A neural network-based method for localizing damage in multilayer heterogeneous guided waves.

Abstract

This invention relates to the field of multilayer heterostructure analysis technology, and provides a neural network-based method for guiding wave damage localization in multilayer heterostructures. First, a finite element simulation module acquires simulated time-domain signal data, and a data processing module calculates a broadband damage index. Next, an integrated neural network model receives the time-domain signal and the broadband damage index as dual inputs. Within this model, a convolutional neural network branch extracts local signal patterns, a long short-term memory network branch models temporal information, and an autoencoder branch performs noise reduction and latent feature extraction. Subsequently, a feature fusion layer and an attention mechanism module merge and weight the outputs of each branch, and a fully connected network outputs the planar coordinates of the defect and its layer number. Finally, a damage probability reconstruction module generates the final defect imaging map based on the output results, achieving precise localization of deep defects in multilayer heterostructures.

Description

Technical Field

[0001] This invention relates to the field of multilayer heterostructure analysis technology, and more specifically, to a method for locating waveguide damage in multilayer heterostructures based on neural networks. Background Technology

[0002] With the continuous development of the global economy and the constant progress of technology, high-end equipment manufacturing, new materials, and intelligent manufacturing have become important focal points of global competition. The rapid development of these fields has not only driven global technological progress but also provided crucial support for national economic restructuring and green sustainable development. As the core of modern industry, high-end equipment manufacturing plays a vital role, particularly in aerospace, shipbuilding, nuclear industry, and energy. Structural health monitoring and damage detection in these industries remain key technologies for ensuring their long-term safety, stability, and economic viability.

[0003] In recent years, with technological advancements, traditional damage detection methods (such as X-ray inspection, eddy current testing, and thermal imaging) have gradually revealed numerous limitations. Especially when dealing with multilayered heterogeneous material structures, traditional non-destructive testing techniques face significant difficulties in identifying deep defects, locating complex damage, and improving accuracy. For example, the application of traditional ultrasonic testing methods in multilayered heterogeneous structures is affected by factors such as material impedance mismatch, complex waveguide modes, and curved wave propagation paths, greatly increasing the difficulty of detecting deep defects and failing to meet the requirements for high precision and efficiency. These problems not only limit the application scope of existing technologies but also increase the cost and risk of engineering maintenance.

[0004] With the development of new-generation information technology, especially the widespread application of artificial intelligence (AI), machine learning (ML), and deep learning (DL), traditional ultrasonic damage detection methods have ushered in a technological revolution. Deep learning models, particularly neural network architectures such as convolutional neural networks (CNN), long short-term memory networks (LSTM), and autoencoders, can effectively process the complex characteristics of ultrasonic guided wave signals, helping to achieve higher-precision damage localization and imaging. This neural network-based damage detection method not only significantly improves detection accuracy but also enables rapid and accurate identification of minute or deep structural damage in complex multilayer heterogeneous structures, demonstrating significant technological advantages.

[0005] In the fields of new energy and intelligent manufacturing, the application of new materials is increasing, especially in new energy vehicles and green energy technologies. The use of new materials presents greater challenges to traditional non-destructive testing (NDT) techniques. For example, the high-precision testing requirements of battery management systems, power battery modules, and electric drive systems in new energy vehicles cannot be met by traditional ultrasonic testing techniques, which demand high accuracy and efficiency. However, adopting AI-based ultrasonic guided wave damage localization technology can effectively improve testing efficiency, reduce testing costs, and provide strong support for ensuring the safety of new energy vehicles. Summary of the Invention

[0006] To overcome the aforementioned shortcomings of existing technologies, this invention provides a neural network-based method for locating guided wave damage in multilayer heterogeneous structures. It constructs a parallel neural network integrating convolutional neural networks, long short-term memory networks, and autoencoders, and introduces a feature fusion layer and an attention mechanism. This method aims to address the technical problem in the prior art where traditional ultrasound struggles to accurately locate deep defects in multilayer heterogeneous structures due to material impedance mismatch and complex guided wave modes. By using a deep learning model to extract the spatiotemporal and latent features of the signal in parallel, it can effectively handle complex wave propagation characteristics, thereby establishing a mapping between the signal and the three-dimensional coordinates of the defect, achieving precise location.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for localizing damage in multilayer heterogeneous guided waves based on neural networks includes the following steps:

[0009] Step 1: Use finite element simulation technology to simulate the guided wave propagation process of the multilayer heterostructure under different damage states, and obtain simulated time-domain signal data containing information on damage location and layer number. Use the simulated time-domain signal data as a training set. Based on the energy of the direct wave in the time-domain signal data, calculate the broadband damage index that characterizes the difference between the reference signal and the detection signal.

