A method for identifying surrounding rock stress stages

Abstract

The present invention relates to the field of computer vision technology, and discloses a method for identifying surrounding rock stress stages, which adopts a neural network including N feature fusion modules, each of which includes a first branch and a second branch, wherein the first branch is used for extracting local features, and the second branch is used for extracting global features; wherein N≥1 and N is an integer. The present invention solves the problems of high equipment cost, high labor cost, low efficiency, etc. in the prior art.

Description

Technical Field

[0001] The invention relates to the technical field of computer vision, and in particular to a method for identifying surrounding rock stress stages. Background Art

[0002] In the engineering field, rock burst is a very common phenomenon. The main reason for its formation is that the accumulation of ground stress inside the rock exceeds the strength and elasticity of the rock and other brittle materials, destroying the original internal equilibrium state and leading to the release of energy in the rock and other brittle materials. Among them, rock burst can be divided into different degrees according to different levels of released energy. Mild rock burst will cause rock flakes to fall off, and severe rock burst will cause earthquakes, tunnel collapse, etc.

[0003] In order to avoid the occurrence of rock bursts, researchers have conducted a large number of related experiments based on existing materials and actual conditions, and based on the stress-strain curves of brittle materials such as rocks ( Figure 1 ), summarized the four stages of stress on the material before rock burst, namely, the compression stage of voids and cracks (OA segment), the stage of stable development from elastic deformation to micro-fractures (AB segment), the stage of unstable fracture development (BC segment) and the stage after destruction (CD segment). Among them, in the OA segment, the curve is concave as a whole, at this time the pores and cracks of the rock are compressed, the density increases, and the elastic modulus increases. In the AB segment, the line approaches a straight line, at this time the elastic deformation and micro-cracks of the sample begin to appear. In the BC segment, the curve is convex as a whole, at this time the local pores and cracks in the sample are connected, and micro-cracks appear. In the CD segment, the curve shows a downward trend, at this time the macroscopic destruction of the sample occurs. The occurrence of rock burst is at the critical point C between the third and fourth stages, at this time the internal stress of the rock reaches its peak, suddenly releasing a large amount of energy, leading to the emergence of dangerous situations.

[0004] Therefore, it is very important to identify and detect the stage of internal stress in rocks in the engineering field. Currently, the commonly used detection methods include acoustic emission technology, microseismic monitoring technology, RA-AF method, etc. However, these methods include but are not limited to high costs, low detection timeliness, low accuracy and other problems:

[0005] AE method:

[0006] That is, the acoustic emission identification method. Specifically, the degree of damage to the surrounding rock is distinguished through acoustic emission signals. However, the main problem is that for acoustic emission signals, the types and components of different surrounding rocks are different, and their acoustic emission signals are also different. Therefore, the type of surrounding rock needs to be clarified before identification, which increases the workload and reduces the timeliness and accuracy of identification.

[0007] Microseismic monitoring method:

[0008] Microseismic monitoring technology is essentially based on acoustic emission. It uses sensors to capture the sound signals generated when the interface in the rock is damaged or fractured, and then records and analyzes them with microseismic detectors to diagnose the location and stage of the fracture. However, the microseismic monitoring system is expensive, and using the system for monitoring requires a high cost, which is not friendly to surrounding rock stress monitoring in tunnel scenarios.

[0009] RA-AF analysis method:

[0010] RA-AF analysis is also a method for distinguishing crack types based on acoustic emission parameters, that is, the ratio of the two acoustic emission parameters, RA value (rise time / amplitude) and average frequency AF value (ring count / duration), is used to distinguish tensile cracks from shear cracks. The inherent tensile strength of the crack causes the movement of the two sides of the crack to produce an acoustic emission signal waveform with a short rise time and a high frequency. The acoustic emission signal waveform produced by shear rupture usually has a longer waveform, resulting in a longer rise time and a lower signal frequency. This may be because most of the energy is transmitted in the form of shear waves, which are slower and have a larger delay in the maximum peak of the waveform compared to the initial longitudinal wave arrival. Many scholars have used the RA-AF analysis method to classify tensile cracks and shear cracks in materials such as concrete, fiber composites, and rocks. The RA-AF analysis method can effectively distinguish the stress level of the surrounding rock under certain conditions. However, the coordinate ratio between the two parameters is not fixed. If the ratio is different, the classification results will also be different, and the normal and movement direction of the crack surface cannot be judged. Therefore, it is necessary to develop a more effective method to judge the crack type. In addition, during the entire process from pressure to destruction of brittle materials, due to different directions and angles, stress will manifest in different forms such as tensile failure and shear failure, so these failures need to be distinguished, resulting in a greater workload.

[0011] Recognition method based on deep learning:

[0012] In recent years, based on the development of deep learning technology, there have been new breakthroughs in the field of computer vision. There have been great achievements in classification, recognition, segmentation and other fields. Therefore, it is possible to consider introducing deep learning methods into the field of surrounding rock stress stage identification.

