Method and system for identifying low-speed impact energy level of composite material
By acquiring the two-dimensional time spectrum of composite materials using a single piezoelectric ceramic sensor and utilizing a residual convolutional network with a multi-scale time-frequency interactive attention module, the problem of sensor quantity and power consumption in low-speed impact identification of composite materials is solved, achieving efficient and robust impact energy level identification, which is suitable for safety assessment and health monitoring of composite materials.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-03-27
AI Technical Summary
Existing data-driven methods rely on multi-point sensor deployment for low-velocity impact identification of composite materials, which increases system weight and power consumption, limiting their practical application on confined platforms such as aircraft.
A single piezoelectric ceramic sensor is used to acquire the two-dimensional time spectrum of transient vibration signals, and the signals are identified by a pre-trained impact energy level recognition model. The model includes a feature extraction module, a residual convolutional backbone network, and a classification and recognition module. The impact features are extracted using a multi-scale time-frequency interactive attention module.
It achieves efficient and robust low-velocity impact energy level identification under single sensor conditions, reduces the number of sensors and system weight, meets the requirements of lightweight and low power consumption, and is suitable for safety assessment and health monitoring of composite material structures.
Smart Images
Figure CN121741014A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of composite material monitoring, in particular to a method and system for identifying low-speed impact energy level of composite material. BACKGROUND
[0002] Low-speed impact caused by debris and hail is one of the main factors threatening the integrity of aircraft composite components. Different energy levels of impact can lead to different degrees of structural degradation from slight damage to delamination and fiber fracture, and low-speed impact often forms almost invisible damage that is difficult to detect by conventional visual inspection. Therefore, accurately identifying impact energy is of great significance to the safety evaluation and health monitoring of composite structures.
[0003] The rapid development of artificial intelligence has promoted the rise of data-driven SHM methods. Deep learning, with its powerful feature automatic extraction and pattern recognition capabilities, has been widely used in composite impact event detection, classification and positioning tasks. In specific composite impact monitoring tasks, Azad et al. proposed a deep learning framework based on Lamb waves and multi-sensor PZT array, which realized impact source positioning and damage severity evaluation, and demonstrated the good recognition ability of CNN, GRU and other models in vibration signals. Mezeix et al. built a deep learning framework to predict composite impact damage, and through training on impact data, realized efficient and accurate damage identification and classification. In addition, Khan et al. proposed a defect detection scheme based on vibration signals and deep residual network, which illustrated the advantages of deep learning in dealing with complex composite dynamic responses and the feasibility of delamination detection. The research of Ouabdou et al. showed that combining vibration pattern analysis with CNN, unsupervised learning and other AI technologies can significantly improve the accuracy and efficiency in composite defect detection tasks.
[0004] However, existing data-driven methods generally rely on multi-point sensor deployment to provide sufficient spatial information to support model training and prediction. This not only increases the weight and power consumption of the system, but also limits the practical application on limited platforms such as aircraft. SUMMARY
[0005] Therefore, it is necessary to provide a method and system for identifying low-speed impact energy level of composite material in view of the above technical problems.
[0006] The embodiment of the present application provides a method for identifying low-speed impact energy level of composite material, which comprises the following steps: obtaining a two-dimensional time-frequency spectrum corresponding to a transient vibration signal of a composite material under the action of low-speed impact and acquired by a single piezoelectric ceramic sensor; The two-dimensional time-frequency spectrum is input into a pre-trained impact energy level identification model, and the pre-trained impact energy level identification model comprises: a feature extraction module, a residual convolution main network and a classification identification module connected in sequence, and the residual convolution main network comprises: a plurality of residual blocks connected in sequence and each embedding a multi-scale time-frequency interaction attention module, and the multi-scale time-frequency interaction attention module comprises: a multi-scale convolution module, a global pooling and convolution module and an interactive fusion module. The feature extraction module is used for performing convolution operation on the two-dimensional time-frequency spectrum to obtain local time-domain features; the multi-scale convolution module is used for performing parallel convolution of different scales on the local time-domain features to extract short-time transient features and long-time trend features in the time dimension and obtain multi-scale time features; the global pooling and convolution module is used for performing global average pooling and convolution operation on the local time-domain features along the time axis to generate adaptive weights representing the importance of the composite material in different frequency bands in the frequency dimension and obtain frequency attention weights; the interactive fusion module is used for performing interactive fusion on the multi-scale time features and the frequency attention weights to obtain a two-dimensional attention map; the two-dimensional attention map is used for weighting the local time-domain features to determine non-stationary features of low-speed impact action and obtain an output feature map; and the classification identification module is used for performing classification prediction on the output feature map to obtain an impact energy level classification result.
[0007] Optionally, the two-dimensional time-frequency spectrum corresponding to the transient vibration signal of the composite material under low-speed impact action acquired by the single piezoelectric ceramic sensor is acquired, and specifically includes: The transient vibration signal of the composite material under low-speed impact action acquired by the single piezoelectric ceramic sensor is acquired, and the transient vibration signal is arranged according to the acquisition time sequence to obtain original time sequence data. The original time sequence data is converted from the transient vibration signal to a time-frequency spectrum representation by short-time Fourier transform based on the following formula: ; Wherein, is the time-frequency spectrum representation, is the impact response signal, is a window function centered at is a window function centered at is a frequency variable; The complex modulus value of the time-frequency spectrum representation is taken based on the following formula, and logarithmic normalization processing is performed to obtain a two-dimensional time-frequency spectrum taking time and frequency as coordinates and taking amplitude as intensity information: ; Wherein, is the two-dimensional time-frequency spectrum.
[0008] Optionally, the feature extraction module is used for performing convolution operation on the two-dimensional time-frequency spectrum to obtain local time-domain features, and specifically includes: The feature extraction module includes: a convolutional layer, a batch normalization layer, and a non-linear activation function layer connected in sequence; The initial feature map is obtained by performing convolution operations on the two-dimensional time spectrum through convolutional layers; The initial feature map is normalized by a batch normalization layer, and the normalized feature map is then nonlinearly transformed by a nonlinear activation function layer to obtain local temporal features.