[0010] Step two: The time-domain signal data and the broadband impairment index are input into an integrated neural network model consisting of a convolutional neural network branch, a long short-term memory network branch, and an autoencoder branch set up in parallel. The convolutional neural network branch is used to extract local spatial features from the time-domain signal data; the long short-term memory network branch is used to model the temporal features of the time-domain signal data to capture propagation delay features between different layers; and the autoencoder branch is used to denoise the time-domain signal data and extract its latent features.

[0011] Step 3: The features extracted from the three branches are fused in the feature fusion layer, and the fused features are weighted by the attention mechanism module to improve the sensitivity to the damaged area. The weighted features are then input into a multi-layer fully connected network for nonlinear mapping to establish the mapping relationship between the time-domain signal data, the broadband damage index, the defect location, and the layer number. Finally, the planar coordinates x and y of the defect on the board and the layer number N are output.

[0012] Step four: Combine the damage probability reconstruction algorithm to process the output planar coordinates x and y and the layer number N to generate an accurate imaging map of the defect.

[0013] As a further aspect of the present invention, the finite element simulation technology simulates a multilayer heterogeneous structure, including establishing a three-layer structure consisting of a copper layer, a steel layer, and an aluminum layer, with epoxy resin used to bond the layers together, and arranging a circular sensor array consisting of sixteen sensors on the structure for excitation and reception of ultrasonic signals.

[0014] As a further aspect of the present invention, the method for calculating the broadband impairment index includes: using a linear frequency modulated signal with a frequency range of 200 kHz to 400 kHz as the excitation signal, dividing the received signal into multiple sub-bands according to frequency, and fusing them into the broadband impairment index, wherein the impairment index value of each sub-band is the energy difference obtained by subtracting the energy of the reference signal from the energy of the defective signal in the sub-band, divided by the maximum value of the energy of the reference signal in all sub-bands.

[0015] As a further embodiment of the present invention, the integrated neural network model comprises a data input layer, a convolutional neural network branch, a long short-term memory network branch, an autoencoder branch, a feature fusion layer, an attention mechanism module, a multi-layer fully connected layer, and an output layer.

[0016] As a further aspect of the present invention, the convolutional neural network branch consists of multiple convolutional layers and pooling layers, used to automatically extract spatial features from the time-domain signal data and perform signal pattern recognition.

[0017] As a further aspect of the present invention, the Long Short-Term Memory (LSTM) network branch models the temporal characteristics of the time-domain signal data to determine the time delay characteristics and fluctuation pattern changes during signal propagation.

[0018] As a further aspect of the present invention, the autoencoder branch employs an unsupervised learning method to learn and extract latent features from the time-domain signal data in order to eliminate noise and enhance key information in the signal.

[0019] As a further aspect of the present invention, the plane coordinate x of the defect on the plate is the abscissa of the position of the defect center point on the plate, and the plane coordinate y is the ordinate of the position of the defect center point on the plate.

[0020] As a further aspect of the present invention, the layer number N where the defect is located is the layer number counted downwards from the surface of the structure.

[0021] As a further aspect of the present invention, the damage probability reconstruction algorithm locates and images defects by constructing a damage probability distribution function, wherein the damage probability distribution function is generated by superimposing elliptical probability distribution functions of all sensing paths, and each elliptical probability distribution function has its corresponding sensing path as its main axis.

[0022] Compared with existing technologies, the beneficial effects of the present invention's method for localizing damage in multilayer heterogeneous guided waves based on neural networks are as follows:

[0023] This invention constructs a deep learning model integrating three parallel branches: a convolutional neural network, a long short-term memory network, and an autoencoder, and optimizes it through a feature fusion layer and an attention mechanism. Compared to traditional ultrasonic detection methods, this multi-dimensional, parallel feature extraction approach can deeply mine effective damage-related information from complex signals that are difficult to analyze using traditional methods, overcoming the information recognition challenges caused by complex waveguide modes and curved propagation paths in multi-layered heterogeneous structures. This lays a solid foundation for precise localization.

