An ultrasonic guided wave damage detection and in-situ calibration method

Abstract

The application discloses an ultrasonic guided wave damage detection and in-situ calibration method, multiple piezoelectric ceramic sensors are fixed on the surface of a carbon fiber reinforced composite material plate in the form of a parallel array, ultrasonic guided wave signals in a transmission path under multiple groups of damage settings are collected to obtain an ultrasonic guided wave data set containing damage in different regions; a dense convolution sparse coding network model is constructed and trained, test set data is evaluated, damage positioning results are obtained, and a positioning accuracy rate is calculated according to damage position classification results; according to the damage positioning results, sensors not involved in the damage region are determined, time delays of multiple groups of guided wave signals are calculated by using cross-correlation, and multi-path time-of-flight consistency in-situ calibration is carried out. By constructing the dense convolution sparse coding network, the carbon fiber reinforced composite material laminate damage detection and positioning based on the ultrasonic guided wave can be realized, meanwhile, the multi-path time-of-flight consistency in-situ calibration is realized, and the overall efficiency of detection and calibration is improved.

Description

Technical Field

[0001] This invention belongs to the field of nondestructive testing and calibration, and more specifically, it relates to a method for ultrasonic guided wave damage detection and in-situ calibration based on deep learning. Background Technology

[0002] Ultrasonic guided wave testing, due to its sensitivity to internal and external damage and wide detection range, has been increasingly applied in the non-destructive testing and monitoring of plate and shell structures in aerospace, petrochemical, and other fields. By arranging a sensor array on the surface of the target structure, the ultrasonic guided wave testing system can sequentially excite the sensors and collect multi-path guided wave signal data, which contains damage-related information. Through damage feature extraction and analysis of the signal data, damage can be identified and located, thereby quantifying the current state of the structure, assessing its future health status, and ensuring the safe operation of equipment. However, the dispersion and multimodal characteristics of ultrasonic guided waves, while enriching the signal information, also increase the difficulty of signal processing, affecting damage location and imaging results. In recent years, deep learning methods have attracted widespread attention in the fields of non-destructive testing and structural health monitoring due to their powerful feature extraction and representation capabilities. Using ultrasonic guided wave signals as input, a deep learning model is trained to extract key damage features, thereby predicting the location and type of damage.

[0003] In practical applications, besides the inherent characteristics of the signal itself, the impact of long-term sensor use on detection accuracy must also be considered. Sensors are affected by factors such as the surrounding environment and the aging of their own hardware, which can cause changes in their output signal, affecting the accuracy of damage detection and consequently compromising the safe and reliable operation of the monitored target. Therefore, regular sensor calibration is of great importance. Although conventional laboratory calibration methods are highly accurate and widely applicable, their long cycles and high costs prevent continuous monitoring of field equipment, resulting in economic losses. Therefore, it is necessary to research a method that can achieve in-situ calibration under specific conditions to solve the problem of regular sensor calibration. However, when the damage condition of the structure under test is unknown, the reliability of in-situ calibration results is also affected.

[0004] Therefore, in response to the need for structural damage detection and in-situ calibration, this invention proposes an ultrasonic guided wave damage detection and in-situ calibration method based on dense convolutional sparse coding networks, which can achieve in-situ calibration while detecting structural damage. Summary of the Invention

[0005] The purpose of this invention is to provide an ultrasonic guided wave damage detection and in-situ calibration method that achieves in-situ calibration while detecting structural damage.

[0006] The technical solution of this invention: A method for ultrasonic guided wave damage detection and in-situ calibration, comprising the following steps:

[0007] S1: Multiple piezoelectric ceramic sensors are fixed on the surface of a carbon fiber reinforced composite material plate in a parallel array. Aluminum blocks are used to simulate the delamination damage of the composite material. Under the damage conditions at different locations, each sensor emits an excitation waveform in sequence and collects the transmission path signal to form an ultrasonic guided wave dataset containing damage in different regions. The dataset is divided into two parts: a training set and a test set.

