Method, system and device for identifying tensile and shear cracks in tight sandstone

Abstract

This application relates to a method, system, and device for identifying tensile and shear cracks in tight sandstone. By collecting the acoustic emission time-frequency map data of the target tight sandstone and inputting it into a trained Transformer model to identify the crack type. The Transformer model consists of an autoencoder neural network with a self-attention mechanism, a fully connected layer, and a softmax layer. In the pre-training, a representation alignment constraint is added to restrict the decoder from learning the encoded information, thereby constraining the encoding task in the self-attention encoder and improving its feature representation ability. At the same time, the self-attention encoder adopts a self-attention module to comprehensively analyze the time-frequency map information through self-correlation operations, enabling the neural network to more fully represent the features of the time-frequency map. The trained model can effectively distinguish between tensile cracks and shear cracks, improving the accuracy and efficiency of crack identification in tight sandstone.

Description

Technical Field

[0001] This application relates to the technical fields of rock mechanics, detection signal processing, artificial intelligence, and machine learning, and particularly relates to a method, system, and device for identifying tensile and shear cracks in tight sandstone. Background Art

[0002] The propagation direction of cracks mainly depends on the stress state on the crack surface. Clearly identifying the crack propagation type during rock failure is of great significance for revealing the rock failure mechanism and clarifying the rock failure type.

[0003] Traditional methods for identifying crack propagation types include microscopic analysis and acoustic emission monitoring analysis. Microscopic analysis is to observe the morphology, direction, and characteristics of cracks using a microscope (such as an optical microscope or an electron microscope) to distinguish tensile cracks and shear cracks; acoustic emission monitoring analysis refers to using acoustic emission sensors to continuously monitor the acoustic emission signals generated by the formation and propagation of internal cracks in rocks during rock tests, and identifying the crack type by analyzing characteristic parameters such as the amplitude and frequency of the acoustic emission signals.

[0004] However, traditional methods for identifying tensile and shear cracks in tight sandstone have problems such as low efficiency and low accuracy. Summary of the Invention

[0005] Based on this, it is necessary to provide a method for identifying tensile and shear cracks in tight sandstone, a system for identifying tensile and shear cracks in tight sandstone, and a computer device to address the above technical problems.

[0006] To achieve the above object, the embodiments of the present invention adopt the following technical solutions:

[0007] On the one hand, a method for identifying tensile and shear cracks in tight sandstone is provided, including:

[0008] Collecting target acoustic emission data of the target tight sandstone; the target acoustic emission data is the acoustic emission time-frequency diagram of the target tight sandstone;

[0009] Inputting the target acoustic emission data into a trained Transformer model; the Transformer model includes an autoencoder neural network based on the self-attention mechanism, a fully connected layer network, and a softmax layer network; during training, a representation alignment constraint is added to the autoencoder neural network;

[0010] Obtaining the target crack type of the target acoustic emission data output by the Transformer model; the target crack type includes tensile cracks and shear cracks.

[0011] On the other hand, a system for identifying tensile and shear cracks in tight sandstone is also provided, including:

[0012] A collection module for collecting target acoustic emission data of target tight sandstone; the target acoustic emission data is the acoustic emission time-frequency diagram of the target tight sandstone.

[0013] An input module for inputting the target acoustic emission data into a trained Transformer model; the Transformer model includes an autoencoder neural network based on the self-attention mechanism, a fully connected layer network, and a softmax layer network; during the training process, a representation alignment constraint is added to the autoencoder neural network.

[0014] An output module for obtaining the target crack type of the target acoustic emission data output by the Transformer model; the target crack types include tensile cracks and shear cracks.

[0015] On the other hand, a computer device is also provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned method for identifying tensile and shear cracks in tight sandstone are implemented.

[0016] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0017] The above-mentioned method, system, and device for identifying tensile and shear cracks in tight sandstone collect the acoustic emission time-frequency diagram data of the target tight sandstone and input it into a trained Transformer model to identify the crack type. The Transformer model includes an autoencoder neural network based on the self-attention mechanism, a fully connected layer network, and a softmax layer network. A representation alignment constraint is added during the pre-training process, that is, when training the autoencoder neural network, the decoder is restricted from learning the encoding information, and thus the encoding task is all constrained in the self-attention encoder, effectively improving the feature representation ability of the self-attention encoder; at the same time, a self-attention module is adopted in the self-attention encoder, and the time-frequency diagram information is comprehensively analyzed through self-correlation operations, enabling the neural network to more fully represent the time-frequency diagram features. Then, a training labeled dataset is obtained through experiments, and then neural network training is carried out to obtain accurate identification of the expansion of tensile and shear cracks. Using the trained model for identification can effectively distinguish between tensile cracks and shear cracks, improve the accuracy and efficiency of crack identification in tight sandstone, achieve accurate classification of crack types, and have strong robustness and real-time identification capabilities. Description of the Drawings

[0018] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0019] Figure 1 It is a schematic flowchart of a method for identifying tensile and shear cracks in tight sandstone in an embodiment;

[0020] Figure 2 It is a schematic diagram of an autoencoder neural network structure based on a self-attention mechanism in an embodiment;

[0021] Figure 3 It is a time-frequency diagram generated based on a pre-trained model in an embodiment, where (a) represents the original picture, (b) represents the input picture, and (c) represents the reconstructed picture;

[0022] Figure 4 It is a schematic diagram of a rock mechanics testing machine for three-point bending test and shear penetration test in an embodiment, where (a) represents the schematic diagram of the testing machine for three-point bending test, and (b) represents the schematic diagram of the testing machine for shear penetration test;

[0023] Figure 5 It is a physical drawing of the processing of a semi-circular disk three-point bending specimen and a shear penetration specimen in an embodiment, where (a) represents the physical drawing of the processing of the semi-circular disk three-point bending specimen, and (b) represents the physical drawing of the processing of the shear penetration specimen;

