Damage identification method for composite materials based on transfer learning and improved residual network

Abstract

The present invention discloses a composite material damage identification method based on transfer learning and improved residual network, and relates to the technical field of acoustic emission data processing and image recognition methods. The method comprises: obtaining an acoustic emission data set of composite material stretching, and converting the acoustic emission data set into Mel spectrum cepstral coefficients to generate a Mel spectrum cepstral coefficient spectrum data set; building a deep residual shrinkage network model based on a ResNet model, calling the pre-trained weights trained on the graph network by the ResNet model, and using the other weights in the pre-trained weights except the residual block to initialize the weights of the deep residual shrinkage network model to obtain a deep residual shrinkage network model based on transfer learning; training with a Mel spectrum cepstral coefficient spectrum data set to obtain a trained deep residual shrinkage network model based on transfer learning; using the trained deep residual shrinkage network model based on transfer learning to identify the damage type of the composite material to be identified.

Description

Technical Field

[0001] The present application relates to the technical field of acoustic emission data processing and image recognition methods, and in particular to a composite material damage recognition method based on transfer learning and improved residual network. Background Art

[0002] At present, composite materials have been widely used in aerospace, construction engineering, transportation, shipbuilding, chemical equipment and medical equipment due to their advantages of high strength, light weight and corrosion resistance. However, during service, composite materials will be affected by various environmental factors and may suffer various damages. Some serious damages may even lead to structural failure, which in turn leads to safety risks and economic losses. Therefore, it is of great significance to study damage identification of composite materials. Among the many non-destructive testing technologies, acoustic emission non-destructive testing has become one of the most commonly used testing methods due to its high reliability, low environmental requirements and wide range of applications.

[0003] Existing technologies rely on manual analysis methods and traditional deep learning network models such as CNN for damage identification. However, traditional manual analysis methods have shortcomings in efficiency, misjudgment rate and cost.

[0004] Traditional CNN network model damage identification requires a large amount of labeled data for training. However, in complex industrial production environments, collecting a large amount of damage data containing labeled information is difficult and time-consuming, which reduces the accuracy and reliability of damage identification. Summary of the invention

[0005] Based on this, it is necessary to provide a composite material damage identification method based on transfer learning and improved residual network to address the above technical problems.

[0006] This manual adopts the following technical solutions:

[0007] This specification provides a composite material damage identification method based on transfer learning and improved residual network, including:

[0008] Acquire an acoustic emission data set of composite material stretching, and perform Mel-spectrum cepstral coefficient conversion on the acoustic emission data set to generate a Mel-spectrum cepstral coefficient spectrum data set;

[0009] Based on the ResNet model, a deep residual shrinkage network model is constructed, the pre-trained weights of the ResNet model trained on the graph network are called, and the weights of the deep residual shrinkage network model are initialized using the other weights in the pre-trained weights except the residual block, so as to obtain a deep residual shrinkage network model based on transfer learning;

[0010] Trained with the Mel-spectrogram cepstral coefficients dataset, a trained deep residual shrinkage network model based on transfer learning was obtained;

[0011] The trained deep residual shrinkage network model based on transfer learning is used to identify the damage type of the composite material to be identified.

[0012] Preferably, the acoustic emission data include: acoustic emission data of the composite material under three damage modes: fiber breakage, matrix cracking and delamination.

[0013] Preferably, the method further comprises: preprocessing the Mel frequency cepstral coefficient spectrum data set, which specifically comprises:

[0014] Scale and crop the image of the input Mel frequency cepstral coefficient map dataset to a specified size to generate image blocks of different sizes;

[0015] Flip the image horizontally and left-right;

[0016] Convert images in PIL image format or array format to PyTorch Tensor format and automatically scale pixel values ​​to the range of [0, 1];

[0017] Normalize the image using the preset mean and standard deviation.