[0024] The direct output of the neural network model in this invention not only includes the planar coordinates of the defect but also explicitly indicates the layer number where the defect is located. Compared to traditional nondestructive testing techniques, which generally suffer from significant limitations in identifying and locating deep defects when dealing with multilayer heterogeneous material structures, traditional methods often fail to accurately determine the specific layer at which the damage occurred. This method addresses this technical challenge by establishing an end-to-end mapping relationship between signal features and the three-dimensional location of the defect, achieving a leap from two-dimensional fuzzy detection to precise three-dimensional localization, thus improving the accuracy and practical value of the detection results.

[0025] This invention employs an improved damage index calculation method based on broadband excitation, using this index and the original time-domain signal as dual inputs to a neural network. Traditional guided wave detection typically uses single-frequency excitation, which has limited sensitivity to defects of varying depths. For example, while high-frequency signals offer good spatial resolution, their energy decays rapidly, making it difficult to detect deep defects; conversely, low-frequency signals have strong penetrating power but insufficient resolution. This invention utilizes a broadband chirp signal and analyzes the energy response of each sub-band to construct a damage index that comprehensively reflects information from multiple frequency bands. By feeding this feature, rich in depth information, along with the original signal into the model, the model's comprehensive ability to assess structural damage is enhanced, further improving the overall detection accuracy and robustness. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a method for locating damage in a multilayer heterogeneous guided wave structure based on a neural network, according to the present invention.

[0027] Figure 2 This is a schematic diagram of the finite element model of a multilayer heterogeneous structure guided wave damage localization method based on neural networks according to the present invention.

[0028] Figure 3 This is a schematic diagram of the integrated neural network model of the multilayer heterogeneous structure guided wave damage localization method based on neural networks according to the present invention.

[0029] Figure 4 This is a schematic diagram illustrating the different layer defect settings of a multilayer heterogeneous structure guided wave damage localization method based on neural networks according to the present invention. Detailed Implementation

[0030] The technical solutions of this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0031] Example 1

[0032] The first step in this embodiment of the invention is data acquisition and preprocessing, the core of which is constructing a high-precision finite element model and processing the simulation data. In this embodiment, Abaqus / EXPLICIT software is used to build a finite element model of a multilayer bonded structure. Figure 2As shown, the model is specifically a three-layer structure consisting of a copper layer (400×400×1mm³, density: 8960kg / m³, Young's modulus: 119GPa, Poisson's ratio: 0.32), a steel layer (400×400×1mm³, density: 7932kg / m³, Young's modulus: 210GPa, Poisson's ratio: 0.29), and an aluminum layer (400×400×2mm³, density: 2700kg / m³, Young's modulus: 70GPa, Poisson's ratio: 0.33). The layers are bonded together with epoxy resin (400×400×0.1mm³, density: 1170 kg / m³, Young's modulus: 1GPa, Poisson's ratio: 0.38). Material nonlinearity and motion nonlinearity are not considered in the simulation. To simulate damage, a circular defect element with a diameter of 10 mm is set in the finite element model, and the depth of the element is set to half the thickness of the layer it is in.

[0033] After constructing the model, each layer needs to be meshed. The choice of mesh size should aim to improve computational efficiency while ensuring computational convergence. Mesh sizes are typically structured. If the mesh size is too large, it will cause significant deviations in the simulation results; conversely, if the mesh size is too small, it will greatly increase the computation time. Therefore, to ensure the accuracy and efficiency of the calculation results, the mesh size must be chosen such that at least 10 wavelengths can propagate within a single mesh; that is, the maximum value of the length and width of the set mesh element must be less than or equal to one-tenth of the minimum wavelength of the ultrasonic Lamb wave. After comprehensive consideration, this embodiment of the invention selects a mesh size of 0.5 mm and sets the mesh type to CPE4R.

[0034] The excitation signal in the simulation is a 5-cycle sinusoidal pulse signal modulated by a Hanning window, the displacement signal amplitude is 1e-6m, and the excitation frequency is set to 300kHz.

[0035] The analysis step mode must be set to explicit dynamics. The choice of analysis step time is equally crucial; too long a time may lead to non-convergence, while too short a time will unnecessarily increase computational load. To ensure the scientific validity and accuracy of the simulation model, the analysis step time must satisfy a specific formula: the analysis step time must be less than or equal to the quotient obtained by dividing the maximum element size in the model's mesh by the maximum ultrasonic wave velocity in the structural component, multiplied by a coefficient of 0.8. Based on this formula, and using the transverse wave velocity in steel of 3215 m / s and the global mesh size of 0.5 mm, the analysis step time should not exceed 1.24e-7 s; therefore, this embodiment selects 1e-7 s.

