*256. The three feature maps are respectively assigned weights through the SE channel attention layer, and then the feature map size becomes 8*8*256 after max-pooling for dimensionality reduction. Then, the features obtained from the three channels are fused to obtain a new feature map with a size of 8*8*256. Thus, the feature extraction of the data is completed.

[0093] The second part is the classification part. First, a global average pooling is performed, and then the signal is classified into two categories through a fully connected layer and a softmax layer. The main function of the fully connected layer is to flatten the two-dimensional feature map obtained in the feature extraction part into a one-dimensional vector for the final classification.

[0094] Specifically, the feature map with a size of 8*8*256 obtained in the feature extraction part is globally averaged and pooled to obtain a feature map with a size of 4*4*256, which is flattened to obtain a one-dimensional vector with a length of 4096. After passing through the first fully connected layer, the vector length becomes 256. After passing through the second fully connected layer, the vector length becomes 128. After passing through the third fully connected layer, the vector length becomes 2. The final softmax layer is used to normalize the obtained one-dimensional numerical vector with a length of 2 into a probability distribution vector. The first value in this probability vector represents the probability that it is a rolling rock microseismic signal, and the second value represents the probability that it is a noise signal (other types of microseismic signals). Thus, the construction of the convolutional neural network is completed.

[0095] Step D: Pick up the signals from the real-time detected seismic data. Take every 3s of data as a data segment to ensure that the entire process of abnormal signals can be completely recorded. That is, each time segment contains 3000 sampling points as a sample (when the sampling rate of the collected signal is 1000 Hz). The Curvelet method is used to denoise this sample. The denoising process is as Figure 3As shown, it specifically includes event sequence signal preprocessing, Curvelet transform, threshold processing and inverse Curvelet transform to obtain the denoised signal. The denoised data fragments are used to pick up abnormal signals using the MER-AIC method, and the abnormal signal position index is obtained in each data fragment.

[0096] The real-time detection data is divided into time segments, each segment lasting 3 seconds, and the segments are normalized so that the signal amplitude is in the range of [-1, 1].

[0097] The MER algorithm is used to identify abnormal valid signals and their first arrival positions. The signals are scanned using two time windows of 200 in length, and the energy ratio of the signals in the two time windows is calculated. When the scanning time window passes through the first arrival position of the abnormal signal, the energy ratio will increase. When the energy ratio threshold is greater than 10, the peak position of the energy ratio is taken as the first arrival time. A time window of 300 in length is opened before and after this time point, and the AIC algorithm is used to pick up the precise first arrival of the signal. The valid signal and noise are distinguished based on the difference in statistical characteristics. The fitting degree of the valid signal and noise at the first arrival position is the lowest, and the precise first arrival time is found by calculating the minimum value of the cost function. Among them, the energy ratio discrimination formula of the MER algorithm and the cost function formula of the AIC algorithm are as follows:

[0098] Among them, R represents the energy change degree of signal x in the window before and after time j, e represents the time window length, k is the index of the peak position of the energy ratio, is the variance of the data before the peak position of the energy ratio, is the variance of the data after the peak position of the energy ratio, N is the total length of the time window, and C is a constant.

[0099] The precise first arrival of the abnormal signal is recorded, and the continuous wavelet transform is used for the sampling points in the segment to generate a signal time-frequency diagram of size 256*256*3 and save it.

[0100] When the MER algorithm is used to make judgments using the energy ratio, if the energy ratio threshold cannot be reached, it is determined that there is no abnormal event in this segment, and the AIC algorithm is no longer used to pick up signals, and the abnormal event of the next segment is directly picked up.

[0101] Taking into account the situation where abnormal signals appear in two time segments, when the first arrival time index of the segment is greater than 2700, the segment is no longer considered as the abnormal signal part, and the following operations are performed: the last 500 points of the segment are intercepted as segment a, and when intercepting the next segment, only 2500 sampling points are intercepted as segment b, and segment a and segment b are sequentially connected to form the next segment for abnormal signal detection.

[0102] Step E: According to the abnormal signal position index result calculated in Step D, calculate the continuous wavelet transform of 3000 sampling points before and after each signal index, generate a signal time-frequency diagram with a size of 256*256*3 and save it. Use this as the input of the convolutional neural network detection model with a dual attention mechanism introduced in Step C. After calculation, two outputs will be generated, respectively representing the probability that it is a rolling stone signal and the probability of other noises. When the probability of the rolling stone signal is greater than the other probability, it is finally determined to be a rolling stone signal. Then, calculate the time when it appears. Repeat the above steps until all time-frequency diagrams are judged, and the recognition result is obtained to achieve automatic detection. Write the position information, time information, and probability information of all judged rolling stone signals into a csv file for recording and saving.