[0018] Preferably, the constructing of the deep residual shrinkage network model includes:

[0019] Using multiple PSA blocks, all convolution blocks in the residual block of the ResNet model are replaced to obtain the EPSANet model;

[0020] A soft threshold denoising module is added after the convolution operation of the second layer PSA block in the EPSANet model to obtain a deep residual shrinkage network model.

[0021] Preferably, the stacking method between the multiple PSA blocks is the stacking method of residual blocks in a deep residual network.

[0022] Preferably, the calling of the pre-trained weights trained on the graph network by the ResNet model and using other weights in the pre-trained weights except the residual block to initialize the weights of the deep residual shrinkage network model specifically includes:

[0023] Call the weights of the 7×7 convolutional layer of ResNet that have been trained on the PyTorch image database to initialize the 7×7 convolutional layer of the deep residual contraction network model;

[0024] Call the weights of the ResNet's maximum pooling layer that have been trained on the PyTorch image database to initialize the maximum pooling layer of the deep residual shrinkage network model;

[0025] Call the weights of the average pooling layer of ResNet that have been trained on the PyTorch image database to initialize the average pooling layer of the deep residual shrinkage network model;

[0026] Call the weights of the fully connected layer of ResNet that have been trained on the PyTorch image database to initialize the fully connected layer of the deep residual shrinkage network model.

[0027] Preferably, the training is performed using a Mel-spectrogram cepstral coefficient spectrum dataset, specifically including:

[0028] The data set is divided into K subsets using the ten-fold cross validation method. K training and validation are performed. Each time, K-1 subsets are used as training sets and the other subset is used as validation sets. The accuracy is calculated.

[0029] When the accuracy reaches the specified iteration cycle and the accuracy no longer continues to rise, the training effect of the deep residual shrinkage network model based on transfer learning reaches the optimal effect and is saved.

[0030] Preferably, the training using the Mel-spectrum cepstral coefficient spectrum data set also includes:

[0031] In the training process of the deep residual shrinkage network model based on transfer learning, the Adam optimization algorithm and the cross entropy loss function are used to optimize the deep residual shrinkage network model based on transfer learning; and the automatic adjustment hyperparameters are set, and the learning rate is halved after every 10 training cycles;

[0032] The Adam optimization algorithm has the following formula:

[0033] ;

[0034] ;

[0035] In the formula, m t is the first-order moment estimate, v t is the second-order moment estimate, β 1 is the exponential decay rate of the first-order moment estimate, β 2 is the exponential decay rate of the second-order moment estimate, g i is the gradient;

[0036] The cross entropy loss function is formulated as:

[0037] ;

[0038] In the formula, is the cross entropy loss function, q is the number of categories, , is the one-hot encoding of the true label, The model predicts i The probability of the class.

[0039] At least one of the above technical solutions adopted in this specification can achieve the following beneficial effects:

[0040] The present invention introduces a deep learning method based on transfer learning in the analysis of acoustic emission data, and constructs a deep residual shrinkage network model based on transfer learning. The model introduces a pre-trained model through a transfer learning strategy, and calls the pre-trained weights of the pre-trained model to improve the deep residual network model, thereby enhancing the model performance under small sample acoustic emission data conditions, so that the composite material damage signal still has a high recognition accuracy when data is scarce.

[0041] In addition, since the damage signal of the composite material itself is very weak, it is difficult for the existing technology to effectively extract and identify the acoustic emission damage features, and the damage signal acquisition of the composite material is usually carried out in a high-noise environment. The traditional model is sensitive to noise, resulting in inaccurate recognition. Therefore, the present invention converts the acoustic emission data set into Mel-spectrum cepstral coefficients, and the model also introduces a pyramid splitting attention module and a soft threshold denoising module to achieve effective extraction of multi-scale damage features, solving the problem of poor performance of damage feature extraction and poor noise resistance in the existing technology when the composite material damage signal is weak. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0043] Figure 1 A schematic diagram of a composite material damage identification method based on transfer learning and improved residual network provided in this specification;

[0044] Figure 2 This specification provides a composite material damage identification method based on transfer learning and improved residual network MFCC map of different damage modes of the composite material;

