Guided wave-based aviation composite material baseline-free multi-mode damage detection method
By adopting a wave-free baseline multimodal damage detection method in the structural health monitoring of aviation composite materials, the Mamba technology and convolutional neural network directly extract features from the guided wave signal, solving the accuracy in the prior art and poor robustness, high-precision damage detection and reduce costs.
Patent Information
- Application Number
- CN202510074227.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-17
- Publication Date
- 2025-02-21
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing damage detection models have poor accuracy, reliability and robustness in structural health monitoring of aerospace composites and rely on dense sensor arrays and baseline signals, limiting their application in baseline-free environments.
Using a wave-guided, a baseline multimodal damage detection method is used to design a multimodal damage detection model through the Mamba technology and convolutional neural network of the state space model, and the time and spectrum characteristics are directly extracted from the guided signal, and through crossover Entropy loss function performs classification tasks.
High-precision damage detection in baseline environments is realized, which reduces the dependence and cost of sensor arrays and enhances the application potential of damage detection in complex environments.
Smart Images
Figure CN119492802A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of structural health monitoring of aviation composite materials, and in particular to a baseline-free multi-modal damage detection method for aviation composite materials based on guided waves. Background Art
[0002] Compared with metal materials, carbon fiber reinforced polymer composites have significant advantages in light weight, high specific strength, high specific stiffness and corrosion resistance, so they are widely used in aerospace and various industrial fields. However, due to time-varying loads, low-speed impacts and harsh working environments, carbon fiber reinforced polymer components are prone to delamination, debonding and other almost invisible damage, which seriously threatens the safety of equipment and personnel. In order to detect structural damage in time, shorten downtime, reduce maintenance costs and prevent catastrophic failures, it is urgent to develop reliable and accurate structural health monitoring and damage detection technologies for aviation composites.
[0003] Among various structural health monitoring methods, guided wave based structural health monitoring technology has been widely used in composite structural damage detection due to its long propagation distance, low attenuation, low cost and high sensitivity to various damages. Traditional physical model-based analysis methods have limitations in quantitatively evaluating the physical behavior of guided waves. They rely on dense sensor arrays, and the detection accuracy is affected by the number and distribution of sensors. The guided wave structural health monitoring method based on deep learning effectively reduces the subjectivity of artificial signal interpretation and significantly improves the reliability and accuracy of the structural health monitoring system. Although the existing deep learning methods have improved the efficiency of damage feature extraction, the current technology only focuses on single-modal signals and ignores the key role of guided wave multimodal signals in damage detection. At the same time, the existing methods' reliance on baseline signals (from healthy or original structural signals) leads to deviations and misdiagnosis in damage identification, limiting their engineering applications in baseline-free environments. Summary of the invention
[0004] The technical problem solved by the present invention is to provide a baseline-free multimodal damage detection method for aviation composite materials based on guided waves, which can solve the problems of poor accuracy, reliability and robustness of existing damage detection models, while reducing dependence on and cost of sensor arrays and enhancing the application potential of damage detection in complex environments.
[0005] The basic solution provided by the present invention is a baseline-free multimodal damage detection method for aviation composite materials based on guided waves, comprising the following steps: Step 1: collect guided wave signals of aviation composite materials with several types of damage, and divide the collected guided wave signals into a training set, a verification set, and a test set; Step 2: Based on the Mamba technique of the state space model and the convolutional neural network, a multi-modal damage detection model MMDD, a GW branch and a CWT branch of the MMDD are designed respectively. The GW branch extracts time features from the guided wavelength sequence data through the Mamba technique based on the state space model, and the CWT branch extracts spectral features from the spectrum diagram of the continuous wavelet transform through the convolutional neural network. Step 3: The features extracted by the GW branch and the CWT branch are fused to obtain multi-dimensional damage features, and then mapped to a new space for classification; Step 4: The cross entropy loss function is used as the objective function of the multimodal damage detection model to perform the classification task by measuring the difference between the predicted score and the true probability that the sample belongs to a certain category; Step 5: Use the validation set to adjust the optimal hyperparameters and network structure to find the best model, and use the test set to verify the damage detection effect of the model and output the performance indicators; The step five comprises the following steps: S5.1: Use the loss function output by the validation set during the training process to evaluate the model performance, continuously adjust and test different hyperparameter combinations and network structures, and select the model configuration with the best performance on the validation set as the best model. The optimal configuration is: 400 iterations, 0.0001 learning rate, 16 batch size, and Adam optimizer. S5.2: Calculate the detection performance index precision (P) of the multimodal damage detection model:
[0006] in, represents the number of states correctly identified, Represents the number of instances that were misidentified.