[0103] Note that the above is only the preferred embodiment of the present invention and the technical principles applied. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described here. Various obvious changes, re-adjustments, and substitutions can be made by those skilled in the art without departing from the protection scope of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments. Without departing from the concept of the present invention, more other equivalent embodiments can be included, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for automatically detecting microseismic signals of rockfall on slopes, characterized in that: The following steps are involved: Step A, using a cable-free node seismograph to collect signals, including rolling stone microseismic signals, natural noise signals, seismic signals and large-scale mechanical microseismic signals, and preprocessing the collected signal data to establish a database; Step B, using continuous wavelet transform to obtain time-frequency diagrams, which are used to prepare training sets and validation sets for classification of rockfall microseismic signals on slopes, and performing data enhancement processing on the training set data; Step C, constructing a convolutional neural network that introduces a dual attention mechanism, reducing the loss value function through training, obtaining the network weight parameter with a loss value less than 0.1 in 50 rounds of training, and using the smallest network weight parameter as the parameter of the convolutional neural network detection model that introduces a dual attention mechanism; The convolutional neural network model with dual attention mechanism is constructed, which includes three parts: convolution module, attention module and classification module. The convolution module is used to extract local features of the signal after time-frequency transformation, and is composed of three convolution blocks, each of which contains three CNN layers, three batch normalization layers and three RELU activation functions; The attention module consists of three parallel multi-head self-attention layers, SE channel attention layer, maximum pooling layer and a feature fusion layer; Among them, the multi-head self-attention layer consists of three parts, namely layer normalization, multi-head self-attention and MLP; MLP consists of the first fully connected layer, the activation function and the second fully connected layer; the process of feature extraction by multi-head self-attention is: the input feature X is generated through three sets of linear layers respectively Q, K, V: Q = XW Q , K = XW K , V = XW V ; Calculate the dot product similarity of Q and K to measure the relevance of each patch to other patches: Multi-head attention splits the input into multiple subspaces through separate calculations to capture features from different angles: MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O , In the above formula, Q, K, and V are query, key, and value respectively; W Q , W K , W V is the weight matrix of linear transformation; d k is the feature dimension of each attention head; h is the number of heads, and each head calculates attention independently; Concat is to concatenate the outputs of h heads; head i is the output of the ith head; W O is a trainable linear projection matrix; Softmax normalizes the similarity to obtain the attention weight; The SE channel attention layer extracts the global features of each channel by combining maximum pooling and average pooling. It learns the importance distribution of each channel through dimensionality reduction and dimensionality increase operations. Finally, the importance of the feature channel is converted into a weight coefficient between 0 and 1 through the sigmoid function. The weight of each channel is redistributed to the input feature map through point multiplication and addition operations. The classification module is composed of an FC module and a softmax function, wherein the FC module is composed of a global average pooling layer, two Dropout layers and three fully connected layers; Step D, picking up signals from the real-time detected seismic data, taking every 3 seconds of data as a data segment, picking up abnormal signals in each data segment using the denoised MER-AIC method, and obtaining an abnormal signal position index in each data segment; In step E, according to the abnormal signal position index result calculated in step D, the continuous wavelet transform of 3000 sampling points before and after each signal index is calculated to form a two-dimensional time-frequency diagram and save it automatically, which is used as the input of the convolutional neural network detection model that introduces the dual attention mechanism in step C to obtain the recognition result for automatic detection.

2. The automatic detection method of microseismic signals of rockfall on slope according to claim 1 is characterized by: Step A, the preprocessing process, includes removing the DC signal from the signal and using a filter to remove noise.

3. The automatic detection method of microseismic signals of rockfall on slope according to claim 1 is characterized by: Step B: training set data augmentation operations, including random horizontal flipping and random cropping.

4. The automatic detection method of microseismic signals of rockfall on slope according to claim 1 is characterized by: In step C, the convolutional neural network model with the dual attention mechanism is trained according to the specified parameters, including cross entropy as the loss function, SGD as the network optimizer, batch size 12, learning momentum 0.9, and the initial learning rate is set to 0.

001. After every 10 rounds of training, the learning rate is reduced to one tenth of the original value.

5. The automatic detection method of microseismic signals of rockfall on slope according to claim 1 is characterized in that: Step D, the steps of picking abnormal signals in each data segment using the MER-AIC method after denoising are as follows: Step D1, denoising the segment signal using the Curvelet method, wherein the denoising process specifically includes signal preprocessing, Curvelet transform, threshold processing and inverse Curvelet transform to obtain a denoised signal; Step D2, normalizing the segment so that the signal amplitude is within the range of [-1, 1]; Step D3, use the MER algorithm to identify the abnormal signal and the first arrival position, use two time windows with a length of 200 before and after to scan the signal, and calculate the energy ratio of the signal in the two time windows; when the scanning time window passes through the first arrival position of the abnormal signal, the energy ratio will increase; when the energy ratio threshold is greater than 10, the peak position of the energy ratio is taken as the first arrival time, and a time window with a length of 300 is opened before and after this time point to use the AIC algorithm to pick up the precise first arrival of the signal, and distinguish the effective signal from the noise according to the difference in statistical characteristics. At the first arrival position, the fitting degree of the effective signal and the noise is the lowest, and the precise first arrival time is found by calculating the minimum value of the cost function; wherein, the MER algorithm energy ratio discrimination formula and the AIC algorithm cost function formula are as follows: Among them, R represents the energy change degree of signal x in the window before and after time j, e represents the time window length, k is the index of the peak position of the energy ratio, is the variance of the data before the peak position of the energy ratio, is the variance of the data after the peak position of the energy ratio, N is the total length of the time window, and C is a constant.

6. The automatic detection method of microseismic signals of rockfall on slope according to claim 1 is characterized by: Step E, adjusting the shape of the time-frequency graph saved in step D to 256*256*3, inputting it into the convolutional neural network detection model that introduces the dual attention mechanism for further judgment, and saving the obtained results, including loading the network weight parameter with the smallest loss value in 50 rounds of training obtained in step C into the convolutional neural network model, and then using the bicubic interpolation method to adjust the size of the two-dimensional time-frequency graph of the abnormal microseismic signal to 256*256*3 to meet the network input requirements, and then inputting it into the convolutional neural network. After calculation, two outputs will be generated, representing the probability of it being a rolling stone signal and the probability of other noises respectively; when the probability of the rolling stone signal is greater than the other probability, it is finally determined to be a rolling stone signal; then, the time of its occurrence is calculated; and the above steps are repeated until all time-frequency graphs are judged.

Images