Automatic side slope rolling stone micro-seismic signal detection method

Abstract

The invention belongs to the technical field of side slope micro-seismic signal analysis and identification, and relates to an automatic side slope rolling stone micro-seismic signal detection method, which comprises the following steps: acquiring signals by using a node seismograph, and establishing a database; acquiring a time-frequency diagram by adopting continuous wavelet transform; building a convolutional neural network introducing a double-attention mechanism, and obtaining a network weight parameter with a loss value small enough by training a loss value reduction function; carrying out abnormal signal pickup on the data detected in real time by using a de-noised MER-AIC method; and taking a pickup signal as the input of the convolutional neural network after continuous wavelet transform, and giving an identification result by a trained model to realize automatic detection. According to the method, abnormal signals can be fully automatically and accurately picked up and side slope rolling stone micro-seismic signals can be detected in data which are collected in real time in a rock side slope and mixed with various kinds of noise, intelligence and automation of real-time detection work of rolling stone events are achieved, dependence on manpower is reduced, and the recognition efficiency is improved.

Description

Technical Field

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

[0002] Rock landslide is a geological disaster caused by the overall sliding of the rock and soil mass along the slope under the action of gravity. It is also one of the most frequent and most harmful geological disasters in the world, and has the characteristics of wide distribution, strong concealment, and serious harm. At present, there are few monitoring equipment available under complex terrain conditions, and it is necessary to develop intelligent monitoring equipment to improve detection capabilities; the occurrence of rolling rocks and falling rocks in steep terrain is the most obvious precursor of geological disasters. The movement of rolling rocks can be used as a warning for large-scale rock landslides. This disaster precursor information can be used to conduct timely and accurate rock slope disaster detection, greatly reducing the loss of human life and property. Therefore, it is particularly important to develop an automatic detection method for rolling rock microseismic events that is high-precision, high-efficiency, and not easily affected by the signal-to-noise ratio.

[0003] Usually, the signals collected by cable-free node seismographs on rock slopes include not only rockfall microseismic signals, but also irrelevant signals such as natural noise signals, seismic signals, and large-scale mechanical microseismic signals. These irrelevant signals interfere with the identification of rockfall microseismic signals. To identify rockfall microseismic signals from the massive amount of detected data, manual identification is difficult and inefficient. At present, the recognition accuracy of some time-domain-based detection methods is still lacking. Therefore, it is necessary to analyze the collected signals in an automated way to realize the automatic detection of rockfall microseismic signals and improve the detection efficiency. Summary of the invention

[0004] The purpose of the present invention is to provide an automatic detection method for rock rolling microseismic signals on slopes based on abnormal event picking and introducing a double-attention mechanism convolutional neural network. Deep learning technology is applied to extract features from rock rolling microseismic signals and other types of microseismic signals, and a convolutional neural network microseismic signal recognition model introducing a double-attention mechanism is constructed and trained. By picking abnormal events from real-time detected slope microseismic signals, the convolutional neural network model is used to identify abnormal microseismic signals on the slope, thereby realizing automatic and intelligent detection of rock rolling microseismic signals on slopes, so as to solve the problems of difficulty in manual identification and low work efficiency.

[0005] The objective of the present invention is achieved through the following technical solutions: An automatic detection method for microseismic signals of rockfall on a slope comprises the following steps: 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; 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, based on the abnormal signal position index result obtained 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. It 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.

[0006] Furthermore, step A, the preprocessing process, includes removing the DC signal from the signal and removing noise using a filter.

[0007] Furthermore, in step B, the training set data is enhanced, including random horizontal flipping and random cropping.

[0008] Furthermore, in step C, the convolutional neural network model introducing 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, and then every 10 rounds of training, the learning rate is reduced to one tenth of the original.