[0007] Furthermore, in step 1, the data proportions of the training set, validation set, and test set are 80%, 10%, and 10%, respectively, and each data set is independent of each other.
[0008] Further, the step 2 comprises the following steps: Step 2.1: The GW branch extracts time features from the guided wavelength sequence data through an average pooling layer, a reshaping layer, a Mamba layer, a reshaping layer, and a fully connected layer in sequence; Step 2.2: The CWT branch extracts spectral features from the spectrum of the continuous wavelet transform through an average pooling layer, a continuous wavelet transform layer, an average pooling layer, two convolutional layers, and a reshaping layer in sequence; the Mamba technique is based on a state space model, and the output of the state model at multiple time points is derived as follows:
[0009] in, express The current state at the moment, represents the input control quantity at time t, represents the system output at time t, represents the state transfer matrix, Indicates the influence of the control quantity on the state quantity, Indicates the impact of the current state quantity on the system output; exist The state equation at this time is shown as follows:
[0010] The convolution transport of the convolutional network is shown in the following formula: .
[0011] Further, the step 2.1 includes the following steps: S2.1.1: The GW branch uses the waveguide signal in step 1 as input data: ,in represents the set of real numbers, represents the batch size, Indicates the sequence length of the GW signal; S2.1.2: Perform average pooling on the data using the following formula:
[0012] Get dimensionality reduction features ; S2.1.3: Sequence After grouping, it is expressed as ; S2.1.4: Input the grouped sequence data into the Mamba layer for feature extraction in, represents the Mamba model based on the state space model; S2.1.5: For serialized features Reshape , and linear processing After that, the final GW features are obtained. .
[0013] Further, the step 2.2 comprises the following steps: S2.2.1: The CWT branch takes the waveguide signal in step 1 as input data: ,in represents the batch size, Indicates the sequence length of the GW signal; S2.2.2: Perform average pooling on the data using the following formula:
[0014] Get dimensionality reduction features ; S2.2.3: Perform continuous wavelet transform on the reduced dimension feature, using a sampling rate of 1e6 and a scale of 65:
[0015] Where t represents the signal The time variable in is the mother wavelet function used to generate the scaled and shifted version used in the transform, express The complex conjugate of is the scaling parameter used to control the width of the wavelet, is the displacement parameter, which determines the position of the wavelet in time; S2.2.4: Resulting CWT output After further processing such as dimensionality reduction, convolution and reshaping, the final CWT feature output is obtained .
[0016] Further, the step three includes the following steps: S3.1: The output features of the GW branch and the CWT branch in step 2 are fused through a connection operation, and damage classification is performed using a multi-layer linear layer:
[0017] in, Indicates a connection operation. , , represents the batch size, Indicates the size of GW feature dimension, Indicates the size of the GWT spectrum feature dimension, Indicates i Linear layers, represents the damage classification result output by the multimodal damage detection model, N Indicates the number of damage level categories.
[0018] Furthermore, in step 4, the objective function of the multimodal damage detection model is:
[0019] in, represents the true probability of belonging to a certain category, Indicates the prediction score of the input sample belonging to a certain category, Represents the Softmax activation function.
[0020] The principles and advantages of the present invention are: 1. The present invention does not rely on baseline signals, but directly inputs the waveguide signal into the model for damage feature recognition and damage index calculation, which greatly enhances the ability to extract high-level features implicit in the signal and is conducive to expanding engineering applications in baseline-free environments.
[0021] 2. The designed GW branch can not only fully extract the temporal features of the guided wavelength sequence data through the Mamba technology based on the spatial state model, but also outperforms Transformers in computational complexity and efficiency.
[0022] 3. At the same time, unlike MGNN, the CWT branch uses a convolutional neural network suitable for local feature extraction to extract CWT spectral features, which significantly enhances the feature extraction ability and robustness of the model.