[0008] S2: Construct and train a dense convolutional sparse coding network model; within the framework of a multi-layer convolutional sparse coding network model, dense connections are introduced to directly connect the feature maps of different layers. A narrowband comb signal is used as the kernel of the first convolutional layer to learn the ultrasonic guided wave signal features that match the guided wave packet waveform. A channel attention module is introduced to further enhance the model's representation ability; Step S2 includes the following steps:

[0009] S2-1. Introduce dense connection theory into the multi-layer convolutional sparse coding network model framework; construct a multi-layer convolutional sparse coding network framework, introduce dense connections by adding an identity matrix to the dictionary, and use an iterative soft thresholding algorithm with dense connections to update the forward propagation process.

[0010] S2-2. Design of convolution kernels considering guided wave characteristics; use a narrowband comb signal matching the experimental excitation frequency as the convolution kernel of the first layer of the network, and match the waveform of the ultrasonic guided wave signal to extract signal features;

[0011] S2-3. Introduce the channel attention mechanism into network feature extraction; the channel attention module includes a global average pooling layer and a fully connected layer. It processes the input feature mapping through the global average pooling operation and compresses the global spatial information into a single channel representation.

[0012] S2-4. Input the training set from step S1 into the dense convolutional sparse coding network model for training;

[0013] S3: Input the test set from S1 into the network model built in S2, evaluate the test set data, obtain the damage localization results, and get the corresponding accuracy.

[0014] S4: Based on the damage localization results in S3, identify the sensors that are not involved in the damage area, calculate the time delay of multiple sets of guided wave signals using cross-correlation, and perform in-situ calibration of multipath time-of-flight consistency.

[0015] In the aforementioned ultrasonic guided wave damage detection and in-situ calibration method, step S2 specifically includes the following steps:

[0016] S2-1. Introducing dense connection theory into the multi-layer convolutional sparse coding network model framework; using the multi-layer convolutional sparse coding network framework, given the acquired ultrasonic guided wave signal, the multi-layer convolutional sparse coding model can be expressed as:

[0017]

[0018] Wherein, the input y of the model is the obtained ultrasonic guided wave signal, and x L D represents the deepest feature map. i It is the dictionary of the i-th layer. This multi-layer expansion enables the model to learn features at multiple levels of abstraction. The forward propagation process of the neural network involves learning these intermediate features x. i The estimate;

[0019] Dense connections are introduced into the framework of multi-layer convolutional sparse coding networks. Dense connections collect all feature maps before the current layer and compute the feature maps of the next layer together with them. The dense connections in the network structure can be represented as follows:

[0020] x i+1 =H(F i+1 ({x0,x1,...,x i})+b i+1 (2)

[0021] Where H(·) is a nonlinear transformation, F i+1 Let x represent the convolution matrix. i Let {x0, x1, ..., x} be the feature map of the i-th layer. i} represents the concatenation of feature maps from each layer, b i+1 For bias terms;

[0022] Dense joins are achieved by adding an identity matrix I to a dictionary, denoted as D. i =[I,(F i ) T The forward propagation process of a network can be represented as follows:

[0023]

[0024] Among them, T λ This is a soft thresholding operator; it connects feature maps from different layers to enhance feature propagation. After replacing the dictionary, the iterative soft thresholding algorithm with dense connections can be used to update the threshold, which can be expressed as follows:

[0025]

[0026] Where c is a learnable constant, x k Let x represent the feature map of the k-th iteration. k+1 This represents the feature map of the (k+1)th iteration;

[0027] S2-2. Design of convolutional kernels considering guided wave characteristics; a narrowband comb signal matching the experimental excitation frequency is used as the convolutional kernel for the first layer of the network, matching the ultrasonic guided wave signal waveform to extract signal features. The waveform of the first-layer convolutional kernel can be expressed as follows:

[0028]

[0029] Among them, f c The center frequency is n, and the number of cycles is n.