[0009] Optionally, multi-scale temporal features can be obtained by performing parallel convolutions at different scales on local temporal features using a multi-scale convolution module based on the following formula: ; in, For each time step, the corresponding time feature is... Indicates the kernel size as One-dimensional convolution operation, For the scale number, It is a local time-domain feature; Based on the following formula, global pooling and convolution modules are used to perform global average pooling on local temporal features along the time axis to obtain an overall description of different frequency bands: ; in, For each frequency band, a comprehensive description is provided. Where T is the frequency and T is the time axis. t This refers to the time step in the timeline; Based on the following formula, a nonlinear mapping of the overall description of different frequency bands is performed through a one-dimensional convolution operation, and the frequency attention weights of each frequency band are determined by the sigmoid activation function: ; in, Frequency attention weights for each frequency band, It is the sigmoid activation function. This is a one-dimensional convolution operation; Based on the following formula, a two-dimensional attention map is obtained by interactively fusing multi-scale temporal features and frequency attention weights through an interactive fusion mechanism: ; in, It is a two-dimensional attention map. For at any time Time feature vector, For frequency The corresponding frequency attention weights.
[0010] Optionally, the output feature map is obtained by weighting the local temporal features using a two-dimensional attention map based on the following formula: ; in, To output the feature map, For the channel, For a moment, For frequency; The impact energy level classification result is obtained by classifying and predicting the output feature map using the following formula through the classification and recognition module: ; in, For category The weight vector, For category The bias.
[0011] Optionally, training an impact energy level recognition model includes: Obtain historical two-dimensional time-frequency spectra and their corresponding historical impact energy level classification results; Inputting the historical two-dimensional time spectrum into the impact energy level identification model yields the predicted impact energy level classification results; The weighted cross-entropy loss between historical and predicted impact energy level classification results is determined based on the following formula: ; in, For category weight coefficients, One-hot encoding of the actual label; The impact energy level identification model is trained with the goal of minimizing the weighted cross-entropy loss, resulting in the trained impact energy level identification model.
[0012] Optionally, before inputting the historical two-dimensional time spectrum into the impulse energy level identification model, the data enhancement preprocessing of the two-dimensional time spectrum is also included. The data enhancement includes randomly masking local areas in the time or frequency dimension to simulate signal loss and noise interference in actual applications.
[0013] This invention provides a system for identifying low-velocity impact energy levels in composite materials, comprising: The data acquisition module is used to acquire the two-dimensional time spectrum corresponding to the transient vibration signal of the composite material under low-speed impact by a single piezoelectric ceramic sensor. The model construction module is configured to input a two-dimensional time-frequency spectrum into a pre-trained impact energy level identification model, and the pre-trained impact energy level identification model comprises a feature extraction module, a residual convolution backbone network and a classification identification module connected in sequence, the residual convolution backbone network comprises a plurality of residual blocks connected in sequence and each residual block embeds a multi-scale time-frequency interaction attention module, and the multi-scale time-frequency interaction attention module comprises a multi-scale convolution module, a global pooling and convolution module and an interactive fusion mechanism. The model processing module is configured to perform a convolution operation on the two-dimensional time-frequency spectrum through the feature extraction module to obtain local time-domain features, perform parallel convolution of different scales on the local time-domain features through the multi-scale convolution module to extract short-time transient features and long-time trend features in the time dimension and obtain multi-scale time features, perform global average pooling and convolution operation on the local time-domain features along the time axis through the global pooling and convolution module to generate adaptive weights representing the importance of the composite material in different frequency bands in the frequency dimension and obtain frequency attention weights, interactively fuse the multi-scale time features and the frequency attention weights through the interactive fusion mechanism to obtain a two-dimensional attention map, weight the local time-domain features through the two-dimensional attention map to determine non-stationary features of low-speed impact action and obtain an output feature map, and perform classification prediction on the output feature map through the classification identification module to obtain an impact energy level classification result.
[0014] Compared with the prior art, the above-mentioned composite material low-speed impact energy level identification method and system provided by the embodiments of the present application has the following beneficial effects: The present application provides a low-speed impact energy level classification framework based on a single PZT sensor, converts a one-dimensional time sequence signal acquired by a single piezoelectric ceramic sensor into a two-dimensional time-frequency spectrum through short-time Fourier transform, so as to fully extract transient time-frequency features of impact in single-point data, and then automatically focuses on key time-domain fragments and sensitive frequency bands of the impact event through a residual convolution backbone network embedded with a multi-scale time-frequency interaction attention module, so as to realize efficient mining of non-stationary features from limited data. This technical process enables the impact energy level identification model to complete impact energy level identification only by using a single piezoelectric sensor, directly reduces the number of sensors, system weight and power consumption, realizes efficient and robust identification of low-speed impact of the composite material, and meets the strict requirements of aircraft and other platforms for lightweight and low power consumption. BRIEF DESCRIPTION OF DRAWINGS
[0015] Figure 1 A flowchart of a composite material low-speed impact energy level identification method provided in an embodiment; Figure 2 A short Fourier transform time-frequency graph of a composite material low-speed impact energy level identification method provided in an embodiment; Figure 3Data enhancement effect comparison chart of a composite low-velocity impact energy level identification method provided in an embodiment, Figure 3 (a) in (a) is an original spectrum chart, Figure 3 (b) in (b) is a random mask chart; Figure 4 Network architecture schematic diagram of a composite low-velocity impact energy level identification method provided in an embodiment; Figure 5 MS-TFCA module architecture schematic diagram of a composite low-velocity impact energy level identification method provided in an embodiment; Figure 6 Simulation model schematic diagram of a composite low-velocity impact energy level identification method provided in an embodiment; Figure 7 Model training result schematic diagram of a composite low-velocity impact energy level identification method provided in an embodiment, Figure 7 (a) in (a) is an actual data loss curve schematic diagram, Figure 7 (b) in (b) is an actual data accuracy rate schematic diagram, Figure 7 (c) in (c) is a simulation data loss curve schematic diagram, Figure 7 (d) in (d) is a simulation data accuracy rate schematic diagram; Figure 8 Confusion matrix schematic diagram of a composite low-velocity impact energy level identification method provided in an embodiment, Figure 8 (a) in (a) is an actual data confusion matrix schematic diagram, Figure 8 (b) in (b) is a simulation data confusion matrix schematic diagram. DETAILED DESCRIPTION
[0016] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.