[0023] 4. In addition, the spectral features extracted by the convolutional neural network and the temporal features extracted by the Mamba technology based on the state-space model are effectively integrated through the connection operation, which comprehensively captures the multi-scale damage characteristics and can accurately construct the high-order nonlinear mapping relationship between the GW signal and the damage location, thus achieving high-precision damage detection even with limited data sets.
[0024] 5. Finally, the damage detection accuracy of the present invention does not need to rely on dense sensor arrays, which significantly reduces the complexity and cost of the sensor network. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 It is a schematic flow chart of an embodiment of a method for baseline-free multi-modal damage detection of aviation composite materials based on guided waves of the present invention; Figure 2 It is a principle block diagram of an embodiment of a method for baseline-free multi-modal damage detection of aviation composite materials based on guided waves of the present invention; Figure 3 A model network structure parameter diagram of an embodiment of a method for baseline-free multimodal damage detection of aviation composite materials based on guided waves of the present invention; Figure 4 A comparison diagram of the training and validation loss curves of the model on the data set in an embodiment of a method for non-baseline multimodal damage detection of aviation composite materials based on guided waves of the present invention; Figure 5This is a performance index comparison diagram of the model in the embodiment of the guided wave-based aviation composite material baseline-free multimodal damage detection method on the aviation composite material damage detection task; Figure 6 This is a comparison diagram of model ablation experiments in an embodiment of a baseline-free multimodal damage detection method for aviation composite materials based on guided waves of the present invention. DETAILED DESCRIPTION
[0026] The following is further described in detail through specific implementation methods: The embodiment is basically as shown in the attached Figure 1 As shown: A baseline-free multimodal damage detection method for aviation composite materials based on guided waves, such as Figure 1 As shown: This method uses the GW branch of the MMDD model to reduce the dimension of the waveguide of the original damage signal, and uses the Mamba technology based on the state space model to extract features and linear mapping, and outputs GW features; the CWT branch performs continuous wavelet transform on the waveguide signal, and then uses a convolutional neural network to extract the transformed spectral features and outputs CWT features; then, the features of the two branches are fused through a connection operation, and a multi-layer linear layer is used for damage classification. During training, the cross entropy loss function is used as the model objective function, and the model is trained and optimized by minimizing the difference between the predicted probability distribution and the true label probability distribution. The validation set is used to adjust the optimal hyperparameters and network structure to find the best model, and the test set samples are used to verify the damage detection effect of the model and output performance indicators to achieve high-precision detection of various different damages of aviation composite materials. In this embodiment, a baseline-free multimodal damage detection method for aviation composite materials based on waveguide specifically includes the following steps: Step 1: Collect the guided wave signals of aviation composite materials with several types of damage, and divide the collected guided wave signals into training set, validation set and test set, where the data proportions of the training set, validation set and test set are 80%, 10% and 10% respectively, and each data set is independent of each other.
[0027] Step 2. Based on the Mamba technique and convolutional neural network of the state space model, a multi-modal damage detection model MMDD, a GW branch and a CWT branch of MMDD are designed respectively. The GW branch extracts time features from the guided wavelength sequence data through the Mamba technique, and the CWT branch extracts spectral features from the spectrum diagram of the continuous wavelet transform through the convolutional neural network.
[0028] The step 2 comprises the following steps: Step 2.1: The GW branch extracts time features from the guided wavelength sequence data through an average pooling layer, a reshaping layer, a Mamba layer, a reshaping layer, and a fully connected layer in sequence; Step 2.2: The CWT branch extracts spectral features from the continuous wavelet transform spectrum map through an average pooling layer, a continuous wavelet transform layer, an average pooling layer, two convolutional layers, and a reshaping layer in sequence; The Mamba technique is based on the state space model, and the output of the state model at multiple time points is derived as follows:
[0029] in, express The current state at the moment, represents the input control quantity at time t, represents the system output at time t, represents the state transfer matrix, Indicates the influence of the control quantity on the state quantity, Indicates the impact of the current state quantity on the system output; exist The state equation at this time is shown as follows:
[0030] The convolution transport of the convolutional network is shown in the following formula:
[0031] Using the above method, the convolution operation can be parallelized, thus greatly improving the computational efficiency of the network.