[0030] S2-3. Introduce the channel attention mechanism into network feature extraction; given the input feature mapping as... Where C represents the number of channels and W represents the length of the feature map, global average pooling is used to compress global spatial information into a single channel representation, thereby generating cross-channel statistical information by leveraging the dependencies between channels. The representation of channel c by z c It can be calculated as

[0031]

[0032] In the formula, x i,c This represents the feature map of the c-th channel. This is a global average pooling operation; the channel attention module retains only one fully connected layer to avoid changes in channel dimensions and improve the extraction quality of cross-channel information, where channel attention A... eSE Calculated using the following formula

[0033] A eSE =σ(W3z) (7)

[0034] In the formula The weights of the fully connected layer; Applied to feature mapping to obtain more efficient information, new feature maps are obtained based on channel multiplication.

[0035]

[0036] in, It is a new feature mapping for the c-th channel. It is the final output of the channel attention module;

[0037] S2-4. Input the training dataset from step S1 into the dense convolutional sparse coding network model for training.

[0038] In the aforementioned ultrasonic guided wave damage detection and in-situ calibration method, the dense convolutional sparse coding network model structure in step S2 includes a convolutional layer with a narrowband comb signal convolution kernel, a max pooling layer, a batch normalization layer, a multi-layer convolutional sparse coding module with embedded dense connections, a channel attention module, a fully connected layer 1, a dropout layer, a fully connected layer 2, and a softmax layer output.

[0039] In the aforementioned ultrasonic guided wave damage detection and in-situ calibration method, in step S2, the parameters of the first layer convolution kernel are set as follows: excitation frequency f c The frequencies were set to 40kHz, 60kHz, and 80kHz, with n being 5, the number of convolutional kernels being 16, and the kernel size being 1×1700.

[0040] In the aforementioned ultrasonic guided wave damage detection and in-situ calibration method, in step S2, the embedded densely connected multilayer convolutional sparse coding module comprises three layers, all of which have a kernel size of 1×3, and the number of kernels in these three layers is set to 32.

[0041] In the aforementioned ultrasonic guided wave damage detection and in-situ calibration method, in step S2, the parameters of the dense convolutional sparse coding network model are set as follows: dropout ratio is set to 0.25, training is performed using the Adam optimizer, the sample batch size is 32, and the initial learning rate is set to 10. -5 After 150 training iterations, the learning rate was reduced by multiplying by 0.75, and cross-entropy was used as the loss function for classification training.

[0042] The beneficial effects of the present invention are as follows: Compared with the prior art, the present invention can realize the detection and localization of damage in carbon fiber reinforced composite laminates based on ultrasonic guided waves by constructing a dense convolutional sparse coding network, and at the same time realize the consistent in-situ calibration of multi-path time of flight, thereby improving the overall efficiency of detection and calibration.

[0043] The present invention has the following advantages:

[0044] (1) The present invention can realize the damage detection and localization of carbon fiber reinforced composite laminate based on ultrasonic guided waves.

[0045] (2) The dense convolutional sparse coding network proposed in this invention can achieve efficient extraction of damage-related features in ultrasonic guided wave signals and complete automated structural damage identification and classification.

[0046] (3) The present invention can achieve multi-path consistency in-situ calibration under the same framework, which improves the overall efficiency of detection and calibration. Attached Figure Description

[0047] Appendix Figure 1This is a flowchart of the ultrasonic guided wave damage localization and in-situ calibration method based on dense convolutional sparse coding network of the present invention.

[0048] Appendix Figure 2 Experimental diagram of carbon fiber reinforced composite laminate and damage arrangement;

[0049] Appendix Figure 3 A schematic diagram showing the sensor layout and location of damage;

[0050] Appendix Figure 4 Output the confusion matrix result image for the network;

[0051] Appendix Figure 5 Image showing the damage localization results of the composite material;

[0052] Appendix Figure 6 The image shows the guided wave signals along five paths in the undamaged region.