[0045] Figure 3 A ResNet network and EPSANet network structure diagram based on transfer learning and improved residual network composite material damage identification method provided in this manual;

[0046] Figure 4A schematic diagram of the structure of a deep residual shrinkage network model based on transfer learning and a composite material damage identification method based on transfer learning and improved residual network provided in this specification;

[0047] Figure 5 A comparison chart of model accuracy and loss function results of different soft threshold denoising module configurations based on transfer learning and improved residual network composite material damage identification method provided in this manual;

[0048] Figure 6 The experimental results of the transfer learning based on transfer learning and improved residual network composite material damage identification method provided in this specification are compared with the fine-tuning of only the last layer of the network and all layers;

[0049] Figure 7 A LIME visualization diagram of a composite material damage identification method based on transfer learning and improved residual network provided in this manual.

[0050] Figure 8 This is a T-SNE dimension reduction visualization diagram of the same stage of the composite material damage identification method based on transfer learning and improved residual network provided in this manual. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of this specification more clear, the technical solutions of this application will be clearly and completely described below in combination with the specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments in the specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0052] The technical solutions provided by various embodiments of the present application are described in detail below in conjunction with the accompanying drawings.

[0053] Figure 1 A schematic diagram of a composite material damage identification method based on transfer learning and improved residual network provided in this specification specifically includes the following steps:

[0054] S101: Acquire acoustic emission data of composite material stretching, and convert the acoustic emission data through Mel spectrum cepstral coefficients to obtain a Mel spectrum cepstral coefficient spectrum data set, specifically including:

[0055] Acoustic emission data of composite material tension, including data of composite materials under three damage modes: fiber fracture, matrix cracking and delamination;

[0056] The acoustic emission data is converted by Mel-spectrometric cepstral coefficients, including:

[0057] Pre-emphasize the collected original acoustic emission signal to enhance the energy of the high-frequency part;

[0058] Divide the pre-emphasized signal into short time frames, usually using overlapping window techniques;

[0059] Fast Fourier Transform (FFT): Apply FFT to each frame of data to convert the time domain signal to the frequency domain;

[0060] Mel filter bank design: Design a set of Mel filters that are evenly distributed on the Mel frequency axis to simulate the human ear's perception of different frequencies;

[0061] Mel filter bank application: The spectrum of each frame is passed through the Mel filter bank to obtain the Mel spectrum.

[0062] Logarithmic operation: Take the logarithm of the Mel spectrum to enhance the ability to distinguish low-energy components;

[0063] Discrete Cosine Transform (DCT): DCT is applied to the log-Mel spectrum of each frame to obtain Mel-frequency Cepstral Coefficients (MFCCs).

[0064] The image preprocessing of the Mel frequency cepstral coefficient spectrum dataset also includes:

[0065] The input image is randomly scaled and cropped to a size of 224x224. This operation is used to generate image patches of different sizes to enhance the generalization ability of the model;

[0066] Randomly flip the image horizontally (left-right flip) to increase data diversity and reduce overfitting of the model;

[0067] Convert images in PIL image or NumPy array format to PyTorch Tensor format, automatically scaling pixel values ​​to the range of [0, 1];

[0068] The images are normalized using the given mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225] to ensure that each channel has a relatively standard distribution during training.

[0069] S102: Based on the ResNet model, a deep residual shrinkage network model is constructed, the pre-trained weights of the ResNet model trained on the graph network are called, and the weights of the deep residual shrinkage network model are initialized using the other weights in the pre-trained weights except the residual block, to obtain a deep residual shrinkage network model based on transfer learning, specifically including:

[0070] Using the pyramid split attention module, the 3×3 convolution in the residual block Bottleneck of the deep residual network model is replaced to obtain an efficient pyramid split attention module;

[0071] According to the stacking method of residual blocks in the deep residual network, an efficient pyramid splitting attention module is stacked in the deep residual network model, and a soft threshold denoising module is added after the convolution operation of the efficient pyramid splitting attention module to obtain a deep residual shrinkage network model;

[0072] The transfer learning method is adopted, the deep residual network model is selected as the pre-trained model, and the pre-trained weights of the pre-trained model in the PyTorch image database are called for the deep residual shrinkage network model to obtain the deep residual shrinkage network model based on transfer learning.

