Intelligent identification method for rock fracture acoustic emission signals under high and low temperature circulation effect

Abstract

The invention relates to the technical field of rock mass fracture identification, in particular to a rock fracture acoustic emission signal intelligent identification method under the action of high and low temperature cycling, which comprises the following steps: making a rock test piece, and obtaining acoustic emission data in the rock fracture process; dividing the acoustic emission data into a temperature damage stage and a stress fracturing stage on the basis of the corresponding relationship between acoustic emission count, amplitude frequency and energy in the acoustic emission data and stress; performing wavelet packet transformation and Hilbert-Huang transformation on waveforms in the acoustic emission data set after stage division, and extracting to obtain waveform modal features; preprocessing waveform modal features; a multi-modal recognition model based on an Inceptionv3 branch network structure is constructed, and the multi-modal recognition model is trained and tested; and inputting acoustic emission data collected in real time under different high and low temperature difference cycles into the tested multi-modal recognition model, and further evaluating the classification performance of the model. According to the method, accurate classification of the rock fracture acoustic emission signals in the temperature damage stage and the stress fracturing stage is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rock mass fracture identification, and particularly relates to a rock fracture acoustic emission signal intelligent identification method under high-low temperature cyclic action. BACKGROUND

[0002] In high-altitude or extreme climate areas, the diurnal temperature difference and seasonal temperature change are very significant. Such periodic temperature fluctuations can lead to internal damage accumulation of the surrounding rock of a tunnel, a mine and the like, and weaken the overall mechanical properties of the rock mass. At the same time, when the high stress concentration area formed in the excavation process acts on the rock mass that has been damaged by temperature, the bearing capacity of the rock mass will be significantly reduced, the cracks will be accelerated to penetrate and evolve into macroscopic fractures, and the surrounding rock is prone to deformation instability. Therefore, it is of great significance to reveal the differences between the temperature damage and stress cracking characteristics of the rock mass and develop an intelligent identification method thereof, to ensure the safety and stability of the rock engineering.

[0003] Acoustic emission technology can dynamically characterize the phased characteristics of crack initiation, propagation and penetration by capturing the elastic wave signals released by rock mass fracture in real time, and the signal time-frequency domain attributes and the fracture mode have a strong correlation, which provides key data support for quantitatively analyzing the temperature damage weakening effect and stress driving action in the rock mass fracture process. In recent years, with the development of artificial intelligence technology, especially the application of deep learning methods, intelligent identification technology based on AE signals has become a research hotspot in the field of rock mechanics. At present, the commonly used AE intelligent identification methods can be divided into parameter-based and waveform-based methods. The parameter-based identification method mainly relies on the feature parameters extracted from the AE signals, and realizes the identification and classification of the signals through the analysis of these parameters. However, the identification method is relatively simple, and its effect is relatively dependent on the selection and processing of the feature parameters, and the high-order information in the signal may be ignored.

[0004] The AE waveform contains a large amount of rock mass fracture information. Through Fourier transform, wavelet transform and Hilbert-Huang transform and the like, the original waveform data is processed to analyze the frequency spectrum characteristics thereof, and in combination with relevant intelligent identification methods, the complex time-frequency domain characteristics in the signal can be fully mined, and the accuracy and robustness are significantly improved. Therefore, compared with the AE parameters, it is more meaningful to use the AE waveform to realize the identification and classification of the signals. The current intelligent identification model based on acoustic emission waveform has been successfully applied to crack mode classification, fracture positioning and instability warning system construction in normal temperature environment.

[0005] However, it is worth noting that the existing research system has obvious limitations, and the intelligent recognition method of rock fracture acoustic emission waveform under extreme environmental action is still a blank field. In addition, the existing rock recognition research is mainly based on single modal characteristics (such as time domain or frequency domain analysis) of acoustic emission waveform, but such methods often have difficulty in fully mining the complementary information between different modalities, resulting in limited feature expression capability, thereby reducing the discrimination accuracy and decision reliability of the model.