[0053] Appendix Figure 7 The envelope diagrams are shown for the five guided wave signals from the undamaged region. Detailed Implementation

[0054] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. The present application will now be described in detail with reference to the accompanying drawings and embodiments.

[0055] Figure 1 This invention illustrates an ultrasonic guided wave damage detection and in-situ calibration method based on a dense convolutional sparse coding network, comprising the following steps:

[0056] S1: Multiple piezoelectric ceramic sensors (PZTs) are fixed on the surface of a carbon fiber reinforced composite material plate in a parallel array. Aluminum blocks are used to simulate delamination damage of the composite material. Under damage conditions at different locations, each sensor emits an excitation waveform in sequence and collects transmission path signals to form an ultrasonic guided wave dataset containing damage in different regions. The dataset is divided into two parts: a training set and a test set.

[0057] In this embodiment, the carbon fiber reinforced composite material plate has dimensions of 500mm × 500mm × 2mm, and the laminate layup is [45 / 0 / -45 / 90 / -45 / 0 / 45 / 90]. s It is quasi-isotropic. Figure 2 The diagram shows an experimental setup of carbon fiber reinforced composite laminates and damage distribution. Figure 3 This is a schematic diagram showing the sensor arrangement and damage location. 12 PZT sensors (T1-T) 12The components were fixed to the structure in a parallel array. An aluminum block with a diameter of 10 mm and a thickness of 3 mm was bonded to the surface of the composite material structure. Different material properties and interface effects were introduced to simulate delamination damage D1-D at 28 different locations. 28 These damages were categorized into seven damage regions, R1-R7, and localized using a network classification system. A five-cycle narrowband comb signal was used as the excitation, with center frequencies set to 40kHz, 60kHz, and 80kHz, respectively. Each PZT sensor sequentially emitted the excitation waveform, acquiring multiple sets of transmission path signals. In each experiment, a simulated damage was set on the structure, and guided wave signals were acquired by changing the damage location. After preprocessing, a dataset for ultrasonic guided wave damage localization was constructed. The dataset in this embodiment consists of 9072 samples, of which 135 samples were randomly selected from each class to form 945 samples as the test dataset.

[0058] S2: Construct and train a dense convolutional sparse coding network model. Specifically, within the framework of a multi-layer convolutional sparse coding network model, dense connections are introduced to directly connect the feature maps of different layers. A narrowband comb signal is used as the kernel of the first convolutional layer to learn the ultrasonic guided wave signal features that match the guided wave packet waveform. A channel attention module is introduced to further enhance the model's representation ability.

[0059] (I) Introducing dense connection theory into the multi-layer convolutional sparse coding network model framework. Using a multi-layer convolutional sparse coding network framework, given the acquired ultrasonic guided wave signal, the multi-layer convolutional sparse coding model can be expressed as follows:

[0060]

[0061] Wherein, the input y of the model is the obtained ultrasonic guided wave signal, and x L D represents the deepest feature map. i It is the dictionary of the i-th layer. This multi-layer expansion enables the model to learn features at multiple levels of abstraction. The forward propagation process of the neural network involves learning these intermediate features x. i The estimate.

[0062] Dense connections are introduced into the framework of multi-layer convolutional sparse coding networks. Dense connections collect all feature maps before the current layer and compute the feature maps of the next layer together with them. The dense connections in the network structure can be represented as follows:

[0063] x i+1 =H(F i+1 ({x0,x1,...,x i})+b i+1 (2)

[0064] Where H(·) is a nonlinear transformation, which in this embodiment is the ReLU function, and F i+1 Let x represent the convolution matrix. i Let {x0, x1, ..., x} be the feature map of the i-th layer. i} represents the concatenation of feature maps from each layer, b i+1 This is the bias term. In the mathematical derivation, dense connections are achieved by adding an identity matrix I to the dictionary, denoted as D. i =[I,(F i ) T The forward propagation process of a network can be represented as follows:

[0065]

[0066] Among them, T λ This is a soft thresholding operator. In this case, feature maps from different layers are concatenated, which promotes feature reuse throughout the network and strengthens feature propagation. After replacing the dictionary, updating with an iterative soft thresholding algorithm that imparts dense connections can be represented as:

[0067]

[0068] Here, c is a learnable constant.