[0036] During data acquisition, a circular array of sixteen sensors was arranged on the front of the model, centered at the model's center and with a radius of 125 mm, with one sensor positioned at 22.5° intervals. Displacement signals along all sensing paths were acquired by sequentially applying excitation to each sensor and having the remaining fifteen sensors receive the signals.

[0037] After obtaining the simulated time-domain signal data, the broadband damage index (DI) needs to be calculated as another input to the neural network. The data processing flow first includes two parts: damage index calculation and an improved method based on broadband excitation. This embodiment of the invention adopts the basic principle of the RAPID (Reconstruction Algorithm for Probabilistic Inspection of Damage) damage probability imaging method. It constructs the damage index (DI) by analyzing the difference between sensor signals in the undamaged and damaged states of the structure, thereby reflecting the probability of defect existence. The specific steps are as follows: when the structure is in a healthy state, a reference signal is acquired for each pair of excitation-reception sensor paths; after the structure may be damaged, the detection signals on the same path are acquired again; the DI value is calculated by comparing the two sets of signals. The DI value typically ranges from 0 to 1, with a larger value indicating a higher probability of damage in that path. To quantify this difference, an energy ratio-based DI calculation method is adopted. Specifically, the energy of the defect signal within a specified time interval is subtracted from the energy of the reference signal within the same interval, and then the difference is divided by the energy of the reference signal within that interval. To further enhance the sensitivity of guided wave detection to defects at different layers, this embodiment of the invention introduces a wideband excitation strategy, using a linear frequency modulated (Chirp) signal with a frequency range of 200-400kHz as the excitation signal. Leveraging its wide frequency coverage and rich mode excitation, it enhances the ability to acquire multimodal responses. The received signal is divided into multiple sub-bands according to frequency and subjected to bandpass filtering. The DI value of each sub-band is then calculated. In detecting blind holes on the bottom surface of multilayer structures, studies have shown that low-frequency bands (e.g., 200–250kHz) have strong penetration capabilities and higher DI values, accurately reflecting the defect location. While high-frequency bands have better spatial resolution, they are susceptible to energy attenuation, limiting their effectiveness in identifying deep defects. Therefore, in actual imaging, a multi-band DI fusion strategy can be adopted, integrating DI information from multiple bands to construct a unified damage probability map, thereby improving overall detection accuracy and image clarity. According to the improved broadband DI definition, its value is the energy difference between the defect signal and the reference signal in the corresponding frequency band, divided by the maximum energy of the reference signal across all frequency bands. For model training, the acquired simulation data is ultimately divided into a training set and a test set, with a ratio of 80% and 20%, respectively, and an additional 20% of the training set is used as a validation set. Simultaneously, to improve model accuracy and generalization ability, all data is standardized to transform it into dimensionless values ​​conforming to an N(0,1) normal distribution, and the order of the data samples is shuffled to prevent overfitting.

[0038] Subsequently, an integrated neural network model is trained and used for prediction. The network model receives the original guided wave time-domain signal and its corresponding broadband impairment index as dual inputs. For example... Figure 3 As shown, this integrated neural network model mainly consists of a data input layer, a parallel feature extraction branch, a feature fusion layer, an attention mechanism module, multiple fully connected layers, and an output layer. The Convolutional Neural Network (CNN) branch is used to identify local patterns in the guided wave signal; the Long Short-Term Memory (LSTM) branch is used to model the signal propagation delay characteristics; and the autoencoder branch is used to denoise the signal and extract its potential nonlinear features. The features extracted by each branch are merged in the feature fusion layer, and the introduced attention mechanism module then weights key features to improve sensitivity to the damaged area. The fused and weighted features are nonlinearly mapped through a multiple fully connected network, ultimately directly outputting the spatial coordinates (x, y) of the damage and the layer number (N). During model training, RMSProp and the Adam optimizer are used to adjust the network parameters, with a learning rate set to 0.001 and mean squared error (MSE) as the loss function, enabling the network to accurately fit the relationship between the ultrasonic guided wave signal and the damage location and layer number.

[0039] Finally, damage probability imaging is performed. After the neural network model outputs precise damage coordinates (x, y) and layer number (N), the RAPID damage probability imaging method is used to visualize the results. This method is achieved by constructing a final damage probability distribution function. Specifically, based on each sensing path, an elliptical probability distribution function is constructed with that path as the principal axis. Then, the distribution functions of all paths are superimposed over the target region to form the final damage probability image. In the final generated image, the areas with high probability values ​​are the potential defect concentration areas, thus achieving accurate localization and intuitive imaging of defects in multilayer heterogeneous structures.