[0006] Therefore, how to construct an identification model that fuses multi-modal features of acoustic emission waveforms based on a convolutional neural network to achieve accurate classification of rock temperature damage and stress-induced acoustic emission signals under high-low temperature cycling is a key technical problem that needs to be solved. To solve the above technical problems, the present application provides an intelligent recognition method for rock fracture acoustic emission signals under high-low temperature cycling. SUMMARY

[0007] In view of the technical problems that the existing acoustic emission intelligent recognition method cannot effectively fuse multi-modal features and lacks the classification ability of rock fracture signals under extreme temperature environment, an intelligent recognition method for rock fracture acoustic emission signals under high-low temperature cycling is provided. The present application constructs a multi-modal recognition model based on an InceptionV3 branch network to extract and fuse features of denoising waveforms (time domain), time-frequency phase planes (frequency domain) and instantaneous energy (energy domain), and embeds a hybrid attention mechanism to increase the attention to key features of the waveforms, thereby achieving the technical effect of classifying temperature damage and stress-induced signals.

[0008] The technical means adopted by the present application are as follows:

[0009] An intelligent recognition method for rock fracture acoustic emission signals under high-low temperature cycling, comprising the following steps:

[0010] S1. Making rock test pieces, and using a high-temperature box and a low-temperature box to perform high-low temperature cycling on the rock test pieces;

[0011] S2. Using an electro-hydraulic servo testing machine to perform indoor uniaxial testing on the rock test pieces after high-low temperature cycling, and using an acoustic emission acquisition instrument and an acoustic emission sensor to obtain acoustic emission data during rock fracture;

[0012] S3. Based on the corresponding relationship between acoustic emission count, amplitude-frequency and energy and stress in the acoustic emission data, dividing the acoustic emission data into temperature damage stage and stress-induced cracking stage, and generating acoustic emission data set after stage division;

[0013] S4. Waveform in the acoustic emission data set after stage division is subjected to wavelet packet transform and Hilbert Huang transform, and waveform modal features are extracted, including denoising waveform image, time-frequency phase plane image and instantaneous energy image;

[0014] S5. Preprocessing the waveform modal features, and dividing the preprocessed waveform modal features into a training set and a test set;

[0015] S6. Constructing a multi-modal recognition model based on an inceptionv3 branch network structure, training the multi-modal recognition model using the training set, and obtaining a trained multi-modal recognition model;

[0016] S7. Testing the trained multi-modal recognition model using the test set;

[0017] S8. Inputting real-time collected acoustic emission data under different high-low temperature difference cycles into the tested multi-modal recognition model, and further evaluating the classification performance of the model.

[0018] Further, the operation of the high-low temperature cycle treatment is:

[0019] The rock sample is subjected to alternating heating and cooling cycle treatment using a high-temperature oven and a low-temperature oven, each cycle including heating the rock to a temperature T1 and maintaining it for a time t, then cooling it to a temperature T2 and maintaining it for a time t, for a total of N cycles.

[0020] Further, the acoustic emission sensor arrangement is: two acoustic emission probes are symmetrically arranged on the upper and lower ends of the rock sample, and vaseline is applied to the contact surface between the probes and the rock sample to enhance signal coupling.

[0021] Further, the network architecture of the multi-modal recognition model is: a first network branch, a concatenation layer, a fully connected layer, and a Softmax function connected in sequence; the first network branch receives a denoised waveform image on one hand and outputs waveform time domain features on the other hand; a second network branch, a concatenation layer, a fully connected layer, and a Softmax function connected in sequence; the second network branch receives a time-frequency phase plane image on one hand and outputs waveform frequency domain features on the other hand; a third network branch, a concatenation layer, a fully connected layer, and a Softmax function connected in sequence; the third network branch receives an instantaneous energy image on one hand and outputs instantaneous energy features on the other hand.

[0022] Further, the workflow of the multi-modal recognition model is:

[0023] First, input the denoised waveform image into the first network branch to output the waveform time domain features;

[0024] Then, input the time-frequency phase plane image into the second network branch to output the waveform frequency domain features;

[0025] Second, input the instantaneous energy image into the third network branch to output the instantaneous energy features;

[0026] The waveform time domain feature, the waveform frequency domain feature and the instantaneous energy feature are input into a concatenation layer to form a multi-modal fusion feature vector;

[0027] The multi-modal fusion feature vector is input into a full connection layer, and a classification result of the rock mass fracture acoustic emission signal is generated through a Softmax function.

[0028] Further, the first network branch, the second network branch and the third network branch have the same architecture, and the specific architecture of the first network branch is: sequentially connected first, second, third convolutional layers, a maximum pooling layer, fourth, fifth, sixth convolutional layers, an Inception module A, a first hybrid attention module, an Inception module B, a second hybrid attention module, an Inception module C, a third hybrid attention module and an average pooling layer.