[0017] Carbon Fiber Reinforced Polymer (CFRP) is a kind of lightweight material with high specific strength and high specific stiffness, which has been widely used in aerospace, automotive manufacturing and rail transportation. Composite structures represented by aircraft components often suffer low-velocity impact caused by debris and hail during service. This kind of impact often does not leave obvious traces on the surface, but can cause invisible damage (Barely Visible Impact Damage, BVID) such as matrix cracking and delamination, thereby weakening the mechanical properties of the material and reducing the safety margin of the structure. If not detected and repaired in time, such hidden damage will pose a potential threat to the safe operation of the aircraft. It is necessary to check the impact damage of aircraft components and maintain them in time.
[0018] With the development of sensing technology and data processing methods, Nondestructive Testing (NDT) and Structural Health Monitoring (SHM) have gradually become an important way to identify impact and assess damage. The monitoring methods currently in use can be roughly divided into three categories: image-based technology (such as infrared thermography, X-ray CT), acoustic signal-based technology (such as ultrasonic guided waves, acoustic emission), and vibration signal-based technology. Among them, image-based methods have high precision, but the equipment is expensive and the operation is complex, making it difficult to adapt to online real-time detection; acoustic methods perform well in impact positioning and damage identification, but rely on multiple sensors, making system deployment complex; in comparison, vibration signals can directly reflect the dynamics of the structure, so they have received more and more attention in the monitoring of low-velocity impact damage of composite materials.
[0019] In recent years, piezoelectric sensors (PZT) have been widely used in vibration signal acquisition due to their high sensitivity, small size, and ease of integration. They are also used to form multi-sensor arrays to detect and locate impact events in composite structures. Santos et al. used a PZT array to evaluate low-speed impact damage in glass fiber composite panels, achieving quantitative damage analysis based on vibration response. Wang et al. proposed a two-dimensional MUSIC imaging-based array signal processing method that maintains high impact source identification accuracy even under environmental disturbances such as temperature changes. In addition, Elkjaer et al. developed a printable PZT array, demonstrating its potential for lightweight and integrated applications. Wang et al. proposed a two-stage collaborative method that combines sparse and dense PZT arrays with virtual feature mapping constructed by finite element simulation to achieve high-precision monitoring of low-speed impact localization on CFRP panels. Zheng et al. designed an integrated MUSIC array for high-precision diagnosis of complex damage in composite materials, which can be robustly located even in complex environments. Carani et al. reviewed the significant potential of embedded piezoelectric sensors in impact monitoring of composite structures, emphasizing their high sensitivity, lightweight, and data-driven analysis advantages.
[0020] The present application embodiment proposes a method for identifying the energy level of low-speed impact on composite materials, comprising: Obtaining a two-dimensional time-frequency spectrum corresponding to the transient vibration signal obtained by a single piezoelectric ceramic sensor under the action of low-speed impact on the composite material.
[0021] Inputting the two-dimensional time-frequency spectrum into a pre-trained impact energy level identification model, the pre-trained impact energy level identification model comprising: a feature extraction module, a residual convolution backbone network, and a classification and identification module connected in sequence, the residual convolution backbone network comprising: a plurality of residual blocks connected in sequence and each residual block embedding a multi-scale time-frequency interaction attention module, the multi-scale time-frequency interaction attention module comprising: a multi-scale convolution module, a global pooling and convolution module, and an interactive fusion module.
[0022] The local time domain feature is obtained by performing convolution operation on the two-dimensional time-frequency spectrum through the feature extraction module. The multi-scale time feature is obtained by performing parallel convolution of different scales on the local time domain feature through the multi-scale convolution module to extract short-time transient feature and long-time trend feature in the time dimension. The frequency attention weight is obtained by performing global average pooling and convolution operation on the local time domain feature along the time axis through the global pooling and convolution module to generate adaptive weight representing the importance of the composite material in different frequency bands in the frequency dimension. The two-dimensional attention map is obtained by interactive fusion of the multi-scale time feature and the frequency attention weight through the interactive fusion module. The output feature map is obtained by weighting the local time domain feature through the two-dimensional attention map to determine the non-stationary feature of the low-speed impact action. The impact energy level classification result is obtained by performing classification prediction on the output feature map through the classification and recognition module.
[0023] Preferably, the impact energy level recognition model is trained, specifically comprising: The historical two-dimensional time-frequency spectrum and the corresponding historical impact energy level classification result are obtained. The historical two-dimensional time-frequency spectrum is input into the impact energy level recognition model to obtain the predicted impact energy level classification result.
[0024] The weighted cross-entropy loss between the historical impact energy level classification result and the predicted impact energy level classification result is determined based on the following formula: ; Wherein, is a category weight coefficient, is the one-hot encoding of the true label.
[0025] The impact energy level recognition model is trained to minimize the weighted cross-entropy loss, and the trained impact energy level recognition model is obtained.
[0026] The specific implementation is as follows: 1. Overall description.
[0027] The composite material low-speed impact energy level classification framework based on a single sensor proposed by the application mainly comprises three modules of data preprocessing, feature extraction and impact classification, as shown in Figure 1 .