[0009] Furthermore, in step C, a convolutional neural network model introducing a dual attention mechanism is constructed, which includes three parts: a convolution module, an attention module, and a classification module; The convolution module is used to extract local features of the signal after time-frequency transformation. It consists of three convolution blocks, each of which contains three CNN layers, three batch normalization layers (BatchNorm, BN) 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 by three sets of linear layers respectively. Q, K, 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: , In the above formula, Q, K, and V are query, key, and value respectively; is the weight matrix of 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 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, it converts the importance of the feature channel into a weight coefficient between 0 and 1 through the sigmoid function, and redistributes the weight of each channel to the input feature map through dot 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.

[0010] Furthermore, in step D, the step of picking up abnormal signals in each data segment using the MER-AIC method after denoising is as follows: Step D1, denoising the segment signal using the Curvelet method, wherein the denoising process specifically includes event sequence signal preprocessing, Curvelet transform, threshold processing and inverse Curvelet transform to obtain the denoised signal; Step D2, normalizing the segment so that the signal amplitude is within the range of [-1, 1]; Step D3, using the MER algorithm to identify the abnormal signal and the first arrival position, using two time windows with a length of 200 before and after to scan the signal, and calculating 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, where 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.

[0011] Furthermore, in step E, the time-frequency graph saved in step D is adjusted to a shape of 256*256*3, and is input into a convolutional neural network detection model that introduces a dual attention mechanism for further judgment, and the obtained results are saved, 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 input 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 a 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; the above steps are repeated until all time-frequency graphs are judged.

[0012] Compared with the prior art, the present invention has the following beneficial effects: The present invention adopts an appropriate method to reduce the interference of background noise in a complex environment of the slope, thereby highlighting abnormal signals and realizing more accurate detection of possible rock rolling microseismic signals; the convolutional neural network with a dual attention mechanism realizes the extraction of local features of the signal through a convolutional layer, picks up global long-distance features through three parallel multi-head self-attention layers, assigns weights to the three channel results through the SE attention layer and then fuses the features, which can effectively realize the picking up of local and global features of the signal; after the abnormal signal is picked up in the time domain, the convolutional neural network with a dual attention mechanism is combined for comprehensive judgment, which improves the detection accuracy, and realizes the intelligent and automated real-time detection of rock rolling events on the slope; the present invention can realize the fully automated and accurate picking up of abnormal signals and detection of rock rolling microseismic signals on the slope in data mixed with various types of noise collected in real time in the rock slope, realizes the intelligent and automated real-time detection of rock rolling events, reduces the dependence on manual labor, and improves the recognition efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0014] Figure 1 It is an overall flow chart of the automatic detection method of the microseismic signal of rock rolling on the slope of the present invention; Figure 2 The original graph and time-frequency graph of a section of rock-rolling microseismic signal and three sections of actual data of non-rock-rolling microseismic signal provided by the embodiment of the present invention; wherein (a) is a time-frequency domain graph of the rock-rolling microseismic signal, (b) is a time-frequency domain graph of the natural noise signal, (c) is a time-frequency domain graph of the large-scale mechanical microseismic signal, and (d) is a time-frequency domain graph of the earthquake signal; Figure 3 is a flow chart of denoising in abnormal signal picking provided by an embodiment of the present invention; Figure 4 is a neural network structure used in an embodiment of the present invention; Figure 5 is a structural diagram of a convolution module in a neural network provided by an embodiment of the present invention; Figure 6 is a structural diagram of the attention layer in a neural network provided by an embodiment of the present invention; Figure 7 It is a structural diagram of the SE self-attention layer in the attention layer provided by an embodiment of the present invention; Figure 8It is a structural diagram of the FC module in the neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The present invention will be further described below in conjunction with embodiments:

[0016] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0017] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0018] The automatic detection method of microseismic signals of rock rolling on slopes of the present invention comprises the following steps: Step A, use cable-free node seismographs to collect signals and establish a database; B, use continuous wavelet transform to obtain time-frequency diagrams, use 80% of the signals as training sets and 20% as validation sets; C, construct a convolutional neural network with 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 with a dual attention mechanism; D, use the denoised MER-AIC method to pick up abnormal signals from the real-time detected data; E, use the picked signal as the input of the convolutional neural network detection model after continuous wavelet transform, and give the recognition result to achieve automatic detection.