[0024] Figure 6 It is a sectional view and a three-dimensional view of the processing of a semi-circular disk three-point bending specimen in an embodiment, where (a) represents the sectional view of the processing of the semi-circular disk three-point bending specimen, and (b) represents the three-dimensional view of the processing of the semi-circular disk three-point bending specimen;

[0025] Figure 7 It is a sectional view and a three-dimensional view of the processing drawing of a shear penetration specimen in an embodiment, where (a) represents the sectional view of the processing drawing of the shear penetration specimen, and (b) represents the three-dimensional view of the processing of the shear penetration specimen Figure 3 Three-dimensional view;

[0026] Figure 8 It is a time-frequency diagram after the transformation of tensile and shear cracks in an embodiment, where (a) represents the time-frequency diagram after the transformation of tensile cracks, and (b) represents the time-frequency diagram after the transformation of shear cracks;

[0027] Figure 9 It is the accuracy rate and loss distribution during the model training process in an embodiment;

[0028] Figure 10 It is a recognition basis diagram for non-acidified tensile cracks in an embodiment. Among them, (a) represents the time-frequency diagram of non-acidified tensile cracks, and (b) represents the weight diagram of non-acidified tensile cracks;

[0029] Figure 11 It is a recognition basis diagram for non-acidified shear cracks in an embodiment. Among them, (a) represents the time-frequency diagram of non-acidified shear cracks, and (b) represents the weight diagram of non-acidified shear cracks;

[0030] Figure 12 It is a recognition basis diagram for acidified tension in an embodiment. Among them, (a) represents the time-frequency diagram of acidified tensile cracks, and (b) represents the weight diagram of acidified tensile cracks;

[0031] Figure 13 It is a recognition basis diagram for acidified shear in an embodiment. Among them, (a) represents the time-frequency diagram of acidified shear cracks, and (b) represents the weight diagram of acidified shear cracks. Detailed implementation manners

[0032] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which this application belongs. The terms used in the description of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0034] It should be noted that referring to "embodiment" herein means that a specific feature, structure or characteristic described in connection with the embodiment may be included in at least one embodiment of the present invention. Displaying this phrase at various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments.

[0035] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0036] In one embodiment, as Figure 1 shown, the embodiment of the present application provides a method for identifying tensile and shear cracks in tight sandstone, including the following processing steps S12-16:

[0037] S12, collecting target acoustic emission data of the target tight sandstone; the target acoustic emission data is the acoustic emission time-frequency diagram of the target tight sandstone.

[0038] It can be understood that the target tight sandstone represents the specific tight sand to be studied and analyzed. By monitoring the acoustic emission signals of tight sandstone samples during tensile or shear tests, these signals are converted into acoustic emission time-frequency diagrams, i.e., time-frequency diagrams, for analyzing and identifying the types of cracks. Specifically, these acoustic emission time-frequency diagrams are input as input data into the trained Transformer model.

[0039] S14, input the target acoustic emission data into the trained Transformer model; the Transformer model includes an autoencoder neural network based on the self-attention mechanism, a fully connected layer network, and a softmax layer network; during training, a representation alignment constraint is added to the autoencoder neural network.

[0040] It can be understood that the collected acoustic emission time-frequency diagrams of the target tight sandstone are input as input data into a trained Transformer model. Among them, the Transformer model includes an autoencoder neural network based on the self-attention mechanism, a fully connected layer network, and a softmax layer network. The self-attention mechanism is used to capture important features in the input data. The autoencoder neural network can encode (extract key features) and decode (reconstruct data) the input data, and the self-attention mechanism helps the model focus on relevant important features when processing data; the fully connected layer network is responsible for further processing the features extracted by the autoencoder neural network. The fully connected layer transforms the features into a more abstract representation through weights and biases, thus providing support for the final classification task; the softmax layer network is the output layer of the Transformer model, which is used to convert the output of the fully connected layer into a probability distribution, thereby realizing classification. The softmax layer will judge which crack type (tensile crack or shear crack) the acoustic emission data belongs to according to the probability values of the input features.

[0041] In addition, as Figure 2 shown: Schematic diagram of the autoencoder neural network structure based on the self-attention mechanism. During training, a representation alignment constraint is added to the autoencoder neural network to restrict the decoder from learning the encoded information, and then all the encoding tasks are constrained in the self-attention encoder, effectively improving the feature representation ability of the self-attention encoder. Specifically, the alignment loss function can use mean squared error, cross-entropy loss, or contrast loss, etc.

[0042] At the same time, a self-attention module is adopted in the self-attention encoder. Through self-correlation operations, the time-frequency diagram information is comprehensively analyzed, enabling the neural network to more fully represent the time-frequency diagram features. In Figure 2 , Embedded Patches represents the embedding layer, layer Norm represents the normalization layer, Multi-HeadAttention represents the self-attention module, and MLP represents the multi-layer perception module.

[0043] S16. Obtain the target crack type of the target acoustic emission data output by the Transformer model; the target crack type includes tensile cracks and shear cracks.

[0044] It can be understood that the trained Transformer model is used to process the target acoustic emission data. The model analyzes the input acoustic emission time-frequency diagram and predicts the crack type corresponding to these data based on the features it has learned. The output of the model is the prediction result of the crack type of these acoustic emission data. The target crack type refers to the crack category that the model needs to identify. In this scenario, cracks are divided into two main types: tensile cracks and shear cracks.

[0045] In the above method for identifying tensile and shear cracks in tight sandstone, the acoustic emission time-frequency diagram data of the target tight sandstone is collected and input into the trained Transformer model to identify the crack type. The Transformer model includes an autoencoder neural network based on the self-attention mechanism, a fully connected layer network, and a softmax layer network. During the pre-training process, a representation alignment constraint is added, that is, when training the autoencoder neural network, the decoder is restricted from learning the encoded information, and thus the encoding task is all constrained in the self-attention encoder, effectively improving the feature representation ability of the self-attention encoder; at the same time, a self-attention module is adopted in the self-attention encoder, and the time-frequency diagram information is comprehensively analyzed through self-correlation operations, enabling the neural network to more fully represent the time-frequency diagram features. Then, a training labeled data set is obtained through experiments, and then neural network training is carried out to obtain accurate identification of the propagation of tensile and shear cracks. Using the trained model for identification can effectively distinguish tensile cracks and shear cracks, improve the accuracy and efficiency of crack identification in tight sandstone, achieve accurate classification of crack types, and has strong robustness and real-time identification ability.