[0069] (II) Design of Convolutional Kernels Considering Guided Wave Characteristics. Since the experimental excitation consists of narrowband comb signals at various center frequencies, the signal response acquired by the receiving device can be considered as a time-shifted and phase-transitioned narrowband comb signal with amplitude attenuation. Therefore, using a narrowband comb signal matching the experimental excitation frequency as the convolutional kernel for the first layer of the network can effectively match the ultrasonic guided wave signal waveform and extract signal features. The waveform of the first-layer convolutional kernel can be represented as follows:

[0070]

[0071] Among them, f c Let f be the center frequency and n be the number of periods. In this embodiment, f c The frequencies were set to 40kHz, 60kHz, and 80kHz, with n being 5. The number of convolutional kernels was set to 16, and the kernel size was set to 1×1700.

[0072] (III) Introducing channel attention mechanism into network feature extraction to fully extract damage information. Given input features mapped as... Where C represents the number of channels and W represents the length of the feature map, global average pooling is used to compress global spatial information into a single channel representation, thereby generating cross-channel statistical information by leveraging the dependencies between channels. The representation of channel c by z c It can be calculated as

[0073]

[0074] In the formula, x i,c This represents the feature map of the c-th channel. This is a global average pooling operation. The channel attention module of this invention retains only one fully connected layer to avoid changes in channel dimensions and improve the extraction quality of cross-channel information. The channel attention A... eSE Calculated using the following formula

[0075] A eSE =σ(W3z) (7)

[0076] In the formula These are the weights of the fully connected layer. Applied to feature mapping to obtain more efficient information, new feature maps are obtained based on channel multiplication.

[0077]

[0078] in, It is a new feature mapping for the c-th channel. It is the final output of the channel attention module.

[0079] (iv) Input the training dataset from step S1 into the dense convolutional sparse coding network model for training.

[0080] In this embodiment, the dense convolutional sparse coding network model structure includes convolutional layers with narrowband comb-shaped signal convolutional kernels, max pooling layers, batch normalization, a multi-layer convolutional sparse coding module with embedded dense connections, a channel attention module, a fully connected layer 1, a dropout layer, a fully connected layer 2, and a softmax layer for output. Each input sample has a size of 1×5100. The multi-layer convolutional sparse coding module with embedded dense connections consists of three layers, all with a kernel size of 1×3, and the number of kernels in each of these three layers is set to 32. Furthermore, the dropout ratio is set to 0.25. Training is performed using the Adam optimizer with a batch size of 32 samples. The initial learning rate is set to 10. -5 After 150 training iterations, the learning rate was reduced by multiplying by 0.75, and cross-entropy was used as the loss function for classification training.

[0081] S3: Input the test set from S1 into the network model constructed in S2, evaluate the test set data, obtain the damage localization results, and obtain the corresponding accuracy.

[0082] In this embodiment, the testing process involves classifying the damaged regions of the samples in the test set. The final confusion matrix and damage localization results are as follows: Figure 4 and Figure 5 As shown in the diagram, the samples in the diagonal region of the confusion matrix are the correctly classified samples. The test accuracy is the ratio of the number of correctly predicted samples to the total number of samples. The test precision is calculated from the prediction results of the test dataset, and the final test accuracy reached 98.94%.

[0083] S4: Based on the damage localization results in S3, identify the sensors that are not involved in the damage area, calculate the time delay through cross-correlation, and perform multipath consistency in-situ calibration.