[0028] Firstly, in the data preprocessing stage, the transient vibration signals collected by a single piezoelectric ceramic sensor (PZT sensor) are taken as input. Since the original time-domain signals are difficult to directly reveal the characteristics of impact energy, the Short-Time Fourier Transform (STFT) is adopted to convert them into time-frequency spectrum, which contains both time and frequency information. This conversion process not only preserves the transient response characteristics induced by impact, but also highlights the frequency distribution evolving over time, providing structured input features for the deep learning model.
[0029] Subsequently, in the feature extraction module, the constructed deep residual convolutional neural network models the time-frequency spectrum hierarchically. The combination of convolution and residual structure enables the network to efficiently capture local patterns and alleviate the gradient vanishing problem, while maintaining stable training of deep networks. On this basis, the Multi-Scale Time-Frequency Cross Attention (MS-TFCA) module is introduced, which extracts dynamic features of different time windows through multi-scale convolution and performs interactive modeling in both time and frequency dimensions, thereby avoiding information loss caused by single-dimensional attention. This design not only improves the comprehensive modeling capability of short-term transient and long-term trends, but also enhances the robustness of the model in noisy environments.
[0030] Finally, in the classification and recognition module, the extracted high-dimensional features are compressed by global average pooling and input into the fully connected layer, and finally the different impact energy levels are distinguished through the Softmax classifier. Combined with the verification of experimental data and finite element simulation data, the proposed framework shows high precision and good generalization performance in the multi-impact energy recognition task. Overall, this method realizes the complete process from the original signal to the energy level classification through the organic connection between modules, ensuring the feasibility under the condition of lightweight sensors, and embodying the effectiveness of the combination of deep feature modeling and attention mechanism.
[0031] 1.1 Data preprocessing.
[0032] Composite material structures generate transient vibration signals under low-velocity impacts. These signals often contain dynamic characteristics related to the impact energy level and potential damage modes. Therefore, measuring and analyzing the dynamic response of composite material plates has become an important means of monitoring their performance evolution and identifying impact damage. However, raw time-series signals (such as voltage signals acquired by piezoelectric ceramic sensors) are often affected by noise and non-stationary characteristics, making it difficult to directly reveal differences in impact energy or the degree of damage. This necessitates the extraction of more discriminative features through appropriate signal processing methods. In composite material impact monitoring, frequency domain analysis can effectively reflect the structural dynamic behavior. While traditional Fourier transforms can provide a global spectrum, they lack the ability to characterize transient features, making it difficult to capture the dynamic changes in frequency over time in low-velocity impact responses. Therefore, time-frequency analysis methods are widely used in composite material structural health monitoring, capable of characterizing signal features simultaneously in both time and frequency dimensions, and are particularly suitable for processing non-stationary vibration signals induced by low-velocity impacts.
[0033] This invention employs Short-Time Fourier Transform (STFT) in the data preprocessing stage to convert the vibration signal acquired by a single piezoelectric ceramic sensor into a time-spectrum representation, thereby obtaining input characteristics with both time and frequency resolution. The transient vibration signal of the composite material under low-speed impact is acquired by a single piezoelectric ceramic sensor, and the transient vibration signal is arranged according to the acquisition time sequence to obtain the original time-series data. The original time-series data is then converted from transient vibration signals into a time-spectrum representation using STFT, which is defined as:
[0034] (1) in, The time-frequency spectrum is represented as follows. For impact response signal, For The window function is centered. The frequency variable is used. By passing the signal through a sliding window and performing Fourier transforms segment by segment, the local changes in the time and frequency domains can be obtained, thus revealing the transient characteristics under the impact of different energy levels.
[0035] To improve the stability and distinguishability of the input features, this invention further takes the complex modulus value of the time-spectrum representation (STFT result) and performs logarithmic normalization: (2) The resulting two-dimensional time spectrum, with time and frequency as coordinates and amplitude as intensity information, can be viewed as a tensor input to a convolutional neural network, such as... Figure 2 As shown, this two-dimensional time spectrum can effectively reveal the non-stationary characteristics in low-speed impulse responses, avoiding the limitation of traditional Fourier transforms that only provide the overall frequency distribution while ignoring transient changes.
[0036] In addition, the present application introduces a data enhancement strategy at the time-frequency spectrum level to further improve the robustness and generalization ability of the model. Specifically, by randomly masking local areas in the time and frequency dimensions, signal loss and noise interference that may occur in actual applications are simulated. The comparison before and after enhancement is shown in Figure 3 , Figure 3 (a) of (a) is the original spectrum, Figure 3 (b) is a random mask. It can be seen that this method significantly enriches the sample features while preserving the main patterns, which helps to alleviate overfitting and improve the adaptability of the model in complex working conditions.
[0037] 1.2 Network architecture.
[0038] The impact energy level identification model (deep residual convolutional neural network framework SIRNet) proposed by the present application is composed of 21 trainable parameter layers, mainly including a feature extraction module composed of a convolutional layer (3×3 convolution, 64 channels), a batch normalization (Batch Normalization, BN) layer and a nonlinear activation layer (RELU activation function), a residual module, an attention module (global average pooling layer) and a fully connected output layer. The overall structure is shown in Figure 4 .
[0039] In the feature extraction part, SIRNet uses a combination of hierarchical convolution and residual connection to represent the input time-frequency spectrum layer by layer. The feature extraction module is composed of a convolutional layer, a batch normalization layer and a nonlinear activation layer, and combines Dropout and L2 regularization during training to enhance the stability and generalization performance of the network. The number of channels of the convolutional block gradually increases with the level, and the receptive field gradually expands, so as to extract the features of the input time-frequency spectrum at different scales and realize the joint representation of local details and global patterns. The residual connection directly superimposes the input features into the block output through an identity mapping, thereby alleviating the gradient vanishing problem and promoting deep feature learning.