[0046] In one embodiment, in the above method for identifying tensile and shear cracks in tight sandstone, the autoencoder neural network includes a self-attention encoder and a decoder. The self-attention encoder includes an embedding layer, a normalization layer, a self-attention module, and a multi-sensory module. The alignment loss function of the self-attention encoder is:

[0047]

[0048] Among them, Z m represents the masked block feature encoding obtained through the self-attention encoder, represents the predicted feature encoding output by the self-attention encoder.

[0049] It can be understood that the autoencoder neural network includes a self-attention encoder and a decoder. Among them, the main function of the self-attention encoder is to extract and encode the input data through the self-attention mechanism. The self-attention mechanism allows the encoder to capture global information and long-range dependencies when processing the sequence, rather than just local context. This mechanism enables each input element to pay attention to other elements in the entire input sequence, thus more comprehensively understanding the patterns and relationships of the data. Through the layer-by-layer processing of the embedding layer, self-attention module, normalization layer, and multi-layer perceptron module, the self-attention encoder can transform the input data into high-dimensional feature representations, which contain rich information of the input sequence and lay the foundation for subsequent decoding or other tasks. The role of the decoder is to decode the feature representation generated by the encoder into the target output sequence or target form. In this application, the decoder is used to reconstruct the acoustic emission time-frequency map.

[0050] By directly comparing the predicted feature encoding output by the encoder and the actual masked block feature encoding, the alignment loss function can ensure that the features learned by the encoder are closer to the true target. This direct supervision signal helps the encoder be more accurate in the feature extraction stage. The alignment loss function promotes the consistency of feature representations, enabling the encoder to better learn and capture the potential patterns and structures of the input data. This optimization process can improve the model's ability to understand the data, thereby achieving higher accuracy and robustness in the task.

[0051] In one embodiment, for the above method for identifying tensile and shear cracks in tight sandstone, the training process of the Transformer model includes a pre-training process, and the pre-training process includes the steps of:

[0052] Perform image segmentation on each existing pre-training acoustic emission data respectively to obtain image patches of each pre-training acoustic emission data, and perform fusion encoding of image encoding and position encoding on the image patches of each pre-training acoustic emission data to obtain a pre-training dataset.

[0053] It can be understood that after the pre-training acoustic emission data is segmented into image patches, each patch is subjected to fusion processing of image encoding and position encoding, thereby generating a fusion feature encoding containing image and position information, and finally constructing a pre-training dataset.

[0054] After randomly masking the pre-training dataset, input the unmasked and corresponding masked block pre-training data into the self-attention encoder respectively to obtain each predicted feature encoding and the corresponding masked block feature encoding.

[0055] It can be understood that after randomly masking the pre-training dataset, the unmasked part and the corresponding masked block are respectively input into the self-attention encoder. The unmasked part generates a predicted feature encoding through the encoder, while the masked block generates a true masked block feature encoding through the encoder.

[0056] Optimize the self-attention encoder according to the alignment loss function, each predicted feature encoding, and the corresponding mask block feature encodings.

[0057] It can be understood that the alignment loss function is used to compare each predicted feature encoding with the corresponding mask block feature encoding, measure the difference between them, and optimize the parameters of the self-attention encoder by minimizing this difference, so that the encoder can more accurately predict the features of the masked part. This process strengthens the encoder's understanding and learning ability of the input features, thereby improving the model's characterization ability of time-frequency map features.

[0058] Reconstruct each pre-trained acoustic emission data according to each predicted feature encoding to obtain each pre-trained reconstructed acoustic emission data.

[0059] It can be understood that according to each predicted feature encoding, it is transformed into the corresponding acoustic emission time-frequency map. That is, the pre-trained acoustic emission data is reconstructed through the decoding process, and finally the pre-trained reconstructed acoustic emission data with a similar structure to the original data is obtained.

[0060] Optimize the autoencoding neural network according to the reconstruction loss function, each pre-trained reconstructed acoustic emission data, and each pre-trained acoustic emission data.

[0061] It can be understood that the reconstruction loss function is used to compare the pre-trained reconstructed acoustic emission data with the original pre-trained acoustic emission data and calculate the difference between them. The parameters of the autoencoding neural network are optimized by minimizing this reconstruction error. This process aims to improve the network's reconstruction ability and make it more accurately learn and reproduce the features of the original acoustic emission data. The decoder only focuses on how to accurately reconstruct the input data without making excessive adjustments to the feature representations generated by the encoder.

[0062] To verify the learning effect of extracting time-frequency map features, the method of missing information restoration is used to test the time-frequency map feature extraction ability of the Transformer model. That is, only part of the time-frequency map information is given, and then the model is required to restore all the time-frequency map information based on the existing prior information. The closer the restoration degree is to the real image, the better the learning effect of the model. The test results are as Figure 3 shown in the time-frequency map generated based on the pre-trained model, where (a) represents the original image, (b) represents the input image, and (c) represents the reconstructed image.

[0063] From Figure 3It can be seen that when a randomly occluded 75% image is input to the model, the pre-trained model utilizes the prior information (the unoccluded part) of the input and reconstructs the time-frequency map based on the learned time-frequency map features. By comparing the original image and the reconstructed image, it can be known that the model can basically restore the entire time-frequency map, indicating that the pre-trained model has learned the variation law of the acoustic emission waveform and can infer information of other parts through partial information. Although there are still some details that are not fully restored, such as the area in the lower part of the high energy in the figure, there are multiple grid-like distributions in the original image, but they are not reflected in the reconstructed image. This may be caused by two reasons. First, the correlation between the input prior information and the waveform information of this part is not strong, and the information of this part cannot be obtained based on the known time-frequency map. In addition, it may also be that the model has not learned the relationship between the information of this part and the prior information. This requires further increasing the sample size of the training data, while increasing the number of model parameters and the dimension of the embedding layer, so as to train a more powerful pre-trained model for time-frequency map information extraction.