[0013] At present, there are two mainstream deep learning methods. One is a model based on convolutional neural network (CNN), and the other is a model based on visual transformer. Convolutional neural network has a simple structure, low computational overhead, and is easy to train. Due to its specific structure, the visual transformer model can capture contextual information within and between targets by establishing an interaction mechanism between network modules. Compared with traditional convolutional neural networks, it often has higher performance. However, this type of model has a complex structure, a large number of parameters, and high computational cost. It cannot be trained and inferred on some edge devices. It can only transmit information to the server through the network for calculation and recognition. In this way, the timeliness of the system cannot be guaranteed, and dangerous situations cannot be warned and handled immediately. It is not friendly to the engineering field.

[0014] Distinguishing the degree of damage to the surrounding rock is an important issue in the engineering field. Predicting the blasting nodes of the surrounding rock is a key technology that researchers focus on. Compared with traditional detection methods, this invention uses deep learning to distinguish the degree of damage to the surrounding rock, issue warnings based on different degrees, conduct risk assessments for impending rockbursts, and issue corresponding warnings based on different risk indexes.

[0015] The current traditional detection methods have the following main problems:

[0016] 1. High equipment cost, high labor cost, and low efficiency: If you want to monitor the stress stage of the surrounding rock in the tunnel, you need to deploy a large number of sensors, build a dedicated network, and rent a server. This greatly increases the cost of monitoring, which is not worth the cost. At the same time, due to the differences in surrounding rock samples, it is necessary to manually collect and organize surrounding rock samples in different tunnels, which increases labor costs and cannot be popularized in the tunnels used.

[0017] Secondly, the efficiency of traditional methods is not high. Since traditional methods need to use acoustic emission to capture the sound signals emitted when the surrounding rock breaks, and then send the collected data to proprietary instruments for monitoring, it requires a lot of manpower and material resources, which wastes resources. Finally, in addition to low efficiency and poor accuracy, traditional monitoring methods rely on manual judgment, resulting in weak learning ability. It takes a long time to find some experience to correct previous mistakes, and it takes a long time to learn some useful experience, which greatly increases the learning cost.

[0018] 2. The existing model has poor accuracy: The core of deep learning is to extract the features of images through convolutional neural networks and summarize the features of images of different categories. However, for traditional convolutional neural networks, the model cannot extract the contextual semantic information in the image, which makes it impossible to continue to improve the accuracy of the model after a certain period of training, and encounters a bottleneck.

[0019] 3. Low real-time performance of the model: Since it is impossible to extract contextual features from images, many works use transformers to improve the accuracy of the model. However, the transformer has too many parameters and consumes a lot of computing resources. It cannot be used on some edge devices and requires the help of networks and servers, which greatly reduces timeliness and does not play a role in timely warning. Summary of the invention

[0020] In order to overcome the deficiencies of the prior art, the present invention provides a method for identifying surrounding rock stress stages, which solves the problems of high equipment cost, high labor cost, low efficiency, etc. in the prior art.

[0021] The technical solution adopted by the present invention to solve the above problems is:

[0022] A surrounding rock stress stage identification method adopts a neural network including N feature fusion modules, each feature fusion module includes a first branch and a second branch, the first branch is used for extracting local features, and the second branch is used for extracting global features; wherein N≥1 and N is an integer.

[0023] As a preferred technical solution, the following steps are included:

[0024] S1, signal acquisition: collecting acoustic emission signal data at different stress stages;

[0025] S2, signal processing and data set construction: converting acoustic emission signal data into image data;

[0026] S3, data enhancement and data preprocessing: perform data enhancement and data preprocessing on image data;

[0027] S4, model building: building a neural network model;

[0028] S5, Training and Testing: Training and testing the neural network model.

[0029] As a preferred technical solution, step S2 includes the following steps:

[0030] S21, construct a mother wavelet, then generate multiple sub-wavelets by scaling and translating, and then superimpose different sub-wavelets to match the input signal. Decompose the acoustic emission signal through continuous wavelet transform to obtain the time-frequency diagram of the acoustic emission signal. The continuous wavelet transform formula is as follows:

[0031]

[0032] Where a represents the scale parameter, b represents the displacement parameter, f(t) represents the original signal, t represents the time variable, and Wf (a, b) represents the continuous wavelet transform coefficient of f(t) under a and b, ψ ab represents the basic wavelet function, represents f(t) and ψ ab The inner product of Represents ψ ab The scaled and shifted wavelet function;

[0033] S22, determining parameters, the determined parameters include: signal length f len , sampling frequency sample rate , mother wavelet function, center frequency sliding step cenfre step ;

[0034] S23, calculate the maximum center frequency Set the current center frequency cenfre now , initialize the time-frequency matrix S;

[0035] S24, constructing a wavelet curve according to the center frequency and the wavelet function, and then convolving it with the original signal to obtain the time distribution vector of the current frequency, and updating the time-frequency matrix;

[0036] S25, determine whether the current center frequency is greater than the maximum center frequency, if so, output the time-frequency matrix S; otherwise, update the current frequency cenfre now =cenfre now +cenfre step , then jump back to step S21;

[0037] S26, creating a data set using all signals that have completed continuous wavelet transform.