[0040] The multi-scale time-frequency interaction attention module MS-TFCA proposed by the present application is embedded in each residual block. This module not only realizes adaptive re-labeling of features in the channel dimension, but also strengthens the comprehensive modeling ability of transient features and global patterns through multi-scale convolution and time-frequency interaction mechanism. The final classification and recognition module is composed of a global average pooling layer, a fully connected layer and a Softmax output layer, which is used to realize accurate identification of multiple impact energy levels.
[0041] 1.2.1 Residual convolutional feature extraction.
[0042] On the basis of the overall architecture, the application further designs a residual convolution backbone network suitable for single sensor time-frequency input. The input data is a two-dimensional time-frequency spectrum obtained by STFT conversion of signals collected by a single PZT sensor , , and represent the number of channels, time steps and discrete frequency sampling points respectively.
[0043] The network first extracts the local time-frequency features at the bottom layer through initial convolution and pooling operations, and then inputs them into the residual convolution backbone network. The residual unit alleviates the gradient vanishing problem in the training of deep networks by introducing cross-layer connections, and its basic form is:
[0044] (3) Where, is a two-dimensional time-frequency spectrum, represents a nonlinear mapping composed of convolution, batch normalization and nonlinear activation function, is the corresponding trainable parameter. This structure not only ensures the effective transmission of features in the deep network, but also improves the training stability and convergence efficiency of the model.
[0045] Unlike the standard ResNet, the application introduces multi-scale convolution kernels in the residual block to process feature representations under different receptive fields in a parallel manner. Large-scale convolution can capture long-term dynamic patterns of the impulse response, while small-scale convolution is more sensitive to local transient features. Through joint representation of multi-scale features, the network can obtain multi-level information across time scales, providing more discriminative input features for the subsequent multi-scale time-frequency interactive attention mechanism.
[0046] 1.2.2 Multi-scale time-frequency interactive attention module.
[0047] To further improve the model's representation ability for non-stationary impact signals, the application embeds a newly designed multi-scale time-frequency interactive attention MS-TFCA module in the residual unit, and the overall process is as shown in Figure 5 The core idea of this module is: first, extract transient and long-term dynamic features in the time dimension through multi-scale convolution, then generate adaptive weights in the frequency dimension to characterize the importance of composite materials in different frequency bands, and finally realize time-frequency joint modeling through interactive fusion mechanism, thereby avoiding information loss caused by single-dimensional attention.
[0048] First, perform multi-scale convolution operation on the local time-domain features in the time dimension: (4) Where, is a convolution kernel with a size of One-dimensional convolution operation, For the scale number, This represents the temporal feature corresponding to each time step.
[0049] By concatenating convolution results at different scales and then weighting and fusing them, multi-scale features are obtained: (5) This design can take into account both short-term transient and long-term trend characteristics, making it more suitable for characterizing the non-stationarity of low-speed impact signals.
[0050] In frequency-dimensional modeling, global average pooling is first performed in the time direction to obtain an overall description of different frequency bands: (6) in, For the overall description corresponding to each frequency band, Where T is the frequency and T is the time axis. t This refers to the time step in the timeline.
[0051] The meaning of this formula is: for each frequency The average response over the entire time axis is calculated to obtain the overall contribution of different frequency bands.
[0052] Next, a one-dimensional convolution operation is used to perform... Nonlinear mapping, and frequency weights generated by the sigmoid activation function: (7) Here This indicates the frequency attention weight of each frequency band; the larger the value, the more critical the frequency component is in the determination of impact energy. It is the sigmoid activation function. This is a one-dimensional convolution operation.
[0053] To achieve the synergistic effect of time and frequency characteristics, this invention designs an interactive gating mechanism: (8) in, Indicates at time Time feature vector, Represents frequency The corresponding attention weights. This formula achieves interaction between the time and frequency dimensions through inner product operations, and then obtains a two-dimensional attention map through sigmoid activation. .
[0054] Finally, the attention map weights the input features point by point: (9) where, is the output feature map, i.e., at each channel , time and frequency , the feature amplitude is adaptively adjusted to highlight the discriminative components related to impact energy.
[0055] Different from SE attention which only weights in the channel dimension, MS-TFCA realizes the joint modeling of time and frequency dimensions; compared with FSatten which only considers the frequency distribution, MS-TFCA further captures transient and trend features through multi-scale time convolution. Thus, this module can more finely depict the non-stationary characteristics of low-speed impact signals, and has significant advantages in classification performance and generalization.
[0056] 1.2.3 Classification prediction and optimization strategy.
[0057] After the feature enhancement by the residual unit and MS-TFCA module, the output feature map needs to be further mapped to the class space to realize the impact energy level discrimination. For this purpose, the global average pooling is introduced at the end of the network to complete the dimension compression, and the vectorization mapping is performed through the fully connected layer. Finally, the softmax classifier is used to output the prediction probability of each class, and its mathematical form is:
[0058] (10) where, is the weight vector of class , and is the bias of class . This function is constrained by normalization to ensure that the sum of all class probabilities is 1, thereby supporting the discrimination of multiple impact energy levels.
[0059] For the problem of uneven data distribution, the weighted cross-entropy loss is used in the training process, which is in the form of: (11) where, is the class weight coefficient, which is used to improve the contribution of minority classes in optimization; is the one-hot encoding of the real label. This loss function can effectively alleviate the bias caused by the difference in the number of samples of different impact classes.
[0060] In terms of training optimization, the application adopts an adaptive strategy combining learning rate preheating and cosine annealing scheduling. This strategy first performs smooth rising in the learning rate updating process to stabilize parameter adjustment, and then realizes dynamic control of the learning rate through periodic decay, thereby balancing the convergence speed and model generalization ability. This method can effectively alleviate the problems of gradient shock and premature convergence, and improve the stability and overall performance of model training.
[0061] 1.2.4 Network configuration and implementation.