[0064] In one embodiment, for the above method for identifying tensile and shear cracks in tight sandstone, the reconstruction loss function is:

[0065]

[0066] where n represents the number of images in the training batch (i.e., the batch size), P represents the image set of the training batch, μ x represents the mean of the processed image, μ y represents the mean of the real image, σ x represents the variance of the processed image, σ y represents the variance of the real image, σ xy represents the covariance between the processed image and the real image, C1 represents the first constant, and C2 represents the second constant (C1 and C2 are set to maintain stability and prevent the denominator from being zero).

[0067] During the training process, we only calculate the loss of the encoding alignment constraint and do not perform gradient propagation. Finally, the total loss function is as follows:

[0068]

[0069] where L recon is the reconstruction loss function, λ is the adjustment coefficient, sg[·] represents not performing gradient propagation, Z m represents the masked block feature encoding obtained through the self-attention encoder, represents the predicted feature encoding output by the self-attention encoder.

[0070] In one embodiment, after the pre-training process of the Transformer model in the above method for identifying tensile and shear cracks in tight sandstone, the following steps are further included:

[0071] Conduct tensile and shear fracture tests on the training tight sandstone, and collect the acoustic emission data of each tensile fracture and the acoustic emission data of each shear fracture of the training tight sandstone as the training labeled data set; the acoustic emission data of tensile fracture is the acoustic emission time-frequency diagram of the tensile fracture test of the training tight sandstone, and the acoustic emission data of shear fracture is the acoustic emission time-frequency diagram of the shear fracture test of the training tight sandstone.

[0072] It can be understood that in different application scenarios, the steps for obtaining the training labeled data set are slightly different. For example, in the scenario of identifying tensile and shear cracks in tight sandstone, only the acoustic emission data of each tensile fracture and the acoustic emission data of each shear fracture of the training tight sandstone need to be collected as the training labeled data set, and the model is trained, and finally the trained model is applied to identify the tensile and shear cracks in tight sandstone. In the scenario of spatio-temporal expansion analysis of CO2 fracturing cracks, it is necessary to collect the acoustic emission data of each tensile fracture and the acoustic emission data of each shear fracture of the training tight sandstone before and after acidification as the training labeled data set (the acidification process is to treat with mud acid (12% HCl + 3% HF) for 24 hours), and finally the trained model is applied to identify the types of tensile and shear cracks during the CO2 fracturing process.

[0073] In this embodiment, taking the application scenario of identifying tensile and shear cracks in tight sandstone as an example, in order to obtain the training labeled data set, the acoustic emission data generated by the three-point bending test and the shear penetration test can be obtained through experiments. The specific test steps can be as follows: (1) Process and manufacture the semi-circular disk three-point bending specimen and the shear penetration specimen of the training tight sandstone respectively and dry them for standby; (2) Conduct the three-point bending test and the shear penetration test, and monitor and save the acoustic emission data during the experiment in real time; (3) Use the wavelet transform method to convert the acoustic emission time-frequency diagram into a two-dimensional image.

[0074] Specifically, a rock mechanics testing machine can be used to conduct the semi-circular disk three-point bending test and the shear penetration test respectively in a displacement loading control manner, and the displacement loading rate can be set at 0.02 mm / min. Figure 4 Represents the rock mechanics testing machine for conducting the three-point bending test and the shear penetration test. Among them, (a) represents the schematic diagram of the testing machine for conducting the three-point bending test, and (b) represents the schematic diagram of the testing machine for conducting the shear penetration test; Figure 5 Represents the physical diagrams of the processed semi-circular disk three-point bending specimen and the shear penetration specimen. Among them, (a) represents the physical diagram of the processed semi-circular disk three-point bending specimen, and (b) represents the physical diagram of the processed shear penetration specimen; Figure 6It shows the processed cross-sectional view and 3D view of the semi-circular disk three-point bending specimen. Among them, (a) shows the processed cross-sectional view of the semi-circular disk three-point bending specimen, and (b) shows the processed 3D view of the semi-circular disk three-point bending specimen; Figure 7 It shows the processed cross-sectional view and 3D view of the shear-through specimen. Among them, (a) shows the processed cross-sectional view of the shear-through specimen, and (b) shows the processed Figure 3 3D view; among them, the shear-through specimen is a cube with a height of 40 mm (it can also be processed into a cylinder). A circular crack with a depth of a = 4 mm and b = 24 mm is prefabricated on two parallel cross-sections of the specimen using a diamond drill bit with a diameter of 25 mm. The crack width t is 1.5 mm, and the diameter ID of the central region cylinder is 22 mm. After performing tensile and shear fracture tests on the trained tight sandstone, Figure 8 It shows the time-frequency diagram after the transformation of tensile and shear cracks. Among them, (a) shows the time-frequency diagram after the transformation of tensile cracks, and (b) shows the time-frequency diagram after the transformation of shear cracks. The abscissa in the figure represents the duration, which is 1024 μs in total, and the ordinate is the frequency, with a range of 0 - 500 kHz. The acoustic emission data obtained from the three-point bending test is marked as tensile cracks, and the acoustic emission data obtained from the shear-through test is marked as shear cracks.

[0075] The training labeled dataset is divided into a training dataset and a test dataset.

[0076] It can be understood that after mixing the generated time-frequency diagrams, most of them are randomly selected as the training set. For example, about 75% - 90% can be selected as the training dataset to train the Transformer model, that is, to let the model learn the features and patterns of the data. The remaining part is used as the test dataset to evaluate the performance of the model and check whether the model can generalize well to unseen data. Specifically, in this embodiment, 80% of the generated time-frequency diagrams are randomly selected as the training dataset, and the remaining 20% is used as the test dataset.

[0077] The Transformer model is trained for multiple rounds according to the training dataset to optimize the parameters of the Transformer model. The Transformer model is verified according to the test dataset to evaluate the accuracy of the Transformer model.