[0073] Among them, the pre-trained weights of the pre-trained model in the PyTorch image database are retrieved, including:

[0074] Pre-trained model, including: 7×7 convolutional layer, maximum pooling layer, average pooling layer and fully connected layer;

[0075] Call the weights of the 7×7 convolutional layer of ResNet that have been trained on the PyTorch image database to initialize the 7×7 convolutional layer of the deep residual contraction network model;

[0076] Call the weights of the ResNet's maximum pooling layer that have been trained on the PyTorch image database to initialize the maximum pooling layer of the deep residual shrinkage network model;

[0077] Call the weights of the average pooling layer of ResNet that have been trained on the PyTorch image database to initialize the average pooling layer of the deep residual shrinkage network model;

[0078] Call the weights of the fully connected layer of ResNet that have been trained on the PyTorch image database to initialize the fully connected layer of the deep residual shrinkage network model.

[0079] S103: Using the Mel frequency cepstral coefficients dataset for training, a trained deep residual shrinkage network model based on transfer learning is obtained, which specifically includes:

[0080] The training set is divided into K subsets using the ten-fold cross validation method. K training and validation are performed. Each time, K-1 subsets are used as the second training set, and the other subset is used as the validation set. The accuracy is calculated.

[0081] When the accuracy reaches the specified iteration cycle and the accuracy no longer continues to rise, the training effect of the deep residual shrinkage network model based on transfer learning reaches the optimal effect and is saved;

[0082] During the training process of the deep residual shrinkage network model based on transfer learning, the Adam optimization algorithm and the cross entropy loss function are used to optimize the deep residual shrinkage network model based on transfer learning; and the automatically adjusted hyperparameters are set, and the learning rate is halved after every 10 training cycles.

[0083] The Adam optimization algorithm has the following formula:

[0084] ;

[0085] ;

[0086] In the formula, m t is the first-order moment estimate, v t is the second-order moment estimate, β 1 is the exponential decay rate of the first-order moment estimate, β 2 is the exponential decay rate of the second-order moment estimate, g i is the gradient;

[0087] The cross entropy loss function is formulated as:

[0088] ;

[0089] In the formula, is the cross entropy loss function, q is the number of categories, , is the one-hot encoding of the true label, The model predicts i Probability of class;

[0090] Among them, the weight parameters of the convolutional layer of the deep residual shrinkage network model based on transfer learning are not randomly initialized, but the pre-trained weights of the pre-trained model are called to fine-tune all layers;

[0091] In addition, the recognition effects of models with different soft threshold denoising module configurations on the training set are compared, including:

[0092] A soft threshold module is added after the convolution operation of the 4-layer efficient pyramid split attention module in the deep residual shrinkage network model based on transfer learning; the number of channels of each layer output is 256, 512, 1024 and 2048 respectively;

[0093] Configuration 1, including: adding a soft threshold denoising module after the convolution operation of all layers of the efficient pyramid split attention module, that is, a multi-layer soft threshold denoising module configuration;

[0094] Configuration 2, including: adding a soft threshold denoising module after the convolution operation of each layer of the efficient pyramid split attention module, that is, a single-layer soft threshold denoising configuration;

[0095] Compare the recognition effects of the deep residual shrinkage network model based on transfer learning on the training set under single-layer soft threshold denoising configuration and multi-layer soft threshold denoising configuration.

[0096] First, the unimproved residual network ResNet50 is compared to verify the effect of the selection of the basic network on the recognition of damage features. The effects of inserting the DRSN_CW module at each stage in S61 are compared, and the accuracy of adding single-layer and multi-layer soft threshold denoising algorithms is compared. The network model with the highest accuracy and the least loss function is selected as the best deep residual shrinkage network model based on transfer learning.