[0038] As a preferred technical solution, step S3 includes the following steps:

[0039] S31, randomly cropping the image;

[0040] S32, enhancing the image by means of a random flip function;

[0041] S33, converting the image into a tensor;

[0042] S34, normalize the tensor.

[0043] As a preferred technical solution, in step S4, the neural network model includes an embedding layer, a first LightNet block, a first merging layer, a second LightNet block, a second merging layer, a third LightNet block, a third merging layer, a fourth LightNet block, a convolutional layer, a global pooling layer, and an FC function layer connected in sequence.

[0044] As a preferred technical solution, the first LightNet block, the second LightNet block, the third LightNet block, and the fourth LightNet block all include a first branch and a second branch connected in parallel, the first branch is used to extract local features, the first branch includes a Pconv layer, a first convolutional layer, and a second convolutional layer connected in sequence, and the second branch includes a circular convolution module for extracting global features.

[0045] As a preferred technical solution, the circular convolution module performs the following processing:

[0046] For a feature vector of dimension CxHxW, first initialize a position coding vector of dimension CxBx1, align the position coding vector of CxBx1 with the feature vector of CxHxW through a bilinear interpolation function to obtain a position coding vector of CxHx1, then align the position coding vector of CxHx1 with the original vector through horizontal expansion to obtain a position coding vector of CxHxW, add and fuse the feature vector of CxHxW with the position coding vector of CxHxW to obtain a feature vector with position coding information; then vertically concatenate the feature vector with position coding information to obtain a vector of Cx2HxW, use a convolution kernel of dimension CxHx1 for convolution, and finally obtain a feature vector of dimension CxHxW to achieve global information exchange in the vertical direction;

[0047] as well as,

[0048] For a feature vector of dimension CxHxW, first initialize a position coding vector of dimension Cx1xB, align the position coding vector of Cx1xB with the feature vector of CxHxW through a bilinear interpolation function to obtain the position coding vector of Cx1xW, then align the position coding vector of Cx1xW with the original vector through vertical expansion to obtain the position coding vector of CxHxW, add and fuse the feature vector of CxHxW with the position coding vector of CxHxW to obtain a feature vector with position coding information; then horizontally splice the feature vector with position coding information to obtain a vector of CxHx2W, use a convolution kernel of dimension Cx1xW for convolution, and finally obtain a feature vector of dimension CxHxW to achieve global information exchange in the horizontal direction.

[0049] As a preferred technical solution, in step S5, the optimizer uses the Adam optimizer during the training process, the learning rate adopts a polynomial decay strategy, and a linear preheating method is used to increase the learning rate.

[0050] As a preferred technical solution, in step S5, the loss function is:

[0051]

[0052] Where p = [p0…p C-1 ] is a probability distribution, each element p i Indicates the probability that the sample belongs to the i-th category; y = [y0…y C-1 ] is the unique hot encoding representation of the sample label. When the sample belongs to the i-th category, y i =1, otherwise y i =0; C is the sample label.

[0053] As a preferred technical solution, step S1 includes the following steps:

[0054] S11, preparation of different rock samples;

[0055] S12, compression test;

[0056] S13, collecting acoustic emission signal data;

[0057] S14, collating the acoustic emission signal data.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] (1) The present invention uses a deep learning method. For different types of surrounding rocks, the corresponding stress curves are different, so the timing of blasting is inconsistent. Therefore, it is necessary to know the type of surrounding rock in advance to make a reasonable prediction. However, in actual applications, the type of surrounding rock of a tunnel is often not fixed. It is a mixed type and there is rarely a single type. Therefore, traditional methods can rarely achieve accurate prediction;

[0060] (2) Based on the traditional CNN model, the present invention is inspired by the transformer and uses global circular convolution (GCConv) to extract global features, which can capture the contextual information in the image and improve the recognition accuracy of the model;

[0061] (3) The present invention reduces the number of parameters by introducing local convolution (PConv) to replace ordinary conventional convolution. At the same time, compared with the tranformer structure, the use of GCConv also reduces the number of parameters while achieving the same performance, and the inference speed of the model can be greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is the rock stress-strain curve;

[0063] Figure 2 Schematic diagram of the device for conducting destruction experiments on rocks;

[0064] Figure 3 It is the spectrum diagram of the collected sound signal;

[0065] Figure 4 is the spectrum diagram after CWT transformation;

[0066] Figure 5 It is a schematic diagram of the main framework of the neural network model;

[0067] Figure 6 Schematic diagram of the core structure of the feature fusion module. DETAILED DESCRIPTION

[0068] The present invention will be further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0069] Example 1

[0070] like Figures 1 to 6 As shown, the present invention introduces the deep learning method into the field of surrounding rock stress stage identification and detection. The deep neural network does not need to pay attention to the specific category of the specific surrounding rock, but directly divides the stress stage it is in, thereby improving efficiency and reducing workload. At the same time, for the deep learning network model, improvements are made on the basis of the traditional algorithm, and a surrounding rock stress identification method based on a lightweight neural network is designed. Compared with traditional methods, this method can efficiently identify and classify the degree of surrounding rock damage in a relatively short time, and give early warning of impending rock bursts. This method combines the advantages of traditional convolutional neural networks and visual transformer models, and improves the performance and accuracy of the model while reducing model parameters and calculations.