[0062] The impact energy level identification model proposed by the application is implemented based on the TensorFlow2.10 framework. Adam is used as the optimizer, and the initial learning rate is 0.0003. The data set is divided into 80% for training and 20% for testing. The test data is composed of samples not included in the training process, which ensures strict evaluation of the model's ability to generalize to unseen samples. Training is performed using 64 batch sizes, and the number of training rounds is set to 100 rounds. After each round, the model is verified using the test set, and the model parameters with the best test accuracy are saved as the final result. In addition, to ensure the robustness of the results, the entire training and testing process is repeated five times on the data set, each time using a different random partitioning method for training and testing. The final experimental results are the average of the five experiments to reduce the impact of accidental factors on the evaluation results.
[0063] 2. Experimental process.
[0064] 2.1 Data set and experimental description.
[0065] The experiment of the application uses the thin plate structure low-speed impact event data set published by Katsidimas et al. The data set was first published by the Patras University team at SenSys 2022, providing a standardized benchmark for low-speed impact detection and positioning research. The experimental object is a piece of acrylic sheet with a size of 300x300x4mm, and four piezoelectric ceramic sensors are pasted on its four corners, model CEB-35D26, used to record the transient vibration response signal caused by impact. The impact load is generated by a steel ball with a diameter of 9.5mm and a mass of 3.53g, which is free-falling, with a falling height from 10cm to 20cm in increments of 0.5cm. To ensure the robustness of the experiment, each impact condition is repeated at least three times, and a total of 771 experimental samples are obtained.
[0066] In terms of data acquisition, sensor signals were sampled at a rate of 100 kHz by an Arduino NANO 33 BLE microcontroller and transmitted to a computer in real time for storage. Each experiment recorded 5000 sampling points, which could completely capture the impact process and the induced structural natural frequency characteristics. The present invention only selected the data of sensor A as input to simulate the impact energy level identification task under the constraint of a single sensor. Different drop heights were further divided into 11 energy level categories, and all signals were converted into time-frequency spectrum by short-time Fourier transform and input into a deep residual convolutional neural network for classification. Unlike the original research, which mainly focuses on impact detection and positioning, the present invention focuses more on classifying impact energy levels under single sensor conditions to explore its application potential on resource-constrained platforms such as aerospace structural components.
[0067] 2.2 Simulation verification.
[0068] Although experimental data can verify the effectiveness of the method under real conditions, the material matrix is acrylic plate, which is significantly different from carbon fiber reinforced composites widely used in the aerospace field. To further evaluate the applicability and generalization ability of the proposed method in engineering materials, the present invention establishes a low-velocity impact finite element model of CFRP laminated plate based on the Abaqus explicit dynamics platform, as shown in Figure 6 The model size is consistent with the experiment, and the plate size is 300x300x4mm. The impact body is a rigid steel ball with a diameter of 9.5mm and a mass of 3.53g. The impact height is set to 10 to 20cm, which corresponds to the experimental conditions completely to ensure the comparability of the load conditions. The material parameters of the CFRP laminated plate are set as follows: the material performance is as follows: Young's modulus E=73.1GPa, Poisson's ratio μ=0.33, density ρ=2780kg / m³. This parameter combination can reasonably reflect the mechanical properties of actual aerospace-grade composites.
[0069] In terms of damage modeling, the Hashin criterion is combined with the progressive damage evolution model to describe the failure process of the material, thereby distinguishing typical modes such as fiber tension, fiber compression, matrix cracking, and compression failure, and realizing damage evolution through stiffness degradation. The plate is meshed with C3D8R elements, and local encryption is performed in the impact area. The boundary conditions are set as four simply supported edges to approximate the actual constraints. The general contact is defined between the steel ball and the plate, and the initial velocity is obtained by converting the free fall height. To simulate the response of the sensor, the acceleration and displacement time history signals are extracted at the corners of the plate, and the sampling rate and sampling length consistent with the experiment are maintained. After the STFT conversion of the obtained simulation signals, the same preprocessing and classification process as the experimental data is adopted, and finally input into the deep residual convolutional neural network proposed in the invention for energy level identification. The double verification of experimental and simulation results shows that the proposed method not only performs well under real acquisition conditions, but also has stable classification ability in CFRP structures, highlighting its good generalization and engineering application potential.
[0070] 2.3 Experimental results and analysis.
[0071] Figure 7 The training loss and classification accuracy curves of SIRNet on the experimental data set and the simulation data set with the number of iterations are shown. Figure 7 (a) in FIG. 1 is a schematic diagram of the actual data loss curve, Figure 7 (b) in FIG. 1 is a schematic diagram of the actual data accuracy, Figure 7 (c) in FIG. 1 is a schematic diagram of the simulation data loss curve, Figure 7 (d) in FIG. 1 is a schematic diagram of the simulation data accuracy. From the results, it can be seen that the model on both data sources shows good convergence characteristics: on the experimental data set, SIRNet reaches a stable accuracy after about 30 epochs and maintains a smooth upward trend; on the simulation data set, although there are some differences in signal features and real data, the model can still converge at about 40 epochs, and the curves of validation accuracy and training accuracy are basically consistent, without obvious overfitting. This shows that the proposed method has strong stability and generalization ability in real scenarios and finite element simulation conditions.
[0072] In the comparative experiment of different models, SIRNet performs the most outstanding in recognition accuracy. Compared with the classic convolutional neural network and deep residual network, SIRNet achieves the highest classification accuracy on experimental data, and also maintains stable performance on simulation data. This result shows that the multi-scale time-frequency interaction attention module can effectively enhance the feature expression ability, making the model still superior to existing methods under the condition of limited sensor input. The specific comparison results are shown in Table 1, further highlighting the performance advantages of the proposed method.
[0073] Table 1 Comparison results table To verify the effect of the improved module, ablation experiments are carried out, and multi-scale modeling, time-frequency interaction attention or both are removed for comparison. The experimental results show that if the multi-scale mechanism is missing, the model accuracy significantly decreases when distinguishing adjacent impact energy levels; if the time-frequency interaction attention is removed, the overall feature extraction capability is weakened, and the performance also decreases. Only when both are combined, the model can balance between capturing global and local features, thus achieving the optimal result. The related ablation comparison results are shown in Table 2.