[0097] Among them, when evaluating the deep residual shrinkage network model, the evaluation indicators accuracy, loss function, precision, recall rate, F1-Score and T-NSE, and confusion matrix are used to comprehensively evaluate the performance of the model:

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] In the formula, ACC is the accuracy, P is the precision, R is the recall rate, F1 is the accuracy index of the binary classification model, TP is the true positive, FP is the false positive, FN is the false negative, and TN is the number of true negative cases.

[0103] S104: Use the trained deep residual shrinkage network model based on transfer learning to identify the damage type of the composite material to be identified.

[0104] This embodiment uses existing acoustic emission detection equipment to connect the composite material plate part to collect the acoustic emission data of the composite material stretching;

[0105] See also Figure 2, the acoustic emission data sets of three damage states of composite materials: fiber breakage, matrix cracking, and delamination are converted into Mel spectrum cepstrum coefficients to generate Mel spectrum cepstrum graphs. The generated graphs constitute 3000 sample images for intelligent network pattern recognition. According to the ratio of 8:2, the test set samples are used for training with 2400 images, 800 images for each category, and the test set has 600 images, 200 images for each category. The specific number is shown in the table below:

[0106]

[0107] Among them, the steps of generating MFCC spectrum are:

[0108] Pre-emphasis: Pre-emphasize the collected original acoustic emission signal to enhance the energy of the high-frequency part.

[0109] Framing: Divide the pre-emphasized signal into short time frames, usually using overlapping window technology;

[0110] Fast Fourier Transform (FFT): Apply FFT to each frame of data to convert the time domain signal to the frequency domain;

[0111] Mel filter bank design: Design a set of Mel filters that are evenly distributed on the Mel frequency axis to simulate the human ear's perception of different frequencies;

[0112] Mel filter bank application: The spectrum of each frame is passed through the Mel filter bank to obtain the Mel spectrum.

[0113] Logarithmic operation: Take the logarithm of the Mel spectrum to enhance the ability to distinguish low-energy components;

[0114] Discrete Cosine Transform (DCT): DCT is applied to the log-Mel spectrum of each frame to obtain Mel-frequency Cepstral Coefficients (MFCCs).

[0115] The generated acoustic emission Mel frequency spectrum cepstrum data set is preprocessed, and the preprocessing includes:

[0116] RandomResizedCrop(224): Randomly scale and crop the input image to a size of 224x224. This operation is used to generate image patches of different sizes to enhance the generalization ability of the model.

[0117] Random Horizontal Flip (RandomHorizontalFlip()): Randomly flip the image horizontally (left to right) to increase data diversity and reduce overfitting of the model.

[0118] Convert to Tensor format (ToTensor()): Convert an image in PIL image or NumPy array format to Tensor format in PyTorch, while automatically scaling pixel values ​​to the range of [0, 1].

[0119] Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])): Normalize the image using the given mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225] to ensure that each channel (RGB) has a relatively standard distribution during training.

[0120] Using the transfer learning method, call the pre-trained model ResNet50, and use the torchvision.models() method to call the pre-trained weights of the ResNet50 model, including:

[0121] The deep residual contraction network model uses the model weights of Convl7*7 in ResNet50 that have been trained on ImageNet;

[0122] The deep residual contraction network model uses the model weights of Max pooling in ResNet50 that have been trained on ImageNet;

[0123] The deep residual contraction network model uses the model weights of Average pooling in ResNet50 that have been trained on ImageNet;

[0124] The deep residual contraction network model calls the model weights of FC in ResNet50 that have been trained on ImageNet.