[0078] It can be understood that when training the Transformer model, by using the training dataset in multiple rounds (Epochs), the model will continuously adjust its internal parameters to minimize the error between the predicted output and the true label. After training, the unseen test dataset is used to verify the model, and the generalization ability and accuracy of the model are measured through evaluation metrics (such as accuracy, etc.).

[0079] Specifically, such as Figure 9As shown by the accuracy and loss distribution during the model training process, where Accuracy represents the correct rate, Loss represents the loss, Train accuracy represents the training accuracy, Valid accuracy represents the validation accuracy, Train loss represents the training loss, and Valid loss represents the validation loss. The training dataset was trained for 60 epochs (training rounds). The model performance was unstable during the first 30 epochs of training, indicating that the model was continuously adjusting the feature weights to seek higher accuracy. After the 35th epoch of training, the accuracy of the model on the test set stabilized at around 95% and no longer increased, while the accuracy of the training set continued to increase. This indicates that the model has overfitted. The neural network model began to use the unique noise information in the acoustic emission time-frequency diagram as the recognition basis to further improve the accuracy. This phenomenon can be alleviated by further increasing the quantity of training data. In this training, the training with the highest accuracy (accuracy of 95.4%) was finally selected as the final model for this training.

[0080] In addition, by obtaining the attention of the Transformer to analyze the feature weight distribution of the neural network, a discrimination basis for the neural network was established, and then the differences between sandstone tensile and shear failures were obtained. The method for the recognition basis is as follows:

[0081] ① Calculate the scores of each attention layer in each layer of the Transformer based on correlation;

[0082] ② By combining correlation and gradient information, use iteration to eliminate negative impacts;

[0083] ③ Integrate the information into the attention map to obtain a class-specific visualization map of the self-attention model;

[0084] ④ Finally, obtain the judgment basis of the self-attention model that can be shown in the generated heat map, and obtain the differences in the time-frequency diagrams of rock fractures with and without acidification.

[0085] Figure 10 Represents the recognition basis map for non-acidified tensile cracks (in the method of this embodiment, the tight sandstone was not acidified. To distinguish it from the recognition basis map of tensile cracks after acidification in other embodiments below, it is named the recognition basis map for non-acidified tensile cracks). Among them, (a) represents the time-frequency diagram of non-acidified tensile cracks, and (b) represents the weight map of non-acidified tensile cracks. Figure 11 Represents the recognition basis map for non-acidified shear cracks. Among them, (a) represents the time-frequency diagram of non-acidified shear cracks, and (b) represents the weight map of non-acidified shear cracks. Figure 10 - Figure 11 The bright areas in it are the important bases for model recognition. From Figure 10 and Figure 11It can be seen that the main identification bases for tensile and shear cracks are distributed in the low-frequency region. Moreover, compared with shear cracks, the acoustic emission time-frequency diagrams generated by tensile cracks have a large amount of high-energy distribution in the high-frequency part. At the same time, the frequencies of the main identification regions of the model for shear cracks are often lower than those of tensile cracks. For example, compared with the tensile cracks in Figure 10 , the main discrimination basis for the time-frequency diagrams of the model for shear cracks ([ Figure 11 ) is in the low-frequency region around the high energy. This may be determined by the fracture mode of shear fissures. Generally, the fissure surfaces of shear fissures are relatively large, and the waveforms generated during fracture have low frequencies and high energies.

[0086] In one embodiment, for the above method for identifying tensile and shear cracks in tight sandstone, the steps of conducting tensile and shear fracture tests on the training tight sandstone and collecting the acoustic emission data of each tensile fracture and each shear fracture of the training tight sandstone as the training labeled data set include: conducting tensile and shear fracture tests on the training tight sandstone before and after acidification, and collecting the acoustic emission data of each tensile fracture and each shear fracture of the training tight sandstone before and after acidification as the training labeled data set; the acidification process is to treat with mud acid (12% HCl + 3% HF) for 24 hours.

[0087] It can be understood that in the scenario of spatio-temporal expansion analysis of CO2 fracturing cracks, it is necessary to collect the acoustic emission data of each tensile fracture and each shear fracture of the training tight sandstone before and after acidification as the training labeled data set (the acidification process is to treat with mud acid (12% HCl + 3% HF) for 24 hours).

[0088] The specific test steps can be as follows: (1) Process and fabricate semi-circular three-point bending specimens and shear-through specimens of the training tight sandstone respectively and dry them for standby; (2) Conduct tensile and shear fracture acoustic emission tests on acid-treated and non-acid-treated tight sandstone, where the acid-treated specimens are treated with mud acid (12HCl + 3% HF) for 24 hours and the acoustic emission data during the experiment are monitored and saved in real time; (3) Use the wavelet transform method to convert the acoustic emission time-frequency diagram into a two-dimensional image.

[0089] Specifically, a rock mechanics testing machine was used to process semi-circular disk three-point bending specimens and shear-through specimens respectively, and tensile and shear crack acoustic emission experiments were carried out before and after acidification. Among them, the acid-treated specimens were treated with mud acid (12% HCl + 3% HF) for 24 hours. The displacement loading mode was adopted for the tensile and shear crack propagation tests, and the loading rate was 0.02 mm / min. The acquired acoustic emission time-frequency diagrams were transformed in time-frequency using the wavelet transform method, and then the one-dimensional features of the acoustic emission waveforms were transformed into two-dimensional features for convenient extraction and utilization in the later stage. The acoustic emission data obtained from the three-point bending test were marked as type I cracks, and the acoustic emission data obtained from the shear-through test were marked as type II cracks. Then, 80% of the randomly selected time-frequency diagrams generated by them were used as the training data set, and the remaining 20% was used as the test data set. The specific data distribution is shown in Table 1: the number of acquired AE waveforms.