[0074] Table 2 Ablation comparison results table Further confusion matrix results reveal the model's classification ability at different impact energy levels. Overall, SIRNet maintains a high recognition accuracy on most classes, especially in high-energy impact recognition, with almost no misjudgment. For some adjacent low-energy classes, the model's confusion rate increases slightly, but it is still better than other comparison methods. This shows that the proposed method has strong discrimination ability in fine-grained classification tasks. Figure 8 The confusion matrix results are shown in Fig. 2, Figure 8 Fig. 2 (a) is a schematic diagram of the actual data confusion matrix, Figure 8 Fig. 2 (b) is a schematic diagram of the simulation data confusion matrix, which intuitively reflects the advantages and disadvantages of the model in multi-class impact energy recognition.
[0075] From the above experiments, it can be seen that SIRNet not only outperforms a variety of mainstream neural network models in overall accuracy, but also exhibits significant advantages in convergence speed, stability and class discrimination ability. Especially with the introduction of the MS-TFCA module, the model's feature representation ability and generalization performance under single sensor conditions are effectively improved, providing a practical solution for composite material low-speed impact monitoring.
[0076] 3. Conclusion.
[0077] The present application proposes a deep residual convolutional neural network framework based on single sensor and multi-scale time-frequency interaction attention module MS-TFCA for automatic identification of impact energy level, aiming at the low-speed impact damage problem that composite material structures may suffer during service. Through experiments and finite element simulation verification, the present application method shows significant advantages in accuracy and generalization. The main contributions can be summarized as follows:
[0078] (1) A single-sensor low-velocity impact energy level identification framework is proposed, which breaks through the limitation of traditional multi-sensor layout. Only using the vibration signals collected by a single piezoelectric ceramic sensor, the impact energy can be effectively classified, and the application potential of the method in resource-limited platforms is significantly improved.
[0079] (2) A multi-scale time-frequency interactive attention module MS-TFCA is designed. By realizing interactive feature modeling in the time and frequency dimensions, and introducing a multi-scale convolution mechanism, the joint representation ability of transient and long-term dynamic features is improved.
[0080] (3) Experiments and simulations are carried out for verification: based on the public thin plate impact data set and the CFRP plate data generated by Abaqus simulation, training and verification are carried out respectively. The results show that the proposed framework not only achieves high-precision classification on experimental data, but also maintains good performance on simulation data, verifying its effectiveness and generalization ability.
[0081] In summary, the single-sensor MS-TFCA residual convolutional neural network proposed in the present application ensures high precision while considering light weight and robustness, providing a new method with application potential for composite low-velocity impact monitoring.
[0082] Based on the same inventive concept, the present application provides a composite low-velocity impact energy level identification system, comprising: A data acquisition module is used to acquire the two-dimensional time-frequency spectrum corresponding to the transient vibration signal of the composite material under the action of low-velocity impact obtained by a single piezoelectric ceramic sensor.
[0083] A model construction module is used to input the two-dimensional time-frequency spectrum into a pre-trained impact energy level identification model. The pre-trained impact energy level identification model includes: a feature extraction module, a residual convolutional backbone network and a classification and identification module connected in turn. The residual convolutional backbone network includes: a plurality of residual blocks connected in turn and each residual block embedded with a multi-scale time-frequency interactive attention module. The multi-scale time-frequency interactive attention module includes: a multi-scale convolution module, a global pooling and convolution module and an interactive fusion mechanism.
[0084] The model processing module is configured to perform a convolution operation on the two-dimensional time-frequency spectrum by the feature extraction module to obtain a local time-domain feature. The local time-domain feature is subjected to parallel convolution of different scales by the multi-scale convolution module to extract short-time transient features and long-time trend features in the time dimension, thereby obtaining multi-scale time features. The local time-domain feature is subjected to global average pooling and convolution operation along the time axis by the global pooling and convolution module to generate adaptive weights representing the importance of the composite material in different frequency bands in the frequency dimension, thereby obtaining frequency attention weights. The multi-scale time features and the frequency attention weights are interactively fused by the interactive fusion mechanism to obtain a two-dimensional attention map. The local time-domain feature is weighted by the two-dimensional attention map to determine non-stationary features of the low-speed impact action, thereby obtaining an output feature map. The output feature map is subjected to classification prediction by the classification and recognition module to obtain an impact energy level classification result.
[0085] The above-described embodiments only express several embodiments of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are within the scope of protection of the present application.
Claims
1. A method for identifying low-velocity impact energy levels in composite materials, characterized in that, include: Two-dimensional time spectrum of transient vibration signal of composite material under low-speed impact is obtained by a single piezoelectric ceramic sensor; The two-dimensional time-frequency spectrum is input into a pre-trained impact energy level recognition model. The pre-trained impact energy level recognition model includes: a feature extraction module, a residual convolutional backbone network, and a classification and recognition module connected in sequence. The residual convolutional backbone network includes: multiple residual blocks connected in sequence, and each residual block embeds a multi-scale time-frequency interactive attention module. The multi-scale time-frequency interactive attention module includes: a multi-scale convolution module, a global pooling and convolution module, and an interactive fusion module. The feature extraction module performs convolution operations on the two-dimensional time spectrum to obtain local time-domain features. A multi-scale convolution module performs parallel convolutions at different scales on these local time-domain features to extract short-term transient features and long-term trend features in the time dimension, resulting in multi-scale time features. A global pooling and convolution module performs global average pooling and convolution operations along the time axis on the local time-domain features to generate adaptive weights representing the importance of the composite material in different frequency bands in the frequency dimension, resulting in frequency attention weights. An interactive fusion module interoperates and fuses the multi-scale time features and frequency attention weights to obtain a two-dimensional attention map. The two-dimensional attention map is then used to weight the local time-domain features to determine the non-stationary characteristics of low-velocity impact, resulting in an output feature map. Finally, a classification and recognition module performs classification prediction on the output feature map to obtain the impact energy level classification result.