[0125] See also Figure 3 , replace the traditional 3×3 convolution in the residual block Bottleneck of the residual network ResNet50 neural network with the pyramid split attention mechanism module (PSA), build the EPSA module, and stack it in the way of the ResNet network to form a new network EPSANet50;

[0126] See also Figure 4 , after the convolution operation of Layer1-Layer2 in each stage of EPSANet50 network, the soft threshold denoising DRSN_CW module is added, and the number of output channels is 256, 512, 1024 and 2048, and the overall framework of the deep residual shrinkage network model based on transfer learning is generated to identify the damage category of composite materials. The specific steps are as follows:

[0127] After the convolution operation of each stage stage0-stage4 of ResNet50, soft threshold denoising module DRSN_CW is added, and the number of output channels is 256, 512, 1024 and 2048; soft threshold denoising module DRSN_CW, respectively, consists of 5 parts. Stage0-Stage4, where Stage0 performs image preprocessing; soft threshold denoising module DRSN_CW is inserted after residual blocks Stage1, Stage2, Stage3, and Stage4 respectively.

[0128] The training image of size (3,224,224) is input into stage0 and after preprocessing, including image resizing and normalization, the input image is prepared for network processing. After convolution and max pooling operations, the output image size is (64,112,112);

[0129] The image of size (64, 112, 112) in step S102 is input into stage1, and after passing through three Bottleneck modules, an image of size (256, 56, 56) is output;

[0130] The above image of size (256, 56, 56) is input into stage2, and after passing through 4 Bottleneck modules, the output image of size (512, 28, 28) is obtained.

[0131] The above image of size (512, 28, 28) is input into stage3, and after passing through 6 Bottleneck modules, the output image size is (1024, 14, 14);

[0132] The above image of size (1024, 14, 14) is input into stage4, and after passing through three Bottleneck modules, the output image is of size (2048, 7, 7);

[0133] The input image is converted into a 1*2048 vector output through the pooling operation of the neural network.

[0134] The following table shows the effect of inserting the soft threshold denoising DRSN_CW module in each stage of the deep residual shrinkage network model, and compares the accuracy of adding single-layer soft threshold denoising and multi-layer soft threshold denoising, and selects the network model with the highest accuracy and the least loss function. As shown in the table: Compared with the initial ResNet50 model, adding the PSA pyramid split attention mechanism and the DRSN_CW module improves the accuracy of the model to a certain extent. Especially when the DRSN_CW module is added after Layer2, the loss function is only 0.035, and the accuracy is as high as 97%, which is 7.3% higher than the accuracy of the ResNet50 model.

[0135]

[0136] For the accuracy and loss function curves of each model, see Figure 5 .

[0137] It is worth noting that when we add the DRSN_CW module after Stage1, Stage2, Stage3, and Stage4 of the model, the accuracy of the model is not as good as when it is added only in a single channel. Despite this, the accuracy of the model is still 4.8% higher than that of the base model. On the contrary, after inserting the SE mechanism after layer1 to layer4 of ResNet50, the accuracy of the model not only did not improve, but decreased by 2.0%. These phenomena show that simply stacking modules in the network is not the best strategy. On the contrary, the network level should be reasonably selected to integrate appropriate network modules according to the characteristics of the dataset and the actual application requirements. Therefore, the selected PRSN network model was finally determined to be the residual network Layer2 followed by the DRSN_CW module, and the network combined with the PSA pyramid attention mechanism and transfer learning was our final network.

[0138] The parameters of the convolutional layer are not randomly initialized, but use the weight parameters in the pre-trained ResNet50 model. In this transfer learning process, all weights except the residual block in the pre-trained model ResNet50 are called, and all layers are fine-tuned.

[0139] During the entire training process, the Adam optimization algorithm formula 1 and the cross entropy loss function formula 2 were used to optimize the model to improve the model fitting effect. Automatic adjustment hyperparameters were also set, and the learning rate was halved after every 10 training cycles to promote rapid convergence of the model.