[0090] Table 1:

[0091]

[0092] The characteristic weight distribution of the neural network can be analyzed by obtaining the attention of the Transformer, and the discrimination basis of the neural network can be established. The identification basis diagram for non-acidified tensile cracks is as shown in Figure 10 ; the identification basis diagram for non-acidified shear cracks is as shown in Figure 11 . Figure 12 It represents the identification basis diagram for acidified tension. Among them, (a) represents the time-frequency diagram of acidified tensile cracks, and (b) represents the weight diagram of acidified tensile cracks; Figure 13 It represents the identification basis diagram for acidified shear. Among them, (a) represents the time-frequency diagram of acidified shear cracks, and (b) represents the weight diagram of acidified shear cracks. Figure 10 - Figure 13 The bright areas in it are the important basis for model identification. As can be seen from Figure 10 - Figure 13 , compared with shear cracks, there is a large amount of high-energy distribution in the high-frequency part of the acoustic emission time-frequency diagram generated by tensile cracks. At the same time, more energy is distributed in the high-frequency part of the acoustic emission waveform generated by the fracture of acidified rock.

[0093] In Figure 10 - Figure 13 the identification basis distribution diagram, it is found that the main identification basis of the model for non-acidified tensile and shear cracks is distributed in the low-frequency region. And the identification basis of the model for acidified cracks is mainly distributed in the high-frequency region, especially the acoustic emission time-frequency diagram of acidified tensile cracks is particularly typical. The frequency of the main identification area of the model for shear cracks is often lower than that of tensile cracks. For example, compared with the non-acidified tension in Figure 10 , the main discrimination basis of the model for the non-acidified shear Figure 11 time-frequency diagram is in the low-frequency region around the high energy. After acidification treatment, the key identification area of the acoustic emission time-frequency diagram generated by shear fractures of the model shifts to the high frequency (Figure 13 )。

[0094] In one embodiment, in the above method for identifying tensile and shear cracks in tight sandstone, the step of collecting target acoustic emission data of the target tight sandstone includes: performing acid treatment and CO2 fracturing tests on the target tight sandstone to obtain the target acoustic emission data of the target tight sandstone. In the step of obtaining the target crack type of the target acoustic emission data output by the Transformer model, the target crack type includes tensile cracks after acidification and shear cracks after acidification; after the step of obtaining the target crack type of the target acoustic emission data, it further includes: performing spatio-temporal expansion analysis according to the target crack type.

[0095] It can be understood that in the scenario of spatio-temporal expansion analysis of CO2 fracturing cracks, after pre-training, in the application process, the steps are specifically: performing acid treatment and CO2 fracturing tests on the target tight sandstone to obtain the target acoustic emission data of the target tight sandstone, obtaining the target crack type (identified as tensile cracks after acidification or shear cracks after acidification) of the target acoustic emission data output by the Transformer model, and finally, spatio-temporal expansion analysis of CO2 fracturing cracks can be performed according to the output crack type.

[0096] Specifically, in this embodiment, cracks mainly occur around 1000s and 1500s. Around 1000s, cracks mainly occur around the borehole, while two cracks mainly occur around 1500s. Among them, acidified tensile mainly occurs in the early stage of macroscopic fracture. For example, the red event near the CO2 injection port occurs around 100s, and the remaining acidified tensile events occur between 1400 - 1500s. This stage is the stage of microcrack propagation inside the sandstone. At the same time, the positions of acidified tensile generated in this stage are mainly concentrated in the area far from the borehole. In addition, acidified shear fractures are also mainly divided into two parts. The first part is the stage of increasing CO2 pressure around 1000s, and the other part is between 1300 - 1500s. This stage is the stage where a series of pressure drops (rapid crack propagation) are generated by fluid pressure. At the same time, during the entire fracturing process, acidified shear events are mainly concentrated after 1300s according to the occurrence time and form two relatively obvious cracks. Non-acidified tensile also contributes a large number of fractures. Non-acidified shear fractures are mainly divided into two parts. The first part is generated during the stress adjustment stage before 200s, and the other part is between 1300 - 1500s. This stage is the stage where a series of pressure drops (rapid crack propagation) are generated by fluid pressure. Based on acoustic emission location analysis, the spatio-temporal distribution of different types of fractures during the entire fracturing process can be obtained.

[0097] It should be understood that although Figure 1The steps in the flowchart are shown in sequence according to the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least a part of the steps in

[0098] In one embodiment, a system for identifying tensile and shear cracks in tight sandstone is provided, including:

[0099] A collection module for collecting target acoustic emission data of target tight sandstone; the target acoustic emission data is the acoustic emission time-frequency diagram of the target tight sandstone;

[0100] An input module for inputting the target acoustic emission data into a trained Transformer model; the Transformer model includes an autoencoder neural network based on the self-attention mechanism, a fully connected layer network, and a softmax layer network; during the training process, a representation alignment constraint is added to the autoencoder neural network;

[0101] An output module for obtaining the target crack type of the target acoustic emission data output by the Transformer model; the target crack type includes tensile cracks and shear cracks.

[0102] The above system for identifying tensile and shear cracks in tight sandstone collects the acoustic emission time-frequency diagram data of the target tight sandstone and inputs it into a trained Transformer model to identify the crack type. The Transformer model includes an autoencoder neural network based on the self-attention mechanism, a fully connected layer network, and a softmax layer network. A representation alignment constraint is added during the pre-training process, that is, when training the autoencoder neural network, the decoder is restricted from learning the encoded information, and thus the encoding task is fully constrained in the self-attention encoder, effectively improving the feature representation ability of the self-attention encoder; at the same time, a self-attention module is adopted in the self-attention encoder, and the time-frequency diagram information is comprehensively analyzed through self-correlation operations, enabling the neural network to more fully represent the time-frequency diagram features. Then, a training labeled data set is obtained through experiments, and then neural network training is carried out to obtain accurate identification of the tensile and shear crack propagation. Using the trained model for identification can effectively distinguish tensile cracks and shear cracks, improve the accuracy and efficiency of tight sandstone crack identification, achieve accurate classification of crack types, and has strong robustness and real-time identification ability.

[0103] In one embodiment, the above-mentioned identification system for tensile and shear cracks in tight sandstone further includes a training module for training the Transformer model.

