Collision load reconstruction method and system

Abstract

The invention discloses a collision load reconstruction method and system, relates to the technical field of load identification, and aims to solve the problem that the prediction result of the current load reconstruction method is not accurate enough, and the method comprises the following steps: S1, collecting an excitation signal and a response signal when a vehicle body is laterally impacted, performing multi-mode decomposition on the excitation signal and the response signal by using a variational mode decomposition method; s2, according to Spearman correlation, selecting a modal component with the maximum correlation as an optimal modal signal base to perform noise adding processing; s3, establishing a noise feature library, generating composite noise conforming to actual working conditions, and expanding a training data set on the premise of retaining effective features; s4, constructing a prediction model of deep learning, performing model training by using a dual-loss function, and generating a reconstructed load signal by using the prediction model, and further comprising a corresponding system; the method and the system are suitable for vehicle body lateral collision load reconstruction of the sparse sensor array, and the precision of load reconstruction is high.

Description

Technical Field

[0001] The present invention relates to the technical field of load identification, and particularly relates to a method and a system for reconstructing collision loads. Background Art

[0002] As one of the structural health monitoring technologies, load reconstruction has always been an important research direction. The mechanical structure, which is the main body of mechanical equipment, is often damaged due to stress concentration, fatigue loads, sudden impacts, and harsh operating environments, resulting in a reduction in the performance and safety of mechanical equipment. In the engineering field, the analysis of the relationship between excitation signals and response signals is of great significance for structural health monitoring, vibration control, etc.

[0003] At present, the maintenance of mechanical equipment such as aerospace aircraft structures usually adopts a regular maintenance method, and it is difficult to detect impact damage in a timely manner. Therefore, there is an urgent need for on-board online real-time monitoring of impact events on mechanical structures. The external dynamic load information of a structure plays an important role in engineering fields such as the overall design of the structure, strength verification, and environmental prediction. In actual engineering, due to factors such as the shape design, sensor installation layout, and external environment, the external dynamic load information received by the structure generally cannot be directly measured by installing force sensors. However, the vibration response of the structure at certain positions can often be directly measured, and the modal information of the structure can also be obtained through finite element simulation or modal tests. Therefore, it has become an effective way to invert the external dynamic load information through the vibration response and the structural modal information.

[0004] In traditional methods, the common problem is that the solution process is ill-posed and prone to non-convergent solutions. In addition, many systems are non-linear under actual engineering conditions, and there is also a non-linear connection between the input load and the corresponding strain. When dealing with the prediction of excitation signals, existing methods often ignore the modal decomposition and noise processing of the signals, resulting in inaccurate prediction results.

[0005] Chinese Patent with Publication No. CN116011244A, although it involves a load identification method, its technical path is significantly different from that of the present application. Summary of the Invention

[0006] The present invention solves the problem that the prediction results of the current load reconstruction method are not accurate enough, and proposes a method and a system for reconstructing collision loads to improve the recognition accuracy of load prediction.

[0007] To achieve the above object, the present invention adopts the following technical solutions: A method for reconstructing collision loads, comprising the following steps: S1, collect the excitation signal and the response signal when an impact occurs on the side of the vehicle body, and perform multi-modal decomposition on the excitation signal and the response signal using the variational mode decomposition method; S2. Select the modal component with the largest correlation as the optimal modal signal basis according to Spearman correlation for noise addition processing; S3. Establish a noise feature library, generate composite noise that conforms to the actual working conditions, and expand the training data set while retaining the effective features; S4. Construct a prediction model of deep learning, use a double loss function for model training, and use the prediction model to generate a reconstructed load signal.

[0008] In this technical solution, first, an excitation signal and a corresponding signal during impact are collected by a data acquisition instrument, and the variational mode decomposition method is used to decompose and obtain several modal components; the modal component with the largest correlation is selected as the optimal modal signal basis for noise addition processing; then a noise feature library is constructed, and through dynamic noise injection, composite noise that conforms to the actual working conditions is generated; finally, a prediction model based on deep learning is constructed, and through this prediction model, a reconstructed load signal can be generated; the method of the present invention can improve the recognition accuracy of load prediction.

[0009] The present invention is further set as follows: The step S1 includes: S11. Select the lateral bending area of the vehicle body as the structural impact measurement area, perform standardization processing on the collected excitation signal and response signal, remove outliers, and eliminate the trend term; S12. Perform multi-scale decomposition on the signal, and use the particle swarm optimization algorithm to optimize the penalty factor and the number of modal layers in the variational mode decomposition algorithm to obtain several modal components.

[0010] In the step S1 of this technical solution, mainly perform standardization processing and multi-modal decomposition on the collected excitation signal and response signal, and optimize the penalty factor and the number of modal layers in the variational mode decomposition algorithm according to the particle swarm optimization algorithm to reduce the modal aliasing phenomenon.