2. The method for identifying low-velocity impact energy levels of composite materials as described in claim 1, characterized in that, The acquisition of the two-dimensional time spectrum corresponding to the transient vibration signal of the composite material under low-speed impact by a single piezoelectric ceramic sensor specifically includes: The transient vibration signals of the composite material under low-speed impact are acquired by a single piezoelectric ceramic sensor, and the transient vibration signals are arranged according to the acquisition time sequence to obtain the original time sequence data; Based on the following formula, the original time-series data is transformed from transient vibration signals into a time-spectrum representation using short-time Fourier transform: ； in, The time-frequency spectrum is represented as follows. For impact response signal, For The window function is centered. For frequency variables; Based on the following formula, taking the complex modulus of the time-frequency spectrum representation and performing logarithmic normalization, we obtain a two-dimensional time-frequency spectrum with time and frequency as coordinates and amplitude as intensity information: ； in, It is a two-dimensional time spectrum.
3. The method and system for identifying low-velocity impact energy levels of composite materials as described in claim 1, characterized in that, The step of performing a convolution operation on the two-dimensional time spectrum through the feature extraction module to obtain local time-domain features specifically includes: The feature extraction module includes: a convolutional layer, a batch normalization layer, and a non-linear activation function layer connected in sequence; The initial feature map is obtained by performing convolution operations on the two-dimensional time spectrum through convolutional layers; The initial feature map is normalized by a batch normalization layer, and the normalized feature map is then nonlinearly transformed by a nonlinear activation function layer to obtain local temporal features.
4. The method and system for identifying low-velocity impact energy levels of composite materials as described in claim 1, characterized in that, Based on the following formula, multi-scale temporal features are obtained by performing parallel convolutions at different scales on local temporal features using a multi-scale convolution module: ； in, For each time step, the corresponding time feature is... Indicates the kernel size as One-dimensional convolution operation, For the scale number, It is a local time-domain feature; Based on the following formula, global pooling and convolution modules are used to perform global average pooling on local temporal features along the time axis to obtain an overall description of different frequency bands: ； in, For each frequency band, a comprehensive description is provided. Where T is the frequency and T is the time axis. t This refers to the time step in the timeline; Based on the following formula, a nonlinear mapping of the overall description of different frequency bands is performed through a one-dimensional convolution operation, and the frequency attention weights of each frequency band are determined by the sigmoid activation function: ； in, Frequency attention weights for each frequency band, It is the sigmoid activation function. This is a one-dimensional convolution operation; Based on the following formula, a two-dimensional attention map is obtained by interactively fusing multi-scale temporal features and frequency attention weights through an interactive fusion mechanism: ； in, It is a two-dimensional attention map. For at any time Time feature vector, For frequency The corresponding frequency attention weights.
5. The method and system for identifying low-velocity impact energy levels of composite materials as described in claim 4, characterized in that, The output feature map is obtained by weighting the local temporal features using a two-dimensional attention map based on the following formula: ； in, To output the feature map, For the channel, For a moment, For frequency; The impact energy level classification result is obtained by classifying and predicting the output feature map using the following formula through the classification and recognition module: ； in, For category The weight vector, For category The bias.
6. The method and system for identifying low-velocity impact energy levels of composite materials as described in claim 1, characterized in that, Training the impact energy level identification model specifically includes: Obtain historical two-dimensional time-frequency spectra and their corresponding historical impact energy level classification results; Input the historical two-dimensional time spectrum into the impact energy level identification model to obtain the predicted impact energy level classification result; The weighted cross-entropy loss between historical and predicted impact energy level classification results is determined based on the following formula: ； in, For category weight coefficients, One-hot encoding of the actual label; The impact energy level identification model is trained with the goal of minimizing the weighted cross-entropy loss, resulting in the trained impact energy level identification model.
7. The method and system for identifying low-velocity impact energy levels of composite materials as described in claim 6, characterized in that, Before inputting the historical two-dimensional time spectrum into the impulse energy level identification model, the data enhancement preprocessing of the two-dimensional time spectrum is also included. The data enhancement includes randomly masking local areas in the time or frequency dimension to simulate signal loss and noise interference in actual applications.
8. A system for identifying low-velocity impact energy levels in composite materials, characterized in that, include: The data acquisition module is used to acquire the two-dimensional time spectrum corresponding to the transient vibration signal of the composite material under low-speed impact by a single piezoelectric ceramic sensor. The model building module is used to input the two-dimensional time-frequency spectrum into the pre-trained impact energy level recognition model. The pre-trained impact energy level recognition model includes: a feature extraction module, a residual convolutional backbone network, and a classification and recognition module connected in sequence. The residual convolutional backbone network includes: multiple residual blocks connected in sequence, and each residual block embeds a multi-scale time-frequency interactive attention module. The multi-scale time-frequency interactive attention module includes: a multi-scale convolution module, a global pooling and convolution module, and an interactive fusion mechanism. The model processing module performs convolution operations on the two-dimensional time-domain spectrum through the feature extraction module to obtain local time-domain features; it then performs parallel convolutions at different scales on the local time-domain features through a multi-scale convolution module to extract short-term transient features and long-term trend features in the time dimension, resulting in multi-scale time features; finally, it performs global average pooling and convolution operations on the local time-domain features along the time axis through a global pooling and convolution module to generate adaptive weights representing the importance of composite materials in different frequency bands in the frequency dimension, resulting in frequency attention weights; an interactive fusion mechanism is used to interactively fuse the multi-scale time features and frequency attention weights to obtain a two-dimensional attention map; the two-dimensional attention map is then used to weight the local time-domain features to determine the non-stationary features of low-velocity impact, resulting in an output feature map; and finally, a classification and recognition module performs classification prediction on the output feature map to obtain the impact energy level classification result.