[0104] In one embodiment, in the above-mentioned identification system for tensile and shear cracks in tight sandstone, the autoencoder neural network includes a self-attention encoder and a decoder. The self-attention encoder includes an embedding layer, a normalization layer, a self-attention module, and a multi-perception module. The alignment loss function of the self-attention encoder is:

[0105]

[0106] where Z m represents the masked block feature encoding obtained through the self-attention encoder, represents the predicted feature encoding output by the self-attention encoder.

[0107] In one embodiment, in the above-mentioned identification system for tensile and shear cracks in tight sandstone, the training module includes a pre-training module, and the pre-training module includes:

[0108] An acquisition sub-module for respectively performing image segmentation on each existing pre-trained acoustic emission data to obtain image patches of each pre-trained acoustic emission data, and performing fusion encoding of image encoding and position encoding on the image patches of each pre-trained acoustic emission data to obtain a pre-training data set;

[0109] An input sub-module for randomly masking the pre-training data set and then respectively inputting the unmasked and corresponding masked-block pre-training data into the self-attention encoder to obtain each predicted feature encoding and the corresponding masked-block feature encoding;

[0110] A first optimization sub-module for optimizing the self-attention encoder according to the alignment loss function, each predicted feature encoding, and the corresponding masked-block feature encoding;

[0111] A reconstruction sub-module for reconstructing each pre-trained acoustic emission data according to each predicted feature encoding to obtain each pre-trained reconstructed acoustic emission data;

[0112] A second optimization sub-module for optimizing the autoencoder neural network according to the reconstruction loss function, each pre-trained reconstructed acoustic emission data, and each pre-trained acoustic emission data.

[0113] In one embodiment, in the above-mentioned identification system for tensile and shear cracks in tight sandstone, the reconstruction loss function is:

[0114]

[0115] where n represents the number of images in the training batch, P represents the image set of the training batch, μ xDenote the mean of the processed image, μ y Denote the mean of the real image, σ x Denote the variance of the processed image, σ y Denote the variance of the real image, σ xy Denote the covariance between the processed image and the real image. C1 represents the first constant, and C2 represents the second constant.

[0116] In one embodiment, for the above-mentioned identification system of tensile and shear cracks in tight sandstone, the training module further includes:

[0117] A training labeled dataset acquisition module, which is used to conduct tensile and shear fracture tests on the training tight sandstone, and collect each tensile fracture acoustic emission data and each shear fracture acoustic emission data of the training tight sandstone as the training labeled dataset; the tensile fracture acoustic emission data is the acoustic emission time-frequency diagram of the tensile fracture test of the training tight sandstone, and the shear fracture acoustic emission data is the acoustic emission time-frequency diagram of the shear fracture test of the training tight sandstone;

[0118] A segmentation module, which is used to divide the training labeled dataset into a training dataset and a test dataset;

[0119] A round training module, which is used to train the Transformer model for multiple rounds according to the training dataset, and optimize the parameters of the Transformer model;

[0120] A verification module, which is used to verify the Transformer model according to the test dataset and evaluate the accuracy of the Transformer model.

[0121] In one embodiment, for the above-mentioned identification system of tensile and shear cracks in tight sandstone, in the training labeled dataset acquisition module, conduct tensile and shear fracture tests on the training tight sandstone before and after acidification, and collect each tensile fracture acoustic emission data and each shear fracture acoustic emission data of the training tight sandstone before and after acidification as the training labeled dataset; the acidification process is to treat with mud acid (12% HCl + 3% HF) for 24 hours.

[0122] In one embodiment, for the above-mentioned identification system of tensile and shear cracks in tight sandstone, in the collection module, conduct acid treatment and CO2 fracturing tests on the target tight sandstone to obtain the target acoustic emission data of the target tight sandstone. In the output module, the target crack types include tensile cracks after acidification and shear cracks after acidification. It further includes an analysis module, which is used to conduct spatio-temporal expansion analysis according to the target crack types.

[0123] For the specific limitations of the identification system for tensile and shear cracks in tight sandstone, reference can be made to the limitations of the identification method for tensile and shear cracks in tight sandstone described above, which will not be elaborated here. Each module in the above-mentioned identification system for tensile and shear cracks in tight sandstone can be implemented in whole or in part by software, hardware, or a combination thereof. The above-mentioned modules can be embedded in the processor of a computer device in hardware form or be independent of it, or can be stored in the memory of the computer device in software form, so as to facilitate the processor to call and execute the operations corresponding to each of the above-mentioned modules.

[0124] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0125] Collect the target acoustic emission data of the target tight sandstone; the target acoustic emission data is the acoustic emission time-frequency diagram of the target tight sandstone;

[0126] Input the target acoustic emission data into the trained Transformer model; the Transformer model includes an autoencoder neural network based on the self-attention mechanism, a fully connected layer network, and a softmax layer network; during the training process, a representation alignment constraint is added to the autoencoder neural network;

[0127] Obtain the target crack type of the target acoustic emission data output by the Transformer model; the target crack type includes tensile cracks and shear cracks.

[0128] It can be understood that in addition to the memory and processor involved above, the above-mentioned computer device also includes other software and hardware components not listed in this specification. Specifically, it can be determined according to the model of the specific computer device in different application scenarios, and will not be listed and elaborated one by one in this specification.

[0129] In one embodiment, when the processor executes the computer program, it can also implement the additional steps or sub-steps in each embodiment of the above-mentioned identification method for tensile and shear cracks in tight sandstone.