[0140] See also Figure 6 The following table shows the comparison between fine-tuning only the last layer and fine-tuning all layers during transfer learning. Figure 6The results show that the TL_PRSN model performs best when all layers are fine-tuned, achieving an accuracy of 97.0% and a loss value as low as 0.035. All evaluation indicators (P, R and F1) reach 0.970. The model converges quickly and performs stably during training. In contrast, the performance of the model that only fine-tunes the last layer drops significantly, with an accuracy of only 64.2% and a loss value of 1.041, even lower than the performance of non-transfer learning, and shows large fluctuations during training. However, the fluctuations after 40 iterations are significantly smaller than those of the above-mentioned PRSN non-transfer learning, which also confirms the stability advantage of transfer learning. This result shows that fully fine-tuning all layers can give full play to the feature extraction capabilities of the pre-trained model and significantly improve the classification performance and stability of the model. This method well verifies the superiority of our proposed method.

[0141]

[0142] Gaussian white noise is added to the acoustic emission signal of the test set to simulate the noise interference in the real situation during the composite material collection process. The noisy acoustic emission signal is converted into a Mel-spectrogram cepstral coefficient map, and the test set samples preprocessed by the Mel-spectrogram cepstral map are input into the PRSN network model based on transfer learning for feature extraction, model training and verification recognition. The performance of the deep residual shrinkage network model of transfer learning is evaluated by accuracy.

[0143] Test set noise addition: Gaussian white noise of 20dB, 15dB, 10dB and 5dB was added to the original acoustic emission signal to simulate different levels of noise. These noise levels represent small, medium and large noise levels. The calculation formula of signal-to-noise ratio (SNR) is as follows:

[0144] ;

[0145] The following table shows the various accuracy rates after adding noise. As can be seen from the table, as the noise level increases, the accuracy rates of both models show a downward trend, but PRSN always performs better than the ResNet50 model in various noise environments, especially in a 5dB high noise environment, where the accuracy rate of PRSN is 61.5%, significantly higher than the 55.16% of ResNet50. On average, the accuracy rate of PRSN is 80%, while the average accuracy rate of ResNet50 is 73%.

[0146]

[0147] Use evaluation indicators such as accuracy, loss function, precision, recall, F1-Score, T-NSE, and confusion matrix to comprehensively evaluate the performance of the model;

[0148] In order to improve the interpretability of deep learning models, this paper introduces the Local Interpretable Model-Agnostic Explanations (LIME) function to analyze the recognition performance of the network. LIME generates different variants of the image by splitting the image into superpixel blocks and enabling or disabling these pixels. By comparing the closeness of these variants to the target instance, the contribution of each superpixel can be evaluated. In this paper, we generated 8,000 neighborhood images, and this increase in number helps to improve the accuracy of the classification area. LIME allows for direct explanations on image samples. See Figure 7 , showing the areas that the model identifies with the most attention in different damage types, which are highlighted with yellow borders. Within the yellow border, yellow and green indicate that the area has a positive impact on the current category, and the more colors, the higher the confidence. Pink and red indicate that the area has a negative impact on the current category. Figures a, e, and i show the spectra of randomly selected fiber breakage, matrix cracking, and delamination signals, respectively. Figures b, c, d, f, g, h, and j, k, and l show the contribution of the model in identifying these three signals as fiber breakage, matrix cracking, and delamination categories, respectively. Through these spectra, we can intuitively observe the model's contribution to the recognition of different types of damage signals, providing a basis for a deep understanding of the model's classification performance.

[0149] Figure 8 This is a diagram of the data visualized by using T-SNE for dimensionality reduction analysis. The Conv1, Maxpool, Avgpool, and Fc layers of the model are visualized. The model shows obvious feature differentiation in the initial stage, thanks to the use of the transfer learning strategy. Although there is still signal confusion in the Conv1 and Maxpool stages, by improving the residual block, the model successfully aggregates the data of the same damage type in the Avgpool and Fc stages, and clearly separates the data of different damage types.

[0150] The PRSN network model is used to perform damage pattern recognition on the Mel-frequency cepstrum of acoustic emission samples.