[0011] The present invention is further set as follows: The step S2 specifically includes: Calculate the Spearman correlation coefficient between each modal component and the original signal; screen out the modal with the largest correlation with the original signal as the signal basis.

[0012] In this technical solution, the modal component with the largest correlation is selected by using Spearman correlation.

[0013] The present invention is further set as follows: The step S3 includes: S31. Perform time-frequency analysis on the noise-added signal, calculate the time-frequency matrix through the Wigner distribution, introduce the Hamming window function, and calculate the time-frequency signal-to-noise ratio of the original signal; S32. Construct a noise feature library, dynamically inject noise to generate composite noise that conforms to the actual working conditions, establish an adaptive noise addition mechanism based on the dynamic signal-to-noise ratio, and dynamically adjust the mixing noise ratio according to the time-frequency characteristics of the signal.

[0014] In this technical solution, in order to ensure the time-frequency synchronization between the noise-added signal and the original signal, perform time-frequency analysis on the signal, calculate the time-frequency matrix through the Wigner distribution, introduce the Hamming window function to suppress cross-term interference, and at the same time maintain the time-frequency resolution of the signal.

[0015] The present invention is further set as: The step S3 further includes: S33. Dynamically adjust the noise energy according to the time-frequency signal-to-noise ratio of the original signal; S34. Define the signal-to-noise ratio threshold. In the low signal-to-noise ratio region, when the signal-to-noise ratio is less than the lowest threshold, reduce the noise injection; in the high signal-to-noise ratio region, when the signal-to-noise ratio is less than the lowest threshold, increase the noise injection; S35. Reconstruct the time-domain signal after noise addition through the inverse Wigner distribution.

[0016] The present invention is further set as: The step S4 includes: S41. Divide the signal into time series windows, construct a time series data set, use the response signal obtained in S3 as the input, and the excitation signal as the output; S42. Adopt a cyclic structure to perform multi-scale feature extraction and feature fusion on the depthwise separable convolution module, Transformer module, and fusion attention mechanism module; S43. Use a dual loss function for training; S44. Use the early stopping method and learning rate decay strategy to optimize the training process.

[0017] In this technical solution, the prediction model includes a depthwise separable convolution module, a Transformer module, and a fusion attention mechanism module. The structural response of the training data obtained in S3 is used as the input, and the impact load time history is used as the output to supervise the training of the deep learning prediction model, and a dual loss function is used to improve the performance of the model.

[0018] The present invention is further set as: It further includes step S5, calculating the average relative error and peak relative error between the reconstructed load signal and the actual result, and evaluating the performance of the prediction model.

[0019] In this technical solution, the average relative error MRE and peak relative error PRE are calculated respectively, and then the performance of the model is evaluated to facilitate subsequent improvement.

[0020] The present invention is further set as: The step S42 includes: S421. Introduce a depthwise separable convolution module to extract local features of the signal; S422. Introduce a Transformer module to capture various long-term dependencies in the signal and enhance the long-term memory ability; S423. Introduce a fusion attention mechanism module to enhance the model's attention to key features.

[0021] In this technical solution, a depthwise separable convolution module is introduced, which combines depth convolution and pointwise convolution to extract local features of the signal and reduce the number of parameters; a Transformer module is introduced, and the multi-head attention mechanism simultaneously captures various long-term dependencies in the signal to enhance the long-term memory ability. The multi-heads share a projection matrix, and each head calculates the attention weights independently; a fusion attention mechanism module is introduced to dynamically adjust the importance of features from two dimensions: feature channels and time steps, enhancing the model's attention to key features.

[0022] The present invention is further set as: the noise feature library includes three types of measured noise samples: road surface excitation noise, environmental electromagnetic noise, and sensor drift noise.

[0023] A collision load reconstruction system, applicable to the above-mentioned collision load reconstruction method, includes: An acquisition module that acquires the excitation signal and response signal during impact and performs preprocessing; A processing module that performs multi-modal decomposition on the signal and selects effective modes for noise addition processing; A training module that trains the built deep learning network model to generate a trained prediction model; A reconstruction module that uses the trained prediction model to output a reconstructed load signal; An evaluation module that evaluates the performance of the prediction model.

[0024] The system of this technical solution mainly includes an acquisition module, a processing module, a training module, a reconstruction module, and an evaluation module. The acquisition module can acquire the structural response and load data during impact; the processing module can perform multi-modal decomposition and noise addition processing; the training module can build a deep learning network model and perform supervised training using a training set; the reconstruction model can input the structural response measured during the impact of the mechanical structure into the prediction model and then output the identified impact load; the evaluation module can evaluate the prediction model based on the average relative error between the reconstructed load signal and the actual result.