[0071] The main steps are as follows:

[0072] 1. Signal acquisition:

[0073] First, prepare different rock samples. Select rock samples with uniform texture and no obvious cracks, holes or other defects. The size and processing accuracy of the specimens refer to the rock strength test standard recommended by the "Standard for Engineering Rock Test Methods". The specimens are all cylinders with a diameter of 48 to 54 mm. The ratio of the specimen height to the diameter is 2.0 to 2.5. The flatness of the bottom of the specimen should be within 0.02 mm, and the maximum deviation distance between the axis of the specimen and the vertical direction should be less than 1 / 1000 of the diameter. The rock samples are then screened to remove specimens with obvious damage or cracks on the surface and those with substandard size, bottom flatness and diameter deviation. After the specimen is prepared, it needs to be placed in a ventilated and dry environment for at least two weeks.

[0074] Next, a compression experiment is performed. The schematic diagram of the entire experiment is shown below. Figure 2As shown. Install the uniaxial compression test fixture on the test system base, apply a thin layer of vaseline on the upper and lower bottom surfaces of the specimen and place it in the center of the base. Place a rigid pad between the upper end of the specimen and the pressure plate of the test machine, install the axial displacement extensometer and radial displacement extensometer, and adjust the position of the axial and radial displacement extensometers so that the displacement extensometer has an initial reading. Adjust the pressure head so that the rigid pad is in uniform contact with the pressure plate of the test machine, so that the specimen is evenly stressed. Load at a rate of 0.1 to 0.5 MPa per second until the specimen is destroyed.

[0075] At the same time, collect acoustic emission signals. Fix the acoustic emission probe on the rock surface and connect the acoustic emission instrument to the computer. During the compression experiment, keep the acoustic emission monitoring system running continuously. For the acoustic emission monitoring system, set its monitoring sampling interval to 10 microseconds and the frequency to between 102 and 104 Hz. In order to reduce noise interference, select a high acoustic signal trigger level of 3.6 or 4.5 millivolts. The loading system collects data at an interval of 4 times per second to automatically record stress, strain and time, and draw a stress-strain curve. At equal time intervals, read the number of acoustic emission events recorded by the acoustic emission instrument, and sample and store the waveforms of the acoustic emission events. The collected waveforms are as follows: Figure 3 shown.

[0076] Finally, the sound signal data was sorted. For the stress-strain curve and the number of acoustic emission events recorded during the experiment, the acoustic emission parameter waveforms were obtained from different stages, and the signal values ​​were collected by the computer, including 1024 data.

[0077] 2. Signal processing and data set construction:

[0078] In step 1, the signal data of different stages are obtained, which contains 1024 values. It needs to be further converted into a two-dimensional image. The method adopted is continuous wavelet transform, which decomposes the signal through continuous wavelet (CWT) and generates a two-dimensional image.

[0079] First, a mother wavelet (i.e., basic wavelet) of finite length or fast decay is constructed, and then multiple sub-wavelets are generated by scaling and translating, and then multiple sub-wavelets are superimposed to match the input signal. The scaling scale and translation parameters correspond to the frequency and time parameters, and finally the time-frequency diagram of the signal is obtained.

[0080]

[0081] Where a represents the scale parameter, which controls the scaling (compression or stretching) of the wavelet function, b represents the displacement (translation) parameter, which controls the movement of the wavelet function in time, f(t) represents the original signal, t represents the time variable, and W f (a, b) represents the continuous wavelet transform coefficient of f(t) under a and b, ψ abrepresents the basic wavelet function, represents f(t) and ψ ab The inner product of Represents ψ ab The wavelet function after scaling and shifting, The scale and position of is determined by a and b, ψ ab The shape of is determined by a and b. In particular, when a=1 and b=0, That is the basic wavelet function;

[0082] Determine the parameters: signal length f len , sampling frequency sample rate , mother wavelet function, center frequency sliding step cenfre step ;

[0083] Calculate the maximum center frequency Set the current center frequency cenfre now , initialize the time-frequency matrix S;

[0084] According to the center frequency and wavelet function, a wavelet curve is constructed, and then convolved with the original signal to obtain the time distribution vector of the current frequency and update the time-frequency matrix;

[0085] Determine whether the current center frequency is greater than the maximum center frequency. If so, output the time-frequency matrix S; otherwise, update the current frequency cenfre now =cenfre now +cenfre step , then jump back to step 2.