[0029] Further, the first hybrid attention module, the second hybrid attention module and the third hybrid attention module have the same architecture, the first hybrid attention module comprises a channel attention module and a spatial attention module connected in series, the channel attention module enhances key channel features through a global average pooling and a full connection layer, and the spatial attention module extracts important region features through a spatial convolution.

[0030] Further, the preprocessed waveform modal feature comprises:

[0031] The waveform modal feature is processed into a standard size of the multi-modal recognition model by using a data standardization method.

[0032] The waveform modal feature of the standard size is subjected to data enhancement through random rotation, translation, scaling and flipping to obtain a preprocessed waveform modal feature.

[0033] Further, in the process of training the multi-modal recognition model by using the training set, an Adam optimization algorithm is used to update the model parameters, a cross-entropy loss function is used to minimize the prediction error to make the model converge, and a learning rate decay strategy is used to improve the convergence speed and stability of the model.

[0034] Further, in the process of testing the trained multi-modal recognition model by using the test set, the test performance of the model is evaluated by using accuracy, precision, recall and F1 score,

[0035] The accuracy is the ratio of the number of correctly predicted samples to the total number of samples, and the calculation formula is as follows:

[0036]

[0037] Wherein, Accuracy is the accuracy rate, TP is the number of samples correctly predicted as positive by the model, TN is the number of samples correctly predicted as negative by the model, FP is the number of samples incorrectly predicted as positive by the model, and FN is the number of samples incorrectly predicted as negative by the model;

[0038] The precision rate is the proportion of true positive samples in the positive samples predicted by the model, and the calculation formula is as follows:

[0039]

[0040] Wherein, precision is the precision rate;

[0041] The recall rate is the proportion of positive samples correctly predicted in the actual positive samples, and the calculation formula is as follows:

[0042]

[0043] Wherein, Recall is the recall rate;

[0044] The F1 score is the harmonic mean of the precision rate and the recall rate, and the calculation formula is as follows:

[0045]

[0046] Wherein, F1-score is the F1 score.

[0047] Compared with the prior art, the present application has the following advantages:

[0048] 1. By analyzing the acoustic emission count, energy evolution law and amplitude-frequency response characteristics of rock failure process under high-low temperature cycle, the temperature damage stage and stress cracking stage are accurately decoupled. Based on the joint analysis method of wavelet packet multi-scale denoising reconstruction and Hilbert Huang transform, the acoustic emission waveform difference mechanism of temperature damage and stress cracking is revealed, which breaks through the cognitive limitation of traditional single-dimensional feature analysis and lays a foundation for building a waveform multi-modal recognition model.

[0049] 2. For the time domain details of the denoised waveform, the frequency domain distribution of the time-frequency phase plane and the nonlinear characteristics of the instantaneous energy, a feature extraction based on the InceptionV3 branch network structure is constructed, and the cross-modal feature efficient fusion is realized through global average pooling and feature vector splicing. At the same time, the channel and spatial double attention mechanism is introduced to enhance the attention of the key area of the waveform and reduce the interference of irrelevant information, effectively improving the model classification accuracy, so as to realize the classification of temperature damage and stress cracking acoustic emission signals. The method provides a theoretical basis for the damage evolution mechanism and stability intelligent monitoring of underground engineering surrounding rock in extreme environment.

[0050] Based on the above reasons, the present application can be widely popularized in the field of rock mass failure identification. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0052] Figure 1 The flow chart for the implementation of the present application.

[0053] Figure 2 The device for treating rock samples by high-low temperature cycle.

[0054] Figure 3 The schematic diagram of the rock sample treated by high-low temperature cycle.

[0055] Figure 4 The rock uniaxial compression acoustic emission test system.

[0056] Figure 5 The original waveform diagram of the rock temperature damage stage.

[0057] Figure 6 The original waveform diagram of the rock stress-induced cracking stage.

[0058] Figure 7 The structure principle diagram of the multi-modal recognition model based on InceptionV3 branch network.