[0032] Step 2.1 specifically includes the following steps: S2.1.1: The GW branch uses the waveguide signal in step 1 as input data: ,in represents the set of real numbers, represents the batch size, Indicates the sequence length of the GW signal; S2.1.2: Perform average pooling on the data using the following formula:
[0033] Get dimensionality reduction features ; S2.1.3: Sequence After grouping, it is expressed as ; S2.1.4: Input the grouped sequence data into the Mamba layer for feature extraction in, represents the Mamba model based on the state space model; S2.1.5: For serialized features Reshape , and linear processing After that, the final GW features are obtained. .
[0034] The step 2.2 specifically includes the following steps: S2.2.1: The CWT branch takes the waveguide signal in step 1 as input data: ,in represents the batch size, Indicates the sequence length of the GW signal; S2.2.2: Perform average pooling on the data using the following formula:
[0035] Get dimensionality reduction features ; S2.2.3: Perform continuous wavelet transform on the reduced dimension feature, using a sampling rate of 1e6 and a scale of 65:
[0036] Where t represents the signal The time variable in is the mother wavelet function used to generate the scaled and shifted version used in the transform, express The complex conjugate of is the scaling parameter used to control the width of the wavelet, is the displacement parameter, which determines the position of the wavelet in time; S2.2.4: Resulting CWT output After further processing such as dimensionality reduction, convolution and reshaping, the final CWT feature output is obtained .
[0037] Step 3: The features extracted by the GW branch and the continuous wavelet branch are fused to obtain multi-dimensional damage features, and then mapped to a new space for classification.
[0038] Specifically, the output features of the GW branch and the CWT branch in step 2 are fused through a connection operation, and damage classification is performed using a multi-layer linear layer:
[0039] in, Indicates a connection operation. , , represents the batch size, Indicates the size of GW feature dimension, Indicates the size of the GWT spectrum feature dimension, Indicates i Linear layers, represents the damage classification result output by the multimodal damage detection model, N Indicates the number of damage level categories.
[0040] Step 4: Use the cross entropy loss function as the objective function of the multimodal damage detection model to perform the classification task by measuring the difference between the predicted score and the true probability that the sample belongs to a certain category. The objective function of the multimodal damage detection model is:
[0041] in, represents the true probability of belonging to a certain category, Indicates the prediction score of the input sample belonging to a certain category, Represents the Softmax activation function.
[0042] Step 5: Use the validation set to adjust the optimal hyperparameters and network structure to find the best model, and use the test set to verify the damage detection effect of the model and output performance indicators.
[0043] Step five includes the following steps: S5.1: Use the loss function output by the validation set during the training process to evaluate the model performance, continuously adjust and test different hyperparameter combinations and network structures, and select the model configuration with the best performance on the validation set as the best model. The optimal configuration is: 400 iterations, 0.0001 learning rate, 16 batch size, and Adam optimizer. S5.2: Calculate the detection performance index precision (P) of the multimodal damage detection model:
[0044] in, represents the number of states correctly identified, Represents the number of instances that were misidentified.
[0045] In this embodiment, the data acquisition platform configuration includes a DG4062 arbitrary waveform generator, an MSO6074 oscilloscope, an ATA-2031 high-voltage amplifier, and four custom ultrasonic sensors, with a sensor height of 3 cm, a diameter of 2 cm, and a center frequency of 80 kHz. The test specimen is a T300 / 7901 CFRP composite plate with a size of 1000 mm × 1000 mm × 2 mm (stacked in the order of The seven damages were simulated by uniformly fixing the mass block on the composite plate using an ultrasonic coupler. The sampling frequency of data acquisition was 10 MHz, with a total of 10,000 sampling points. The collected pilot signals were input into Figure 2 By effectively combining the features obtained by the GW branch and the CWT branch in MMDD, the multi-scale damage features are fully captured. The specific network structures of the two branches are as follows: Figure 3 As shown in Figure 2, the features of the two branches are fused by designing a connection operation, and multi-layer linear layers are used for damage classification.