[0040] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0041] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for locating damage in multilayer heterogeneous guided waves based on neural networks, characterized in that, Includes the following steps: Step 1: Use finite element simulation technology to simulate the guided wave propagation process of the multilayer heterostructure under different damage states, and obtain simulated time-domain signal data containing information on damage location and layer number. Use the simulated time-domain signal data as a training set. Based on the energy of the direct wave in the time-domain signal data, calculate the broadband damage index that characterizes the difference between the reference signal and the detection signal. Step two: The time-domain signal data and the broadband impairment index are input into an integrated neural network model consisting of a convolutional neural network branch, a long short-term memory network branch, and an autoencoder branch set up in parallel. The convolutional neural network branch is used to extract local spatial features from the time-domain signal data; the long short-term memory network branch is used to model the temporal features of the time-domain signal data to capture propagation delay features between different layers; and the autoencoder branch is used to denoise the time-domain signal data and extract its latent features. Step 3: The features extracted from the three branches are fused in the feature fusion layer, and the fused features are weighted by the attention mechanism module. The weighted features are then input into a multi-layer fully connected network for nonlinear mapping to establish the mapping relationship between the time-domain signal data, the broadband damage index, the defect location, and the layer number. Finally, the planar coordinates x and y of the defect on the board and the layer number N are output. Step four: Combine the damage probability reconstruction algorithm to process the output planar coordinates x and y and the layer number N to generate an accurate imaging map of the defect.

2. The method for locating damage in a multilayer heterogeneous guided wave structure based on a neural network according to claim 1, characterized in that, The finite element simulation technology simulates a multilayer heterogeneous structure, including the establishment of a three-layer structure consisting of a copper layer, a steel layer, and an aluminum layer, with epoxy resin used to bond the layers together, and a circular sensor array consisting of sixteen sensors arranged on the structure for excitation and reception of ultrasonic signals.

3. The method for locating damage in a multilayer heterogeneous guided wave structure based on a neural network according to claim 1, characterized in that, The method for calculating the broadband impairment index includes: using a linear frequency modulated signal with a frequency range of 200 kHz to 400 kHz as the excitation signal, dividing the received signal into multiple sub-bands according to frequency, and fusing them into the broadband impairment index. The impairment index value of each sub-band is the energy difference obtained by subtracting the energy of the reference signal from the energy of the defective signal in the sub-band, divided by the maximum value of the reference signal energy in all sub-bands.

4. The method for locating damage in a multilayer heterogeneous guided wave structure based on a neural network according to claim 1, characterized in that, The integrated neural network model consists of a data input layer, a convolutional neural network branch, a long short-term memory network branch, an autoencoder branch, a feature fusion layer, an attention mechanism module, a multi-layer fully connected layer, and an output layer.

5. The method for locating damage in a multilayer heterogeneous guided wave structure based on a neural network according to claim 1, characterized in that, The convolutional neural network branch consists of multiple convolutional layers and pooling layers, used to automatically extract spatial features from the time-domain signal data and perform signal pattern recognition.

6. The method for locating damage in a multilayer heterogeneous guided wave structure based on a neural network according to claim 1, characterized in that, The Long Short-Term Memory (LSTM) network branch models the temporal characteristics of the time-domain signal data to determine the time delay characteristics and fluctuation pattern changes during signal propagation.

7. The method for locating damage in a multilayer heterogeneous guided wave structure based on a neural network according to claim 1, characterized in that, The autoencoder branch employs an unsupervised learning approach to learn and extract latent features from the time-domain signal data, thereby eliminating noise and enhancing key information in the signal.

8. The method for locating damage in a multilayer heterogeneous guided wave structure based on a neural network according to claim 1, characterized in that, The x-coordinate of the defect on the plate is the horizontal coordinate of the defect center point on the plate, and the y-coordinate is the vertical coordinate of the defect center point on the plate.

9. The method for locating damage in a multilayer heterogeneous guided wave structure based on a neural network according to claim 1, characterized in that, The layer number N where the defect is located is the layer number counted downwards from the surface of the structure.

10. The method for locating damage in a multilayer heterogeneous guided wave structure based on a neural network according to claim 1, characterized in that, The damage probability reconstruction algorithm locates and images defects by constructing a damage probability distribution function. The damage probability distribution function is generated by superimposing elliptical probability distribution functions of all sensing paths, and each elliptical probability distribution function has its corresponding sensing path as its main axis.

Images