[0059] In the figure: 1, high temperature box; 2, low temperature box; 3, rock test piece; 4, electro-hydraulic servo testing machine; 5, control console; 6, first electronic computer; 7, acoustic emission sensor; 8, acoustic emission amplifier; 9, acoustic emission acquisition instrument; 10, second electronic computer. DETAILED DESCRIPTION

[0060] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0061] It should be noted that the terms "first", "second", and the like in the description and claims of the application and the above drawings are used to distinguish between similar objects, and are not necessarily used to describe a particular sequential or chronological order. It should be understood that the use of such terms as "first", "second", and the like, if used herein, are used to distinguish one from another, and are not necessarily used to describe a particular sequential or chronological order. It should be understood that the use of such terms as "first", "second", and the like, if used herein, are used to distinguish one from another, and are not necessarily used to describe a particular sequential or chronological order. In addition, the terms "comprising" and "having" and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product or apparatus that includes a list of steps or units as processes, methods, systems, products or apparatuses are not necessarily limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or apparatuses.

[0062] See Figure 1 , Figure 1 is a flowchart of the implementation of the application. The application provides a rock fracture acoustic emission signal intelligent identification method under high and low temperature cycles, the steps of which are:

[0063] S1. Make rock test piece 3, and use high temperature box 1 and low temperature box 2 to perform high and low temperature cycle treatment on rock test piece 3.

[0064] Specifically, in this embodiment, the rock test piece 3 is a 50mm*100mm standard cylindrical basalt sample. After the rock test piece is placed into the high temperature box 1, the temperature is raised to 100℃ at a rate of 5℃ / min, and after 24h in this environment, the rock sample 3 is quickly taken out and placed into the low temperature box 2, and is also kept at -80℃ for 24h, to form one heating and cooling cycle, and a total of 15 cycles are performed; see Figure 2 and Figure 3 , Figure 2 is a device diagram for high and low temperature cycle treatment of rock samples, Figure 3 is a high and low temperature cycle diagram for rock samples.

[0065] S2. Use electro-hydraulic servo testing machine 4 to perform indoor uniaxial test on the rock test piece 3 after high and low temperature cycle treatment, and combine acoustic emission acquisition instrument 9 and acoustic emission sensor 7 to obtain acoustic emission data in the rock fracture process.

[0066] Specifically, the uniaxial compression acoustic emission test system is composed of electro-hydraulic servo testing machine 4, control console 5, first electronic computer 6, acoustic emission sensor 7, acoustic emission amplifier 8, acoustic emission acquisition instrument 9 and second electronic computer 10; see Figure 4 , Figure 4 is a rock uniaxial compression acoustic emission test system.

[0067] The AE collection instrument 9 in this embodiment is PCI-2, and the sampling frequency and the amplification threshold are set to 2 MHz and 40 dB respectively. Two nano30 acoustic emission sensors 7 are fixed on the upper and lower ends of one side of the rock sample 3 by using adhesive tape, and vaseline is applied on the rock sample 3 and the sensors 7 to ensure close contact.

[0068] S3. The corresponding relationship between the emission count, amplitude-frequency and energy in the acoustic emission data and the stress is explored, the acoustic emission data is divided into a temperature damage stage and a stress-induced cracking stage based on the corresponding relationship between the acoustic emission count, amplitude-frequency and energy in the acoustic emission data and the stress, and an acoustic emission data set after the division of the stages is generated.

[0069] S4. The waveforms in the acoustic emission data set after the division of the stages are subjected to wavelet packet transform and Hilbert Huang transform, and waveform modal features are extracted, including denoised waveform images, time-frequency phase plane images and instantaneous energy images. Please refer to Figure 5 and Figure 6 , Figure 5 and Figure 6 are the original waveform images of the rock temperature damage stage and the stress-induced cracking stage respectively.

[0070] Specifically, the AE time sequence waveform accurately represents the intensity distribution and morphological evolution of the signal through time domain features, the time-frequency phase plane analyzes the distribution characteristics of the signal components from the frequency domain dimension to identify the frequency response of different fracture modes, and the instantaneous energy quantifies the energy release intensity in the local time window to reveal the transient energy change law in the fracture evolution process. Among them, the AE signal inevitably contains a large amount of background noise and external interference signals, which masks the true characteristics of rock fracture behavior, resulting in loss or extraction error of key information. Therefore, compared with the original waveform containing noise, the denoised waveform is more effective as the acoustic emission waveform time domain feature extraction. Effective combination of the above three waveform modal features can improve the accuracy of acoustic emission signal classification. The typical acoustic emission original waveforms of the two stages are selected, the denoised waveform, the time-frequency phase plane and the instantaneous energy three waveform modal features are obtained according to the wavelet packet transform and the Hilbert Huang transform, and the differences in the acoustic emission waveform features of the temperature damage and the stress-induced cracking are accurately revealed.