[0086] All the transformed raw data are used to make a data set. Since there are four stages of stress development in the surrounding rock, the final recognition results are four categories. The signals of different stages are divided into these four different stages, and four folders are created. Each folder stores the pictures of the stage, and the name of the folder is the label of the picture. In addition, the data set must be divided into a training set and a test set, that is, 80% of the data is randomly selected from the data set for training, and the remaining 20% ​​of the data is used for testing, so that the construction of the data set is completed.

[0087] 3. Data enhancement and data preprocessing:

[0088] Based on the unique learning method of deep neural networks, the model needs to continuously extract features from image data, then abstract the features, and finally identify and classify based on the abstracted features. At the same time, the model is sensitive to the content of the input data, and data of different scales usually have different results. Therefore, after completing the construction of the data set, some preprocessing operations are required on the data.

[0089] First, for the original 1024x1024 size picture, it is necessary to perform random cropping to crop the image to 224x224 size. Reducing the size of the image can make the model focus on specific locations more quickly, speeding up the learning process. At the same time, by cropping in a random manner, the generalization ability of the model can be improved, so that the recognition ability of the final trained model will not differ too much due to different cropping positions.

[0090] Secondly, the data needs to be further enhanced through random flipping functions. That is, for an image, a certain probability value is set, and the image is flipped according to the probability, which can be horizontal or vertical. The purpose of doing this is also to improve the generalization ability of the model, and to analyze and process waveforms of different shapes. Similar operations include adding noise, rotating, etc., all of which aim to improve the generalization ability of the model and make the model more robust to data collected in different environments.

[0091] Finally, for neural networks, the final form of training data must be in the form of tensors, so a function needs to be called to convert the image into a tensor. In addition, the most important step is to normalize the tensor. The purpose of this is to adjust some effective feature sizes to a similar range. Among them, the normalized mean is set to (0.485, 0.456, 0.406), and the standard deviation is set to (0.229, 0.224, 0.225). If normalization is not performed, the learning effect of the model will change with the fluctuation of the eigenvalue. If the eigenvalue is large, the gradient value will be larger. On the contrary, if the eigenvalue is small, the gradient will be smaller. During back propagation, when the learning rate is small, the parameter update will be slow, and if the learning rate is too large, the model will not converge easily.

[0092] Through the above data processing, the generalization ability of the model can be enhanced and the robustness of the model can be improved.

[0093] 4. Model building:

[0094] The overall structure of the model is as follows Figure 5As shown in the figure, there are roughly four stages, each of which consists of an embedding layer (or merging layer) and a LightNet block. The embedding layer is used for spatial downsampling and expansion of the number of channels, and the LightNet block is used to extract features. Finally, the extracted features are classified. The input of the model is a two-dimensional RGB image, which is converted into a three-dimensional vector of 3x224x224 after preprocessing. After the embedding layer of the first stage, a 40x56x56 vector is obtained. Blocks are stacked once to extract features, and the output feature dimension remains unchanged. After the merging layer of the second stage, a 80*28*28 vector is obtained. Blocks are stacked twice to extract features, and the output feature dimension remains unchanged. After the merging layer of the third stage, a 160x14x14 vector is obtained. Blocks are stacked 8 times to extract features, and the output feature dimension remains unchanged. After the merging layer of the fourth stage, a 320x7x7 vector is obtained. Blocks are stacked 8 times to extract features, and the output feature dimension remains unchanged. Finally, a classifier is used to classify the extracted features.

[0095] The core of this model is the Block used for feature extraction. Its main design ideas are as follows: Figure 6 As shown in the figure, it consists of two parallel branches, which process the input data at the same time and merge them at the end. The first branch is a circular convolution module for extracting context information. Inspired by the visual transformer, learning global features from the entire spatial position is better than extracting features from the local area alone, so a circular convolution module is designed to extract global features. The specific implementation method is: for a feature vector with a dimension of CxHxW, first initialize a position encoding vector with a dimension of CxBx1, align it with the feature vector through a bilinear interpolation function, and obtain a position encoding vector of CxHx1. Then align it with the original vector through horizontal expansion to obtain a position encoding vector of CxHxW. Add and fuse the two to obtain a feature vector with position encoding information. Then vertically splice the feature vector to obtain a vector of Cx2HxW, use a convolution kernel with a dimension of CxHx1 for convolution, and finally obtain a feature vector with a dimension of CxHxW, so that global information exchange in the vertical direction is realized. Similarly, global information exchange in the horizontal direction can be realized. Finally, the information in the two directions is fused to obtain a feature vector with global information. The second branch is used to process local information and extract features to enhance local information. The main implementation method is: for the feature vector of CxHxW, only the first Cp channels are convolved to obtain the feature vector of CpxHxW, and the features of the remaining (C-Cp) channels are convolved point by point to obtain the feature vector of CxHxW.