[0084] In this embodiment, sensor array signal data with an excitation signal center frequency of 40kHz and a D1 damage setting are selected for analysis. After damage localization using a dense convolutional sparse coding network, the damaged region is determined to be R1, and the undamaged regions are determined to be T1-T5 and T7-T1. 11 The monitoring area is enclosed by sensors. Therefore, the paths T1-T7, T2-T8, T3-T9, and T4-T are selected. 10 T5-T 11 As a consistency calibration path, since these five paths propagate in the same direction, they have the same group velocity and the same distance. Therefore, the flight time of the guided wave direct-to-wave packet in the acquired signal should theoretically be consistent. The signals obtained from these five paths and their corresponding envelopes are as follows: Figure 6 As shown, the time delays of the five signal groups were calculated using cross-correlation. Taking paths T1-T7 as the baseline, the relative time delays for the other paths were -7μs, 6.8μs, 7.3μs, and -7.8μs, respectively. These time delay data can be used to adjust the sensor data and achieve in-situ calibration for multi-path time-of-flight consistency, thereby ensuring the accuracy and reliability of damage localization based on ultrasonic guided waves.

[0085] The foregoing description provides a detailed account of embodiments of the present invention and is not intended to limit the invention in any way. Those skilled in the art can make various optimizations, improvements, and modifications based on this invention. Therefore, the scope of protection of this invention should be defined by the appended claims.

Claims

1. A method for ultrasonic guided wave damage detection and in-situ calibration, characterized in that: It includes the following steps: S1: Multiple piezoelectric ceramic sensors are fixed on the surface of a carbon fiber reinforced composite material plate in a parallel array. Aluminum blocks are used to simulate the delamination damage of the composite material. Under the damage conditions at different locations, each sensor emits an excitation waveform in sequence and collects the transmission path signal to form an ultrasonic guided wave dataset containing damage in different regions. The dataset is divided into two parts: a training set and a test set. S2: Construct and train a dense convolutional sparse coding network model: Under the framework of a multi-layer convolutional sparse coding network model, dense connections are introduced to directly connect the feature maps of different layers. A narrowband comb signal is used as the kernel of the first convolutional layer to learn the ultrasonic guided wave signal features that match the guided wave packet waveform. A channel attention module is introduced to further improve the model's representation ability. Step S2 includes the following steps: S2-1. Introduce dense connection theory into the multi-layer convolutional sparse coding network model framework; construct a multi-layer convolutional sparse coding network framework, introduce dense connections by adding an identity matrix to the dictionary, and use an iterative soft thresholding algorithm with dense connections to update the forward propagation process. S2-2. Design of convolution kernels considering guided wave characteristics; use a narrowband comb signal matching the experimental excitation frequency as the convolution kernel of the first layer of the network, and match the waveform of the ultrasonic guided wave signal to extract signal features; S2-3. Introduce the channel attention mechanism into network feature extraction; the channel attention module includes a global average pooling layer and a fully connected layer. It processes the input feature mapping through the global average pooling operation and compresses the global spatial information into a single channel representation. S2-4. Input the training set from step S1 into the dense convolutional sparse coding network model for training; S3: Input the test set from S1 into the network model built in S2, evaluate the test set data, obtain the damage localization results, and get the corresponding accuracy. S4: Based on the damage localization results in S3, identify the sensors that are not involved in the damage area, calculate the time delay of multiple sets of guided wave signals using cross-correlation, and perform in-situ calibration of multipath time-of-flight consistency.