[0071] In this embodiment, the original waveform is processed by wavelet packet to obtain the denoised waveform and the time-frequency phase plane, wherein the data sampling frequency is 2 MHz, the Nyquist sampling frequency is 1 MHz, the Db4 wavelet basis function is selected for 4-layer decomposition, and there are 16 (2 4 ) frequency bands, and the frequency range of each frequency band is 62.5 kHz. The original waveform is processed by Hilbert Huang transform to obtain the instantaneous energy, and the instantaneous features in the signal are captured.

[0072] S5. The waveform modal features are preprocessed, and the preprocessed waveform modal features are divided into a training set and a test set.

[0073] Specifically, in the embodiment, the noise reduction waveform, the time-frequency reduction plane and the instantaneous energy are input images of a multi-modal model, 2000 waveform signals in the temperature damage and stress cracking stages are selected for training and evaluating the performance of the model. An independent input layer is created for each modal data, and the data standardization method is used to process the images to a standard size of 229x229x3. In order to improve the generalization ability of the model, random rotation (20°), translation (20% movement), scaling (20% enlargement), flipping (horizontal and vertical flipping) and the like are used to enhance the images. The data set is divided into 80% training set and 20% test set, the training set is used for training and parameter adjustment of the model, and the test set is used for verifying the classification effect of the model.

[0074] S6. A multi-modal recognition model based on the inceptionv3 branch network structure is constructed, and the multi-modal recognition model is trained by using the training set to obtain the trained multi-modal recognition model.

[0075] Specifically, the multi-modal recognition model extracts waveform time domain, frequency domain and instantaneous energy features. Since the signals are unevenly distributed in channels and space, a mixed attention mechanism is embedded to enhance the attention of the key area of the waveform. The network architecture of the multi-modal recognition model is: a first network branch, a splicing layer, a full connection layer and a Softmax function connected in sequence; the first network branch receives the noise reduction waveform image on one hand and outputs the waveform time domain feature on the other hand; a second network branch, a splicing layer, a full connection layer and a Softmax function connected in sequence; the second network branch receives the time-frequency phase plane image on one hand and outputs the waveform frequency domain feature on the other hand; a third network branch, a splicing layer, a full connection layer and a Softmax function connected in sequence; the third network branch receives the instantaneous energy image on one hand and outputs the instantaneous energy feature on the other hand.

[0076] The first network branch, the second network branch and the third network branch have the same architecture, and the specific architecture of the first network branch is: a first convolutional layer, a second convolutional layer, a third convolutional layer, a maximum pooling layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, an Inception module A, a first mixed attention module, an Inception module B, a second mixed attention module, an Inception module C, a third mixed attention module and an average pooling layer connected in sequence.

[0077] The first mixed attention module, the second mixed attention module and the third mixed attention module have the same architecture, the first mixed attention module includes a channel attention module and a spatial attention module connected in series, the channel attention module enhances the key channel feature through global average pooling and a full connection layer, and the spatial attention module extracts important area features through spatial convolution.

[0078] The workflow of the multi-modal recognition model is as follows:

[0079] Firstly, the denoising waveform image is input into the first network branch, and the waveform time domain feature is output.

[0080] Subsequently, the time-frequency phase plane image is input into the second network branch, and the waveform frequency domain feature is output.

[0081] Secondly, the instantaneous energy image is input into the third network branch, and the instantaneous energy feature is output.

[0082] Then, the waveform time domain feature, the frequency domain feature and the instantaneous energy feature are passed through a splicing layer to form a multi-modal fusion feature vector.

[0083] Finally, the multi-modal fusion feature vector is input into a fully connected layer, and then a classification result of the rock mass fracture acoustic emission signal is generated through a Softmax function.

[0084] Specifically, the input modal image is first subjected to 3x3 convolution with a step size of 2 and 1, and then subjected to 3x3 convolution with a step size of 2 and 1 after 3x3 pooling with a step size of 2, so as to extract low-level features such as image edges and textures; subsequently, three groups of Inception modules (Inception A, B and C) are passed through, and each group of modules further extracts intermediate and high-level features through multiple layers of convolution and pooling. In order to enhance the feature expression of the key area of the waveform mode, a mixed attention composed of channel attention and spatial attention is embedded behind the Inception A, B and C modules. Finally, each modal network branch outputs a feature size of 8x8x2048. Global average pooling is used to compress the features into a one-dimensional vector, and then the extracted one-dimensional vector is spliced together to form a multi-modal fusion feature vector. The fused feature vector is input into a fully connected layer to further extract high-order abstract features through nonlinear transformation, and the ReLU activation function enhances the nonlinear expression ability of the model. The one-dimensional feature is output through the Softmax function to obtain a binary classification result; please refer to Figure 7 , Figure 7 The structure principle diagram of the multi-modal recognition model based on the InceptionV3 branch network.