[0096] 5. Training and testing:

[0097] Setting of training parameters: During the training process, the optimizer uses the Adam optimizer, the learning rate is set to 0.0005, and the momentum coefficient of the Adam optimizer is set to (0.9, 0.99). The weight decay is set to 0.01 to regularize the model parameters. The learning rate adopts a polynomial decay strategy, and the linear warm-up method is used to gradually increase the learning rate in the initial stage of training to help the model adapt to the training data more effectively. The learning rate is warmed up within 50 iterations, and the warm-up learning rate is 1e of the initial learning rate. -6 , the exponent of the polynomial decay is 1.0.

[0098] Setting of loss function: This method uses cross entropy to correct the classification of each category. The loss function is calculated as follows:

[0099]

[0100] where p = [p0…p C-1 ] is a probability distribution, each element p i Indicates the probability that the sample belongs to the i-th category; y = [y0…y C-1 ] is the onehot encoding representation of the sample label. When the sample belongs to the i-th category, y i =1, otherwise y i = 0; C is the sample label. This loss function will judge the category of each image, and optimize the model parameters and improve the model accuracy by minimizing the loss.

[0101] For each round of training, after updating the weights, the model will be tested on the test set to monitor the learning effect of the model. For the test set, the data processing is slightly different from the training set. This includes scaling to 256x256, cropping to 224x224 at the center, converting to tensors, normalizing, etc. The difference is that the model needs to be adjusted to test mode during testing. The difference is that in this mode, the model only makes predictions and does not update parameters.

[0102] The present invention proposes a recognition method based on a lightweight neural network for the task of identifying the surrounding rock stress stage. The core of the method is the construction of a feature extraction block. The feature extraction block is composed of two parallel feature extraction branches. The first feature extraction branch is used to extract local feature information. The traditional CNN convolution kernel usually performs convolution on all channels. Although this can effectively extract features, it greatly increases the parameters of the model. Based on this, some scholars have proposed deep separable convolution, which wants to reduce the number of parameters by reducing the number of channels of convolution, but the effectiveness of feature extraction is greatly reduced. At the same time, according to the results of the visualization results analysis during feature extraction, the first few channels often have higher feature performance, and the features of the latter few channels show redundant characteristics. Therefore, we use local convolution to balance the parameter quantity and extraction accuracy in these two aspects. The specific method is as described above. For the CxHxW feature map, by performing ordinary convolution on the first Cp channels and point-by-point convolution on the latter (C-Cp) channels, a T-shaped convolution kernel is formed to extract features. The experiment also proves the effectiveness of the convolution kernel. The second feature extraction branch is used to extract global information. Traditional global information extractors often use visual transformers and their variants, but they have many parameters and large computational complexity, making them unsuitable for terminal devices. Therefore, a global circular convolution mechanism is introduced. For feature vectors, convolution operations are performed in the horizontal and vertical directions, respectively, so that global context information is extracted. Finally, the features extracted by the two branches are fused.

[0103] The steps of processing, image conversion, and preprocessing of the acoustic emission signal of the present invention are similar to the general processing method. However, the construction of the deep learning model, the recognition efficiency, and the accuracy of the model are different.

[0104] The key to this method is to use a local-global dual-branch feature extraction mode for data processing, where local convolution is used to extract local information and global circular convolution is used to extract global information, and finally the fused information is analyzed. Compared with other deep learning models, this model reduces parameters, reduces the amount of calculation, and improves recognition accuracy.

[0105] Compared with the prior art, the present invention has the following advantages, including but not limited to:

[0106] 1. Reduced application costs:

[0107] Traditional rockburst monitoring systems require the deployment of a large number of sensors in the tunnel, the use of the network for data transmission, and the rental of servers for data processing. This not only consumes a lot of manpower and material resources, but the final recognition effect will also vary greatly due to different types of surrounding rocks. After the introduction of deep learning methods, the use and deployment of sensors have been greatly reduced, reducing costs. At the same time, deep learning models have strong generalization capabilities. Compared with traditional methods, they do not need to understand the specific types of the corresponding surrounding rocks to achieve a higher level of recognition.

[0108] 2. Improved the recognition accuracy of the model:

[0109] Currently, the most widely used CNN model is MobileNetV2. Based on this model, we proposed a new CNN model. By introducing a global circular convolution module that can extract global information, the model's ability to extract global features is improved, allowing the model to achieve higher recognition accuracy in the final recognition stage.

[0110] For the public dataset AE_waves, the final recognition efficiency of the MobileNetV2 model is 0.97, while the recognition efficiency of the model we proposed reaches 0.98, an increase of 1 percentage point.

[0111] 3. Improved the recognition efficiency of the model:

[0112] For CNN models, transformer-based models usually have better performance, but they have a very serious disadvantage, which is that they will greatly increase the number of model parameters and the amount of calculation. On this basis, inspired by transformers, we introduced a global recurrent convolution module, which allows the model to extract global information in the process of feature extraction without using a large number of parameters. In addition, local convolution is introduced into the model, that is, in the process of convolution, convolution is not performed on all channels, but on specific channels with obvious features. This also reduces the number of model parameters and improves recognition efficiency.