2. The ultrasonic guided wave damage detection and in-situ calibration method according to claim 1, characterized in that: Step S2 specifically includes the following steps: S2-1. Introducing dense connection theory into the multi-layer convolutional sparse coding network model framework; using the multi-layer convolutional sparse coding network framework, given the acquired ultrasonic guided wave signal, the multi-layer convolutional sparse coding model can be expressed as: Wherein, the input y of the model is the obtained ultrasonic guided wave signal, and x L D represents the deepest feature map. i It is the dictionary of the i-th layer. This multi-layer expansion enables the model to learn features at multiple levels of abstraction. The forward propagation process of the neural network involves learning these intermediate features x. i The estimate; Dense connections are introduced into the framework of multi-layer convolutional sparse coding networks. Dense connections collect all feature maps before the current layer and compute the feature maps of the next layer together with them. The dense connections in the network structure can be represented as follows: x i+1 =H(F i+1 ({x0,x1,...,x i })+b i+1 ) (2) Where H(·) is a nonlinear transformation, F i+1 Let x represent the convolution matrix. i Let {x0, x1, ..., x} be the feature map of the i-th layer. i } represents the concatenation of feature maps from each layer, b i+1 For bias terms; Dense joins are achieved by adding an identity matrix I to a dictionary, denoted as D. i =[I,(F i ) T The forward propagation process of a network can be represented as follows: Among them, T λ This is a soft thresholding operator; it connects feature maps from different layers to enhance feature propagation. After replacing the dictionary, the iterative soft thresholding algorithm with dense connections can be used to update the threshold, which can be expressed as follows: Where c is a learnable constant, x k Let x represent the feature map of the k-th iteration. k+1 This represents the feature map of the (k+1)th iteration; S2-2. Design of convolutional kernels considering guided wave characteristics; a narrowband comb signal matching the experimental excitation frequency is used as the convolutional kernel for the first layer of the network, matching the ultrasonic guided wave signal waveform to extract signal features. The waveform of the first-layer convolutional kernel can be expressed as follows: Among them, f c The center frequency is n, and the number of cycles is n. S2-3. Introduce the channel attention mechanism into network feature extraction; given the input feature mapping as... Where C represents the number of channels and W represents the length of the feature map, global average pooling is used to compress global spatial information into a single channel representation, thereby generating cross-channel statistical information by leveraging the dependencies between channels. The representation of channel c is z c It can be calculated as In the formula, x i,c This represents the feature map of the c-th channel. This is a global average pooling operation; the channel attention module retains only one fully connected layer to avoid changes in channel dimensions and improve the extraction quality of cross-channel information, where channel attention A... eSE Calculated using the following formula A eSE =σ(W3z) (7) In the formula The weights of the fully connected layer; Applied to feature mapping to obtain more efficient information, new feature maps are obtained based on channel multiplication. in, It is a new feature mapping for the c-th channel. It is the final output of the channel attention module; S2-4. Input the training dataset from step S1 into the dense convolutional sparse coding network model for training.

3. The ultrasonic guided wave damage detection and in-situ calibration method according to claim 1, characterized in that: The dense convolutional sparse coding network model structure in step S2 includes a convolutional layer with a narrowband comb signal convolution kernel, a max pooling layer, a batch normalization layer, a multi-layer convolutional sparse coding module with embedded dense connections, a channel attention module, a fully connected layer 1, a dropout layer, a fully connected layer 2, and a softmax layer output.

4. The ultrasonic guided wave damage detection and in-situ calibration method according to claim 2, characterized in that: In step S2, the parameters of the first convolutional kernel are set as follows: activation frequency f c The frequencies were set to 40kHz, 60kHz, and 80kHz, with n being 5, the number of convolutional kernels being 16, and the kernel size being 1×1700.

5. The ultrasonic guided wave damage detection and in-situ calibration method according to claim 2, characterized in that: In step S2, the embedded densely connected multilayer convolutional sparse coding module contains three layers, all of which have a kernel size of 1×3, and the number of kernels in these three layers is set to 32.

6. The ultrasonic guided wave damage detection and in-situ calibration method according to claim 2, characterized in that: In step S2, the parameters of the dense convolutional sparse coding network model are set as follows: dropout ratio is set to 0.25, training is performed using the Adam optimizer, the batch size is 32, and the initial learning rate is set to 10. -5 After 150 training iterations, the learning rate was reduced by multiplying by 0.75, and cross-entropy was used as the loss function for classification training.

Images