[0019] Specifically, the method includes the following steps: Step A, using a cable-free node seismograph to collect signals, including rolling stone microseismic signals, natural noise signals, earthquake signals and large-scale mechanical microseismic signals, and preprocessing the collected signal data to establish a database.

[0020] Specifically, the preprocessing process includes removing the DC signal and using a filter to remove noise.

[0021] Step B, using continuous wavelet transform to obtain time-frequency diagram, which is used to prepare training set and verification set for slope rockfall microseismic signal classification, and data enhancement processing is performed on the training set data.

[0022] Specifically, the training set data augmentation operations include random horizontal flipping and random cropping.

[0023] Step C, construct a convolutional neural network that introduces 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 that introduces a dual attention mechanism.

[0024] Specifically, 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.

[0025] Among them, the convolutional neural network model that introduces the dual attention mechanism includes three parts: convolution module, attention module and classification module.

[0026] The convolution module is used to extract local features of the signal after time-frequency transformation. It consists of three convolution blocks, each of which contains three CNN layers, three batch normalization layers (BatchNorm, BN) and three RELU activation functions.

[0027] The attention module consists of three parallel multi-head self-attention layers, an SE channel attention layer, a maximum pooling layer and a feature fusion layer.

[0028] The multi-head self-attention layer consists of three parts: 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 passed through three sets of linear layers to generate Q, K, and 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: , In the above formula, Q, K, and V are query, key, and value respectively; is the weight matrix of 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 a trainable linear projection matrix; Softmax normalizes the similarity to obtain the attention weight.

[0029] 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, it converts the importance of the feature channel into a weight coefficient between 0 and 1 through the sigmoid function, and redistributes the weight of each channel to the input feature map through dot multiplication and addition operations.

[0030] 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.

[0031] Step D, picking up signals from 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 the abnormal signal position index in each data segment.

[0032] Specifically, the steps of using the MER-AIC method after denoising to pick up abnormal signals in each data segment are as follows: Step D1, the segment signal is denoised using the Curvelet method. The denoising process is as follows: Figure 3 As shown, it specifically includes event sequence signal preprocessing, Curvelet transform, threshold processing and inverse Curvelet transform to obtain the 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. Among them, 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.

[0033] 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.

[0034] Specifically, the time-frequency graph saved in step D is adjusted to a shape of 256*256*3 and input into the convolutional neural network detection model that introduces the dual attention mechanism for further judgment, and the obtained results are saved, 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 input 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. When the probability of a 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; the above steps are repeated until all time-frequency graphs are judged.

[0035] If the probability of a rolling stone signal in the output of the neural network is greater than another probability, it is finally determined to be a rolling stone signal; then, the time of its occurrence is calculated; the above steps are repeated until all time-frequency graphs are judged, and the position information, time information and probability information of all judged rolling stone signals are written into a csv file for record preservation.

[0036] The interference noise accounts for a large proportion of the signals collected from actual slope microseismic events, and the monitoring environment is relatively harsh, so the signal-to-noise ratio of the acquired data is low. Therefore, an abnormal signal picking method is needed to quickly and accurately detect possible rockfall microseismic events in continuous waveform signals.

[0037] The location of abnormal microseismic signals is detected from the collected slope signals, and the continuous wavelet transform method is used to generate time-frequency diagrams of the abnormal microseismic signals. The convolutional neural network with a double attention mechanism is called to identify them and retain the final rockfall microseismic signal detection results.

[0038] Embodiment 1: A method for automatically detecting microseismic signals of rockfall on a slope, such as Figure 1 As shown, the following steps are included: Step A, using a cable-free node seismograph to collect microseismic 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; According to the frequency domain characteristics of the rockfall microseismic signal, the appropriate frequency band is selected to perform bandpass filtering on the collected signal to filter out the clutter and complex background noise of the slope that is significantly lower than the rockfall microseismic signal frequency band, thereby improving the signal-to-noise ratio. Figure 2 Middle (a) - Figure 2 As shown in (d), Figure 2 (a) is the time-frequency domain diagram of the rolling stone microseismic signal. Figure 2 (b) is the natural noise signal and the time-frequency domain diagram. Figure 2 Middle (c) is the time-frequency domain diagram of the large-scale mechanical microseismic signal; Figure 2 (d) is the time-frequency domain diagram of the seismic signal; the bandpass filter adopts a direct FIR bandpass filter designed with a Hanning window, with a low-frequency cutoff frequency of 50 Hz, a high-frequency cutoff frequency of 150 Hz, and an order of 128.