[0113] The MobileNetV2 model has 3.4M parameters, while the model we proposed has only 2.6M parameters. This allows it to be applied on some edge devices and get results in a very short time, improving real-time performance.

[0114] The technical points that play an important role in the present invention are as follows:

[0115] 1. In view of the high price and low precision of the traditional sensor-based recognition method, the deep learning method is introduced to reduce labor costs. At the same time, it is not necessary to know the prior knowledge of rock samples in advance, which improves the generalization ability of recognition.

[0116] 2. Aiming at the problem of low accuracy of traditional CNN network, inspired by the self-attention mechanism, a global circular convolution structure is introduced, which includes two forms, one is vertical circular convolution and the other is horizontal circular convolution. The vertical and horizontal receptive fields cover all pixels in the same column and row respectively. Using them together can extract global features from all input pixels. Therefore, the recognition accuracy of the model is improved.

[0117] 3. In order to solve the problem that traditional models have large number of parameters and cannot be deployed on edge devices, a local convolution mechanism is introduced. Experimental verification shows that this is an effective way to reduce computational redundancy. It is only necessary to apply conventional convolution to a part of the input channel for spatial feature extraction, and keep the remaining channels unchanged. For continuous or regular memory access, the first or last continuous Cp channel is regarded as the representative of the entire feature map for calculation. Without loss of generality, the input and output feature maps are assumed to have the same number of channels. This allows the model to use as few parameters as possible when extracting features, reducing computational overhead, thereby increasing the inference speed of the model. At the same time, the global loop convolution described in the second point can achieve the same effect as the traditional transformer mechanism, but with a greatly reduced number of parameters, improved computational efficiency, and reduced inference time.

[0118] Example 2

[0119] like Figures 1 to 6 As shown, as a further optimization of Example 1, based on Example 1, this embodiment also includes the following technical features:

[0120] Step 1: Data collection, by conducting compression experiments on different rock samples under different working conditions, collecting the acoustic emission signals of the rock during the compression process, and dividing the different signals into four stages. Put the signals of the same stage in the same folder, and finally get four folders, and the name of the folder is the label of each category.

[0121] Step 2: Data conversion and division, using cwt transformation, the collected acoustic emission signals are converted into two-dimensional image signals, the image size is 1024x1024. At the same time, two folders are created, named train and val, and the converted data are placed in the train and val folders in a ratio of 4:1.

[0122] Step 3: Input the prepared data set into the model for training. The specific method is as follows:

[0123] Step 1: Data enhancement, including the following processes: Resize, RandomCrop, RandomFilp, PhotoMetricDistortion, Normalize, Pad. A three-channel RGB image with a size of 224*224 is obtained. It is worth noting that these enhancement methods are not the only ones, and other enhancement methods can be used, but the final effect may not be good.

[0124] Step 2: Perform a 3x3 convolution on the preprocessed image with a step size of 4 and an output channel of 40 to obtain a 40x56x56 feature map.

[0125] Step 3: Input the result of the previous step to the LightNet block. In this block, first convolution is performed using a local convolution kernel of size 3x3, with a stride of 1 and 40 output channels. Then convolution is performed using a regular 1x1 convolution kernel with a stride of 1 and 80 output channels. Then, a regular 1x1 convolution kernel is used for convolution, with a step size of 1 and 40 output channels. At the same time, the 40x1x1 Embedding layer is adjusted to a size of 40x56x1 through a bilinear interpolation function, and then horizontally copied 56 times and spliced ​​to obtain a vector of size 40x56x56, which is added to the feature map obtained in the previous step, and the added 40x56x56 vector is vertically spliced ​​to obtain a vector of size 40x112x56. A convolution kernel of size 40x56x1 is used to perform a convolution operation on it to obtain a feature vector of size 40x56x56. This feature map is the feature vector after extracting vertical information; similarly, a feature vector of size 40x56x56 can be obtained after extracting horizontal information. Finally, these three feature vectors of size 40x56x56 are added to obtain a feature vector that combines local information and global information. In this step, the LightNet block is reused once.

[0126] Step 4: Input the result of step 3 and perform convolution using a regular 2x2 convolution kernel with a step size of 2 and an output channel of 80 to obtain a feature vector of size 80x28x28.

[0127] Step 5: Input the result of step 4 and use the same LightNet block as step 3 for feature extraction. The difference is that the number of output channels of each convolution is doubled. The feature extraction block is stacked twice to finally obtain a feature vector of size 80x28x28.

[0128] Step 6: Input the result of step 5 and perform convolution using a regular 2x2 convolution kernel with a step size of 2 and an output channel of 160 to obtain a feature vector of size 160x14x14.

[0129] Step 7: Take the result of step 6 as input and use the same LightNet block as step 5 for feature extraction, except that the number of output channels of each convolution is doubled. Stack the feature extraction block eight times to get a feature vector of size 160x14x14.

[0130] Step 8: Use the result of step 7 as input and perform convolution using a regular 2x2 convolution kernel with a step size of 2 and an output channel of 320 to obtain a feature vector of size 320x7x7.