[0046] To evaluate the effectiveness of the MMDD method, it was compared with multiple deep learning models, including LSTM, BCNN, FCN, CNN, AECNN, DCSCNet, and MGNN. Figure 4 The comparison of the training and validation loss curves of the model on the data set is shown. MMDD reaches the optimal value after 400 training rounds. Compared with other methods, the validation set loss is the smallest and the convergence is more stable during the training process. Figure 5 The performance index comparison chart of the model shown in the figure on the aviation composite material damage detection task. It can be seen from the figure that compared with other models, MMDD has the best performance index and has achieved 100% accuracy. In addition, in order to verify the effectiveness of different components of the MMDD model, targeted ablation experiments were carried out. Figure 6 It can be seen that the accuracy of LSTM and GRU both reached 71.43%, which is lower than the accuracy of the proposed method. The results with and without the CWT spectrum were compared. Without the CWT spectrum, the accuracy of the proposed method was only 57.14%, while after adding the CWT spectrum, the accuracy increased to 100%. This improvement is attributed to the fact that the CWT spectrum can effectively capture the spectral characteristics of the waveguide, enabling the damage detection framework to more accurately distinguish damage features of different frequencies.
[0047] The MMDD model of this embodiment does not rely on baseline signals (signals from healthy or original structures), but directly uses the complete waveguide time series signal as network input. First, the time-frequency characteristics of the waveguide signal are analyzed at multiple scales using continuous wavelet transform, and high-resolution time-frequency diagrams of the signal are drawn to extract hidden patterns and structural features, thereby enhancing the learning and feature extraction capabilities of the network model. Subsequently, a convolutional neural network model is used to deeply mine and optimize the waveguide and damage-related features in the CWT spectrum, and the Mamba technology based on the state space model is used to parallelly calculate and extract the time series features of the waveguide. By effectively integrating the spectral features extracted by the convolutional neural network and the time series features extracted by the Mamba technology based on the state space model, MMDD realizes a comprehensive analysis of the hidden damage features in the waveguide and extracts high-level damage-related features from multiple dimensions. The model is trained and optimized by adopting the cross entropy loss function as the model objective function to minimize the difference between the predicted probability distribution and the true label probability distribution. Finally, the best model is found using the output of the validation set, and high-precision damage detection is achieved through the correlation between the waveguide signal and the damage location accurately established by the model. Due to the model's excellent long-range feature extraction capability, this method can achieve accurate damage detection even with a limited dataset of only 140 damage signals.
[0048] Traditional methods rely on baseline signals, which can easily lead to deviations and misdiagnosis in damage identification, and require dense sensor arrays to increase the weight and cost of the structural health monitoring system. At the same time, they rely heavily on damage feature extraction, thereby losing the potential multi-scale damage information in the waveguide signal. The present invention does not rely on baseline signals, and directly uses the waveguide signal as the model input for damage feature identification and damage index calculation, which greatly enhances the ability to extract high-level features implicit in the signal, and the damage detection accuracy does not need to rely on dense sensor arrays. In addition, MMDD combines the Mamba technology based on the state space model with convolutional neural networks to promote efficient parallel computing, comprehensively capture multi-scale damage features, and accurately construct a high-order nonlinear mapping relationship between the signal and the damage location to achieve high-precision damage location information output. Compared with traditional methods, the MMDD model shows excellent effectiveness, accuracy and robustness in damage detection under baseline-free environments and limited data conditions, while significantly reducing the complexity and cost of the sensor network, providing a more efficient and economical solution for structural health monitoring of aviation composites.
[0049] The above are only embodiments of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme is not described in detail here. Ordinary technicians in the relevant field know all the common technical knowledge in the technical field to which the invention belongs before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.
Claims
1. A baseline-free multimodal damage detection method for aviation composite materials based on guided waves, characterized by: The following steps are involved: Step 1: collect guided wave signals of aviation composite materials with several types of damage, and divide the collected guided wave signals into a training set, a verification set, and a test set; Step 2: Based on the Mamba technique of the state space model and the convolutional neural network, a multimodal damage detection model MMDD, a GW branch and a CWT branch of the MMDD are designed respectively. The GW branch extracts time features from the guided wavelength sequence data through the Mamba technique based on the state space model, and the CWT branch extracts spectral features from the spectrum diagram of the continuous wavelet transform through the convolutional neural network. Step 3: The features extracted by the GW branch and the CWT branch are fused to obtain multi-dimensional damage features, and then mapped to a new space for classification; Step 4: The cross entropy loss function is used as the objective function of the multimodal damage detection model to perform the classification task by measuring the difference between the predicted score and the true probability that the sample belongs to a certain category; Step 5: Use the validation set to adjust the optimal hyperparameters and network structure to find the best model, and use the test set to verify the damage detection effect of the model and output the performance indicators; The step five comprises the following steps: S5.1: Use the loss function output by the validation set during the training process to evaluate the model performance, continuously adjust and test different hyperparameter combinations and network structures, and select the model configuration with the best performance on the validation set as the best model. The optimal configuration is: 400 iterations, 0.0001 learning rate, 16 batch size, and Adam optimizer. S5.2: Calculate the detection performance index precision (P) of the multimodal damage detection model: in, represents the number of states correctly identified, Represents the number of instances that were misidentified.