[0039] Step B, using continuous wavelet transform to prepare training set and verification set for classification of rockfall microseismic signals on slopes, performing data enhancement processing on the training set and verification set data, generating a signal time-frequency diagram of size 256*256*3 and saving it as input for the training model; Step C, construct a convolutional neural network that introduces 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 that introduces a dual attention mechanism.

[0040] A convolutional neural network is constructed based on the pytorch framework. Its overall network structure and the output size of feature maps at each level are as follows: Figure 4 As shown in Figure 2. The optimized convolutional neural network in this method consists of a convolution module, an attention module, and a classification module. Figure 5 As shown in Figure 1, the deep convolution module architecture consists of three convolution blocks, each of which contains three CNN layers, three batch normalization layers (BatchNorm, BN), and three RELU activation functions. Figure 6 As shown in Figure 1, the attention module consists of three parallel multi-head self-attention layers, SE channel attention layer, maximum pooling layer and a feature fusion layer. Figure 7 As shown in , the SE channel attention layer uses maximum pooling and average pooling to extract global features, generates channel weights through dimensionality reduction, dimensionality increase and sigmoid function, and redistributes the channel importance of the input feature map in a channel-by-channel weighted manner. Figure 8 As shown in Figure 1, the classification module consists of an FC module and a softmax function, where the FC module consists of a global average pooling layer, two Dropout layers, and three fully connected layers.

[0041] The neural network constructed by this method is mainly composed of two parts. The first part is the feature extraction part, which extracts the features of the training set data through the combination of convolution module and attention module.

[0042] Specifically, the network input is a 256*256*3 RGB image, and the first convolution layer is set as follows: 3 channels are input, 16 channels are output, the convolution kernel size is 3*3, the step size is 1 pixel, the padding is 1 pixel, and the output feature map size becomes 256*256*16. The output passes through the batch normalization layer and the RELU activation function and enters the next convolution layer.

[0043] The second convolutional layer is set as follows: 16 channels of input, 16 channels of output, 3*3 kernel size, 2 pixels of stride, 1 pixel of padding, and the output feature map size becomes 128*128*32. The output passes through the batch normalization layer and the RELU activation function and enters the next convolutional layer.

[0044] The third convolutional layer is set as follows: 32 channels are input, 32 channels are output, the convolution kernel size is 3*3, the step size is 1 pixel, the padding is 1 pixel, and the output feature map size becomes 128*128*32; after the output passes through the batch normalization layer and the RELU activation function, the Dropout unit is used to improve the generalization ability of the model and prevent overfitting. At this time, the feature map size becomes 128*128*32.

[0045] The fourth convolutional layer is set as follows: 32 channels are input, 64 channels are output, the convolution kernel size is 3*3, the step size is 2 pixels, the padding is 1 pixel, and the output feature map size becomes 64*64*64. The output passes through the batch normalization layer and the RELU activation function and enters the next convolutional layer.

[0046] The fifth convolutional layer is set as follows: 64 channels of input, 64 channels of output, 3*3 kernel size, 1 pixel step, 1 pixel padding, and the output feature map size becomes 64*64*64. The output passes through the batch normalization layer and the RELU activation function and enters the next convolutional layer.

[0047] The sixth convolutional layer is set as follows: 64 channels are input, 128 channels are output, the convolution kernel size is 3*3, the step size is 2 pixels, the padding is 1 pixel, and the output feature map size becomes 32*32*128; the output passes through the batch normalization layer and the RELU activation function and then passes through the Dropout unit to improve the generalization ability of the model and prevent overfitting. At this time, the feature map size becomes 32*32*128.