[0131] Step 9: Take the result of step 8 as input and use the same LightNet block as step 7 for feature extraction. The difference is that the number of output channels of each convolution is doubled. The feature extraction block is stacked twice to finally get a feature vector of size 320x7x7.

[0132] Step 10: Use the result of step 9 as input and perform convolution using a 1x1 convolution kernel with a step size of 1 and an output channel of 1280 to obtain a vector of size 1280x7x7.

[0133] Step 11: Use average pooling to averagely sample the vector obtained in step 10 to obtain a vector of size 1280x1.

[0134] Step 12: Use the fully connected layer to process the vector in step 11 to obtain a vector of size 4x1. Each value represents the probability of the image corresponding to a different category, and the category with the highest probability is the category of the image.

[0135] Step 4: After the training in step 2 is completed, the images in the test set are tested using the trained recognition model to obtain the test results.

[0136] As described above, the present invention can be preferably implemented.

[0137] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.

[0138] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. According to the technical essence of the present invention, within the spirit and principles of the present invention, any simple modification, equivalent replacement and improvement made to the above embodiment still falls within the protection scope of the technical solution of the present invention.

Claims

1. A method for identifying surrounding rock stress stages, characterized in that: The following steps are involved: S1, signal acquisition: collecting acoustic emission signal data at different stress stages; S2, signal processing and data set construction: converting acoustic emission signal data into image data; S3, data enhancement and data preprocessing: perform data enhancement and data preprocessing on image data; S4, model building: building a neural network model; S5, training and testing: training and testing neural network models; In step S4, the neural network model includes an embedding layer, a first LightNet block, a first merging layer, a second LightNet block, a second merging layer, a third LightNet block, a third merging layer, a fourth LightNet block, a convolution layer, a global pooling layer, and an FC function layer connected in sequence; the first LightNet block, the second LightNet block, the third LightNet block, and the fourth LightNet block all include a first branch and a second branch connected in parallel, the first branch is used for extracting local features, the first branch includes a Pconv layer, a first convolution layer, and a second convolution layer connected in sequence, and the second branch includes a circular convolution module for extracting global features; The circular convolution module performs the following processing: For the feature vector, convolution operations are performed in the horizontal and vertical directions respectively. The vertical and horizontal receptive fields cover all pixels in the same column and row respectively. Finally, the information in the two directions is fused to obtain a feature vector with global information.

2. A surrounding rock stress stage identification method according to claim 1, characterized in that: Step S2 includes the following steps: S21, construct a mother wavelet, then generate multiple sub-wavelets by scaling and translating, and then superimpose different sub-wavelets to match the input signal. Decompose the acoustic emission signal through continuous wavelet transform to obtain the time-frequency diagram of the acoustic emission signal. The continuous wavelet transformation formula is as follows: Where a represents the scale parameter, b represents the displacement parameter, f(t) represents the original signal, t represents the time variable, and W f (a, b) represents the continuous wavelet transform coefficient of f(t) under a and b, ψ ab represents the basic wavelet function, <f-ψ ab > represents f(t) and ψ ab The inner product of Represents ψ ab The scaled and shifted wavelet function; S22, determining parameters, the determined parameters include: signal length f len , sampling frequency sample rate , mother wavelet function, center frequency sliding step cenfre step ; S23, calculate the maximum center frequency Set the current center frequency cenfre now , initialize the time-frequency matrix S; S24, constructing a wavelet curve according to the center frequency and the wavelet function, and then convolving it with the original signal to obtain the time distribution vector of the current frequency, and updating the time-frequency matrix; S25, determine whether the current center frequency is greater than the maximum center frequency, if so, output the time-frequency matrix S; otherwise, update the current frequency cenfre now =cenfre now +cenfre step , then jump back to step S21; S26, creating a data set using all signals that have completed continuous wavelet transform.

3. A surrounding rock stress stage identification method according to claim 2, characterized in that: Step S3 includes the following steps: S31, randomly cropping the image; S32, enhancing the image by means of a random flip function; S33, converting the image into a tensor; S34, normalize the tensor.

4. A surrounding rock stress stage identification method according to claim 1, characterized in that: In step S5, during the training process, the optimizer uses the Adam optimizer, the learning rate adopts a polynomial decay strategy, and the learning rate is increased by using a linear warm-up method.

5. A surrounding rock stress stage identification method according to claim 1, characterized in that: In step S5, the loss function is: Where p = [p0…p C-1 ] is a probability distribution, each element p i Indicates the probability that the sample belongs to the i-th category; y = [y0…y C-1 ] is the unique hot encoding representation of the sample label. When the sample belongs to the i-th category, y i =1, otherwise y i =0; C is the sample label.

6. A surrounding rock stress stage identification method according to any one of claims 1 to 5, characterized in that: Step S1 includes the following steps: S11, preparation of different rock samples; S12, compression test; S13, collecting acoustic emission signal data; S14, collating the acoustic emission signal data.

Images