2. The method for baseline-free multimodal damage detection of aviation composite materials based on guided wave according to claim 1, characterized in that: In the step 1, the data proportions of the training set, validation set and test set are 80%, 10% and 10% respectively, and each data set is independent of each other.
3. The method for baseline-free multimodal damage detection of aviation composite materials based on guided wave according to claim 1, characterized in that: In the step 2, the GW branch and the CWT branch of the multimodal damage detection model include the following steps: Step 2.1: The GW branch extracts time features from the guided wavelength sequence data through an average pooling layer, a reshaping layer, a Mamba layer, a reshaping layer, and a fully connected layer in sequence; Step 2.2: The CWT branch extracts spectral features from the continuous wavelet transform spectrum map through an average pooling layer, a continuous wavelet transform layer, an average pooling layer, two convolutional layers, and a reshaping layer in sequence; The Mamba technique is based on the state space model, and the output of the state model at multiple time points is derived as follows: in, express The current state at the moment, represents the input control quantity at time t, represents the system output at time t, represents the state transfer matrix, Indicates the influence of the control quantity on the state quantity, Indicates the impact of the current state quantity on the system output; exist The state equation at this time is shown as follows: The convolution transport of the convolutional network is shown in the following formula: 。 4. The method for baseline-free multimodal damage detection of aviation composite materials based on guided wave according to claim 3, characterized in that: The step 2.1 comprises the following steps: S2.1.1: The GW branch uses the waveguide signal in step 1 as input data: ,in, represents the set of real numbers, represents the batch size, Indicates the sequence length of the GW signal; S2.1.2: Perform average pooling on the data using the following formula: Get dimensionality reduction features ; S2.1.3: Sequence After grouping, it is expressed as ; S2.1.4: Input the grouped sequence data into the Mamba layer for feature extraction in, represents the Mamba model based on the state space model; S2.1.5: For serialized features Reshape , and linear processing After that, the final GW features are obtained. .
5. The method for baseline-free multimodal damage detection of aviation composite materials based on guided waves according to claim 4, characterized in that: The step 2.2 comprises the following steps: S2.2.1: The CWT branch takes the waveguide signal in step 1 as input data: ,in represents the batch size, Indicates the sequence length of the GW signal; S2.2.2: Perform average pooling on the data using the following formula: Get dimensionality reduction features ; S2.2.3: Perform continuous wavelet transform on the reduced dimension feature, using a sampling rate of 1e6 and a scale of 65: Where t represents the signal The time variable in is the mother wavelet function used to generate the scaled and shifted version used in the transform, express The complex conjugate of is the scaling parameter used to control the width of the wavelet, is the displacement parameter, which determines the position of the wavelet in time; S2.2.4: Resulting CWT output After further processing such as dimensionality reduction, convolution and reshaping, the final CWT feature output is obtained .
6. The method for baseline-free multimodal damage detection of aviation composite materials based on guided wave according to claim 5, characterized in that: The step three comprises the following steps: S3.1: The output features of the GW branch and the CWT branch in step 2 are fused through a connection operation, and damage classification is performed using a multi-layer linear layer: in, Indicates a connection operation. , , represents the batch size, Indicates the size of GW feature dimension, Indicates the size of the GWT spectrum feature dimension, Indicates i Linear layers, represents the damage classification result output by the multimodal damage detection model, N Indicates the number of damage level categories.
7. The method for baseline-free multimodal damage detection of aviation composite materials based on guided wave according to claim 1, characterized in that: In step 4, the objective function of the multimodal damage detection model is: in, represents the true probability of belonging to a certain category, Indicates the prediction score of the input sample belonging to a certain category, Represents the Softmax activation function.