[0048] The seventh convolutional layer is set as follows: 128 channels are input, 128 channels are output, the convolution kernel size is 3*3, the step size is 1 pixel, the padding is 1 pixel, and the output feature map size becomes 32*32*128. The output passes through the batch normalization layer and the RELU activation function to enter the next convolutional layer.

[0049] The eighth convolutional layer is set as follows: 128 channels are input, 256 channels are output, the convolution kernel size is 3*3, the step size is 2 pixels, the padding is 1 pixel, and the output feature map size becomes 16*16*256. The output passes through the batch normalization layer and the RELU activation function and enters the next convolutional layer.

[0050] The ninth convolutional layer is set as follows: input 256 channels, output 256 channels, convolution kernel size is 3*3, step size is 1 pixel, padding is 1 pixel, and the output feature map size becomes 16*16*256.

[0051] The output passes through the batch normalization layer and the RELU activation function, and then passes through the Dropout unit to improve the generalization ability of the model and prevent overfitting. At this time, the feature map size becomes 16*16*256.

[0052] Next, the feature map output by the convolution module is flattened as the input of the multi-head self-attention layer to extract global features. The specific operation is: first, the feature map is flattened as the output of the three multi-head self-attention layer branches. 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, where path_h and patch_w represent the length and width of each block after the feature map is split, respectively, and num_heads represents the number of heads of the attention mechanism.

[0053] The outputs of the three multi-attention layer branches are folded to restore the feature map size to 16*16*256. The three feature icons are weighted through the SE channel attention layer respectively, and then the maximum pooling operation is performed to reduce the dimension to 8*8*256. Then, the features obtained from the three channels are fused to obtain a new feature map size of 8*8*256. The feature extraction of the data is completed.

[0054] 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 by the feature extraction part into a one-dimensional vector for the final classification.

[0055] Specifically, the feature map of size 8*8*256 obtained from the feature extraction part is globally average pooled to obtain a feature map of size 4*4*256, which is flattened to obtain a one-dimensional vector of length 4096. After the first fully connected layer, the vector length becomes 256, after the second fully connected layer, the vector length becomes 128, and after the third fully connected layer, the vector length becomes 2. The final softmax layer is used to normalize the obtained one-dimensional numerical vector of length 2 into a probability distribution vector. The first value in the probability vector represents the probability of being a rolling stone microseismic signal, and the second value represents the probability of being a noise signal (other types of microseismic signals). The construction of the convolutional neural network is now completed.

[0056] Step D, pick up the real-time detected seismic data, take every 3s of data as a data segment to ensure that the entire process of abnormal signal can be fully recorded, that is, each time segment contains 3000 sampling points as a sample (when the sampling rate of the collected signal is 1000 Hz), and use the Curvelet method to denoise the sample. The denoising process is as follows: Figure 3 As 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.

[0057] 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].

[0058] 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: 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.

[0059] 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.

[0060] 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.

[0061] 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.

[0062] 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 and save the signal time-frequency diagram of size 256*256*3, and use it as the input of the convolutional neural network detection model with dual attention mechanism introduced in step C. 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 a rolling stone signal is greater than the other probability, it is finally determined to be a rolling stone signal; then, calculate the time of its occurrence; repeat the above steps until all time-frequency diagrams are judged, and the recognition results are obtained to realize automatic detection, and the position information, time information and probability information of all judged rolling stone signals are written into a csv file for record and preservation.

[0063] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, 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; 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 by: Step C, constructing a convolutional neural network model that introduces a dual attention mechanism, including three parts: a convolution module, an attention module, and a 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 by three sets of linear layers respectively. Q, K, 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: ， In the above formula, Q, K, and V are query, key, and value respectively; is the weight matrix of 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 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, it converts the importance of the feature channel into a weight coefficient between 0 and 1 through the sigmoid function, and redistributes the weight of each channel to the input feature map through dot 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.

6. 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, using the MER algorithm to identify the abnormal signal and the first arrival position, using two time windows with a length of 200 before and after to scan the signal, and calculating 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.

7. 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