[0130] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0131] The embodiments described above merely represent several implementation manners of the present application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A method for identifying tensile and shear cracks in tight sandstone, characterized in that, Including steps: Collecting target acoustic emission data of the target tight sandstone; the target acoustic emission data is the acoustic emission time-frequency diagram of the target tight sandstone; Inputting the target acoustic emission data into a trained Transformer model; the Transformer model includes an auto-encoding neural network based on the self-attention mechanism, a fully connected layer network, and a softmax layer network; during the training process, an alignment constraint is added to the auto-encoding neural network; the auto-encoding neural network includes a self-attention encoder and a decoder, the self-attention encoder includes an embedding layer, a normalization layer, a self-attention module, and a multi-perception module, and the alignment loss function of the self-attention encoder is: ； Among them, represents the masked block feature encoding obtained by the self-attention encoder, represents the predicted feature encoding output by the self-attention encoder; Obtaining the target crack type of the target acoustic emission data output by the Transformer model; the target crack type includes tensile cracks and shear cracks; The training process of the Transformer model includes a pre-training process, and the pre-training process includes steps: performing image segmentation on each existing pre-training acoustic emission data respectively to obtain image patches of each pre-training acoustic emission data, performing fusion coding of image coding and position coding on the image patches of each pre-training acoustic emission data to obtain a pre-training data set; after randomly masking the pre-training data set, inputting the unmasked and corresponding masked block pre-training data into the self-attention encoder respectively to obtain each predicted feature coding and the corresponding masked block feature coding; optimizing the self-attention encoder according to the alignment loss function, each predicted feature coding, and the corresponding masked block feature coding; reconstructing each pre-training acoustic emission data according to each predicted feature coding to obtain each pre-training reconstructed acoustic emission data; optimizing the auto-encoding neural network according to the reconstruction loss function, each pre-training reconstructed acoustic emission data, and each pre-training acoustic emission data; The reconstruction loss function is: ； Among them, n represents the number of images in the training batch, P represents the image set of the training batch, represents the mean of the processed images, represents the mean of the real images, represents the variance of the processed images, represents the variance of the real images, represents the covariance between the processed images and the real images, C 1 represents the first constant, C 2 represents the second constant.

2. The method for identifying tensile and shear cracks in tight sandstone according to claim 1, wherein After the Transformer model performs the pre-training process, it further includes steps: Performing tensile and shear fracture tests on the training tight sandstone, and collecting each tensile fracture acoustic emission data and each shear fracture acoustic emission data of the training tight sandstone as a training labeled data set; the tensile fracture acoustic emission data is the acoustic emission time-frequency diagram of the tensile fracture test of the training tight sandstone, and the shear fracture acoustic emission data is the acoustic emission time-frequency diagram of the shear fracture test of the training tight sandstone; Dividing the training labeled data set into a training data set and a test data set; Performing multiple rounds of training on the Transformer model according to the training data set to optimize the parameters of the Transformer model; Verifying the Transformer model according to the test data set to evaluate the accuracy of the Transformer model.

3. The method for identifying tensile and shear cracks in tight sandstone according to claim 2, characterized in that The step of performing tensile and shear fracture tests on the training tight sandstone and collecting each tensile fracture acoustic emission data and each shear fracture acoustic emission data of the training tight sandstone as a training labeled data set includes: Tensile and shear fracture tests are carried out on the trained tight sandstone before and after acidification, and each tensile fracture acoustic emission data and each shear fracture acoustic emission data of the trained tight sandstone before and after acidification are collected as a training labeled data set; the acidification process is to treat with mud acid for 24 hours; the mud acid is a solution containing 12% HCl and 3% HF.

4. The method for identifying tensile and shear cracks in tight sandstone according to claim 3, wherein the step of collecting the target acoustic emission data of the target tight sandstone includes: performing acid treatment and CO2 fracturing test on the target tight sandstone to obtain the target acoustic emission data of the target tight sandstone; in the step of obtaining the target crack type of the target acoustic emission data output by the Transformer model, the target crack type includes tensile cracks after acidification and shear cracks after acidification; after the step of obtaining the target crack type of the target acoustic emission data, it further includes: performing spatio-temporal expansion analysis according to the target crack type.

5. An identification system for tensile and shear cracks in tight sandstone, characterized in that It includes: a collection module for collecting the target acoustic emission data of the target tight sandstone; the target acoustic emission data is the acoustic emission time-frequency diagram of the target tight sandstone; an input module for inputting the target acoustic emission data into a trained Transformer model; the Transformer model includes a self-encoding neural network based on a self-attention mechanism, a fully connected layer network, and a softmax layer network; during training, a representation alignment constraint is added to the self-encoding neural network; the self-encoding neural network includes a self-attention encoder and a decoder, and the self-attention encoder includes an embedding layer, a normalization layer, a self-attention module, and a multi-perception module. The alignment loss function of the self-attention encoder is: ； Among them, represents the masked block feature encoding obtained by the self-attention encoder, represents the predicted feature encoding output by the self-attention encoder; an output module for obtaining the target crack type of the target acoustic emission data output by the Transformer model; the target crack type includes tensile cracks and shear cracks; A training module, the training module includes a pre-training module, and the pre-training module includes: an acquisition sub-module, configured to perform image segmentation on each existing pre-training acoustic emission data respectively to obtain image patches of each pre-training acoustic emission data, and perform fusion encoding of image encoding and position encoding on the image patches of each pre-training acoustic emission data to obtain a pre-training data set; an input sub-module, configured to randomly mask the pre-training data set, and then input the unmasked and corresponding masked pre-training data into the self-attention encoder respectively to obtain each of the predicted feature encodings and the corresponding masked block feature encodings; a first optimization sub-module, configured to optimize the self-attention encoder according to the alignment loss function, each of the predicted feature encodings and the corresponding masked block feature encodings; a reconstruction sub-module, configured to reconstruct each pre-training acoustic emission data according to each of the predicted feature encodings to obtain each pre-training reconstructed acoustic emission data; a second optimization sub-module, configured to optimize the auto-encoding neural network according to the reconstruction loss function, each of the pre-training reconstructed acoustic emission data and each of the pre-training acoustic emission data; The reconstruction loss function is: ； Among them, n represents the number of images in the training batch, P represents the image set of the training batch, represents the mean of the processed images, represents the mean of the real images, represents the variance of the processed images, represents the variance of the real images, represents the covariance between the processed images and the real images, C 1 represents the first constant, C 2 represents the second constant.

6. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, the steps of the method for identifying tensile and shear cracks in tight sandstone according to any one of claims 1 to 4 above are implemented.

Images