[0085] Specifically, the Adam optimization algorithm is used to update the model parameters during the training process. The cross-entropy loss function minimizes the prediction error to make the model converge. The learning rate decay strategy is used to improve the convergence speed and stability of the model. The loss value and accuracy are used to evaluate the model training process.

[0086] In this embodiment, the Epoch and batch size are set to 100 and 16, the initial learning rate is set to 0.01, and when the loss value does not improve for 3 consecutive Epochs, the learning rate is reduced to 1 / 10 of the previous one, and the minimum learning rate is limited to 1x10-6 The model tends to be stable at about 21 epochs, and the loss value of training convergence is 0.09, and the accuracy is 0.93.

[0087] S7. Test the trained multi-modal recognition model with the test set.

[0088] In the process of testing the trained multi-modal recognition model with the test set, the test performance of the model is evaluated by accuracy, precision, recall and F1 score, and finally the classification of temperature damage and stress cracking acoustic emission signals is realized.

[0089] The accuracy is the ratio of the number of samples correctly predicted by the model to the total number of samples, and the calculation formula is as follows:

[0090]

[0091] Where Accuracy is the accuracy, TP is the number of samples correctly predicted as positive by the model, TN is the number of samples correctly predicted as negative by the model, FP is the number of samples incorrectly predicted as positive by the model, and FN is the number of samples incorrectly predicted as negative by the model;

[0092] The precision is the proportion of true positive samples in the positive samples predicted by the model, and the calculation formula is as follows:

[0093]

[0094] Where precision is the precision;

[0095] The recall is the proportion of positive samples correctly predicted in the actual positive samples, and the calculation formula is as follows:

[0096]

[0097] Where Recall is the recall;

[0098] The F1 score is the harmonic mean of precision and recall, and the calculation formula is as follows:

[0099]

[0100] Where F1-score is the F1 score.

[0101] S8. Input the real-time collected acoustic emission data under different high-low temperature difference cycles into the multi-modal recognition model after testing, and further evaluate the classification performance of the model.

[0102] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent identification method for acoustic emission signals of rock fracture under high and low temperature cycles, characterized in that: The following steps are involved: S1. Making rock specimens, using a high temperature box and a low temperature box to cycle the rock specimens through high and low temperature treatment; S2. Conduct indoor uniaxial tests on rock specimens subjected to high- and low-temperature cycling using an electro-hydraulic servo testing machine. Acoustic emission data from the rock fracture process is obtained using an acoustic emission acquisition instrument and acoustic emission sensor. S3. Based on the correspondence between the acoustic emission count, amplitude frequency and energy and stress in the acoustic emission data, the acoustic emission data is divided into the temperature damage stage and the stress cracking stage to generate an acoustic emission data set after the stage division; S4. performing wavelet packet transform and Hilbert-Huang transform on the waveforms in the acoustic emission data set after the division phase to extract waveform modal features, the waveform modal features including a denoised waveform image, a time-frequency phase plane image, and an instantaneous energy image; S5. Preprocessing the waveform modal features, dividing the preprocessed waveform modal features into a training set and a test set; S6. Construct a multimodal recognition model based on the inceptionv3 branch network structure, and train the multimodal recognition model using the training set to obtain a trained multimodal recognition model; S7. Testing the trained multimodal recognition model using a test set; S8. Input the real-time acoustic emission data collected under different high and low temperature differential cycles into the tested multimodal recognition model to further evaluate the classification performance of the model.

2. The intelligent identification method of rock fracture acoustic emission signals under high and low temperature cycles according to claim 1 is characterized in that: The operation of the high and low temperature cycle treatment is as follows: The rock specimens were subjected to alternating heating and cooling cycles using a high-temperature chamber and a low-temperature chamber. Each cycle involved heating the rock to temperature T1 and maintaining it for time t, and then cooling it to temperature T2 and maintaining it for time t, for a total of N cycles.