[0151] The computer device used in the present invention comprises a memory and a processor, wherein the memory contains a computer program, and the processor executes the program to implement a composite material damage identification method based on transfer learning and improved residual network.

[0152] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, 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, they should be considered to be within the scope of this specification.

Claims

1. A composite material damage identification method based on transfer learning and improved residual network, characterized in that: include: Acquire an acoustic emission data set of composite material stretching, and perform Mel-spectrum cepstral coefficient conversion on the acoustic emission data set to generate a Mel-spectrum cepstral coefficient spectrum data set; The acoustic emission data include: acoustic emission data of the composite material under three damage modes: fiber fracture, matrix cracking and delamination; Based on the ResNet model, a deep residual shrinkage network model is constructed. The pre-trained weights of the ResNet model trained on the graph network are called. The weights of the deep residual shrinkage network model are initialized using the weights other than the residual block in the pre-trained weights to obtain a deep residual shrinkage network model based on transfer learning. The construction of the deep residual shrinkage network model includes: using multiple PSA blocks to replace all convolution blocks in the residual block of the ResNet model to obtain an EPSANet model; adding a soft threshold denoising module after the convolution operation of the second layer of PSA blocks in the EPSANet model to obtain a deep residual shrinkage network model; The method of calling the pre-trained weights trained on the graph network of the ResNet model and using the other weights in the pre-trained weights except the residual block to initialize the weights of the deep residual shrinkage network model includes: calling the weights of the 7×7 convolutional layer of the ResNet model that have been trained on the PyTorch image database; calling the weights of the maximum pooling layer of the ResNet model that have been trained on the PyTorch image database to initialize the maximum pooling layer of the deep residual shrinkage network model; calling the weights of the average pooling layer of the ResNet model that have been trained on the PyTorch image database to initialize the average pooling layer of the deep residual shrinkage network model; calling the weights of the fully connected layer of the ResNet model that have been trained on the PyTorch image database to initialize the fully connected layer of the deep residual shrinkage network model; Training a deep residual shrinkage network model based on transfer learning, including: using a ten-fold cross-validation method, dividing the Mel-spectrum cepstral coefficient map data set into K subsets, performing K training and validation, each time using K-1 subsets as training sets and another subset as validation sets, and calculating the accuracy; when the accuracy reaches a specified iteration cycle and the accuracy no longer continues to rise, the training effect of the deep residual shrinkage network model based on transfer learning reaches the optimal value, and the trained deep residual shrinkage network model based on transfer learning is saved; Among them, in the training process of the deep residual shrinkage network model based on transfer learning, the Adam optimization algorithm and the cross entropy loss function are used to optimize the deep residual shrinkage network model based on transfer learning; and the automatically adjusted hyperparameters are set, and the learning rate is halved after every 10 training cycles; The formula of the Adam optimization algorithm is: ； ； In the formula, m t is the first-order moment estimate, v t is the second-order moment estimate, β 1 is the exponential decay rate of the first-order moment estimate, β 2 is the exponential decay rate of the second-order moment estimate, , g i is the gradient; The formula of the cross entropy loss function is: ； In the formula, is the cross entropy loss function, q is the number of categories, , is the one-hot encoding of the true label, The model predicts i Probability of class; The trained deep residual shrinkage network model based on transfer learning is used to identify the damage type of the composite material to be identified.

2. The composite material damage identification method based on transfer learning and improved residual network according to claim 1 is characterized in that: Also includes: The Mel frequency cepstral coefficient spectrum data set is preprocessed, which specifically includes: Scale and crop the image of the input Mel frequency cepstral coefficient map dataset to a specified size to generate image blocks of different sizes; Flip the image horizontally; Convert images in PIL image format or array format to PyTorch Tensor format and automatically scale pixel values ​​to the range of [0, 1]; Normalize the image using a preset mean and standard deviation.

3. The composite material damage identification method based on transfer learning and improved residual network according to claim 1, characterized in that: The stacking method between the multiple PSA blocks is the stacking method of residual blocks in a deep residual network.

Images