3. The intelligent identification method of rock fracture acoustic emission signals under high and low temperature cycles according to claim 1 is characterized in that: The acoustic emission sensor is arranged in such a way that two acoustic emission probes are symmetrically arranged at the upper and lower ends of the rock specimen, and vaseline is applied on the contact surface between the probes and the rock specimen to enhance signal coupling.

4. The intelligent identification method of rock fracture acoustic emission signals under high and low temperature cycles according to claim 1 is characterized in that: The network architecture of the multimodal recognition model is: a first network branch, a splicing layer, a fully connected layer and a Softmax function connected in sequence; the first network branch receives a denoised waveform image on the one hand and outputs waveform time domain features on the other hand; a second network branch, a splicing layer, a fully connected layer and a Softmax function connected in sequence; the second network branch receives a time-frequency phase plane image on the one hand and outputs waveform frequency domain features on the other hand; a third network branch, a splicing layer, a fully connected layer and a Softmax function connected in sequence; the third network branch receives an instantaneous energy image on the one hand and outputs instantaneous energy features on the other hand.

5. The intelligent identification method of rock fracture acoustic emission signals under high and low temperature cycles according to claim 4 is characterized in that: The workflow of the multimodal recognition model is as follows: First, the denoised waveform image is input into the first network branch, and the waveform time domain features are output; Subsequently, the time-frequency phase plane image is input into the second network branch, which outputs the waveform frequency domain features; Secondly, the instantaneous energy image is fed into the third network branch to output the instantaneous energy features; The waveform time domain features, waveform frequency domain features and instantaneous energy features are passed through the splicing layer to form a multimodal fusion feature vector; The multimodal fusion feature vector is input into the fully connected layer and then passes through the Softmax function to generate the rock fracture acoustic emission signal classification result.

6. The intelligent identification method of rock fracture acoustic emission signals under high and low temperature cycles according to claim 4 is characterized in that: The first network branch, the second network branch and the third network branch have the same architecture. The specific architecture of the first network branch is: the first convolutional layer, the second convolutional layer, the third convolutional layer, the maximum pooling layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer, the Inception module A, the first hybrid attention module, the Inception module B, the second hybrid attention module, the Inception module C, the third hybrid attention module and the average pooling layer connected in sequence.

7. The intelligent identification method of rock fracture acoustic emission signals under high and low temperature cycles according to claim 6 is characterized in that: The first hybrid attention module, the second hybrid attention module and the third hybrid attention module have the same architecture. The first hybrid attention module includes a channel attention module and a spatial attention module connected in series. The channel attention module enhances key channel features through global average pooling and fully connected layers, and the spatial attention module extracts important area features through spatial convolution.

8. The intelligent identification method of rock fracture acoustic emission signals under high and low temperature cycles according to claim 1 is characterized in that: The pre-processed waveform modal features include: Using a data normalization method, the waveform modal features are processed into a standard size of the multimodal recognition model; The waveform modal features of the standard size are subjected to data enhancement by random rotation, translation, scaling and flipping to obtain preprocessed waveform modal features.

9. The intelligent identification method of rock fracture acoustic emission signals under high and low temperature cycles according to claim 1 is characterized in that: In the process of training the multimodal recognition model using the training set, the Adam optimization algorithm is used to update the model parameters; the cross entropy loss function prediction error is minimized to make the model converge; and the learning rate decay strategy is adopted to improve the convergence speed and stability of the model.

10. The intelligent identification method of rock fracture acoustic emission signals under high and low temperature cycles according to claim 1 is characterized in that: In the process of testing the trained multimodal recognition model using the test set, the test performance of the model is evaluated by accuracy, precision, recall and F1 score. The accuracy is the ratio of the number of samples correctly predicted by the model to the total number of samples. The calculation formula is as follows: Among them, Accuracy is the accuracy rate, TP is the number of samples correctly predicted as positive by the model, TN is the number of samples correctly predicted as negative by the model, FP is the number of samples incorrectly predicted as positive by the model, and FN is the number of samples incorrectly predicted as negative by the model; The precision rate is the ratio of positive samples to positive samples predicted by the model. The calculation formula is as follows: Among them, precision is the accuracy rate; The recall rate is the proportion of correct predictions of positive samples in the actual positive samples, and the calculation formula is as follows: Among them, Recall is the recall rate; The F1 score is the harmonic mean of precision and recall, and is calculated as follows: Among them, F1-score is the F1 score.

Images