[0025] The present invention can bring the following beneficial effects: 1. A collision load reconstruction method involved in the present invention proposes a signal decomposition mechanism that combines PSO-VMD and correlation analysis. Compared with traditional EMD / VMD decomposition methods, it adaptively determines decomposition parameters through the particle swarm optimization algorithm, solves the mode mixing problem, and improves the accuracy. 2. Based on the modal screening strategy of Spearman correlation coefficient, by quantitatively evaluating the correlation degree between each IMF component and the original signal, it accurately locks the key signal carriers, and improves the target modal recognition accuracy compared with the traditional energy entropy screening method. 3. A multi-dimensional noise injection system is constructed, and a dynamic noise injection mechanism is designed for the signal characteristics of composite materials, breaking through the limitation of single noise enhancement and improving the data enhancement efficiency. 4. An heterogeneous fusion architecture of depthwise separable convolution and Transformer is adopted, a multi-scale attention collaboration mechanism is deployed, channel attention and temporal attention modules are respectively deployed in the spatial domain and the temporal domain, and a residual cross-connection structure is developed to achieve the deep fusion of local features and global dependencies, making the correlation coefficient between the reconstructed signal and the real excitation higher, further indicating that this method can accurately identify the impact load acting on the mechanical structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a schematic flow chart of a collision load reconstruction method of this application.

[0027] Figure 2 is a schematic diagram of the experimental signal acquisition scenario of this application.

[0028] Figure 3 is a schematic diagram of the deep learning prediction model of this application.

[0029] REFERENCE SIGNS 1. First strain gauge 2. Second strain gauge 3. Third strain gauge 4. Fourth strain gauge 5. Impact force hammer 6. Impact point 7.Computer. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only the best embodiments of the present invention, which are only used to explain the present invention and do not limit the protection scope of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0031] Embodiment 1 This embodiment proposes a collision load reconstruction method, referring to Figure 1 , which mainly includes the following steps.

[0032] Step S1, collect the excitation signal and response signal of the vehicle body side during impact, and then use the variational mode decomposition method to perform multi-modal decomposition on the above excitation signal and response signal to generate several modal components.

[0033] In this embodiment, a simulation scheme for the vehicle body side structure suffering from low-speed impact is designed, and a data collector is used to collect the corresponding excitation signal and response signal, and preprocess the signal. After the preprocessing is completed, the variational mode decomposition method is used to perform multi-modal decomposition on the signal.

[0034] For step S1, more specifically, it mainly includes the following sub-steps.

[0035] Step S11, first select the vehicle body transverse bending area as the structural impact measurement area, standardize the collected excitation signal and response signal, remove outliers, and eliminate the trend term.

[0036] More specifically, refer to Figure 2 , arrange 4 strain gauges and 1 force sensor on the door surface. Apply a low-speed impact with a force hammer, and the data collector synchronously records the excitation signal and response signal. Set the sampling frequency to 1 kHz, and collect the structural response signal and impact signal collected by the strain gauges and force sensor. Specifically, in the above impact measurement area, there is an impact point 6 in the middle area. A first strain gauge 1, a second strain gauge 2, a third strain gauge 3, and a fourth strain gauge 4 are respectively arranged around the impact point 6. The distances from the first strain gauge 1, the second strain gauge 2, the third strain gauge 3, and the fourth strain gauge 4 to the impact point 6 are all equal. Apply the corresponding impact load with the impact force hammer 5, and the first strain gauge 1, the second strain gauge 2, the third strain gauge 3, the fourth strain gauge 4, and the force sensor transmit the collected data to the computer 7 for data analysis.

[0037] Step S12, perform multi-scale decomposition on the signal after the above standardization process, and then use the particle swarm optimization algorithm (PSO) to optimize the penalty factor and the number of modal layers in the variational mode decomposition algorithm (VMD) to obtain multiple modal components.

[0038] In step S1 of this technical solution, the collected excitation signal and response signal are mainly standardized and multi-modally decomposed, and the penalty factor and the number of modal layers in the variational mode decomposition algorithm are optimized according to the particle swarm optimization algorithm to reduce the modal aliasing phenomenon.

[0039] Further, for the above step S12, its specific process is as follows: The improved permutation entropy is selected as the fitness function of the particle swarm optimization algorithm, and the optimized fitness function is used to evaluate the quality of the VMD parameters (α and K) corresponding to each particle. The specific formula is as follows: where m is the embedding dimension, representing the time series segmentation length in permutation entropy calculation, p is the probability of the occurrence of a specific permutation pattern of the modal component. u is the weight coefficient, K is the total number of modal components, is the standard deviation of the instantaneous frequency of the k-th modal component, is the average value of the instantaneous frequency of the k-th modal component.

[0040] The PSO is used to optimize the penalty factor and the number of modes. Let the size of the particle swarm be N, and the position of each particle be p i , and the velocity be v i . The PSO updates the velocity and position of the particle through the following formula: where w is the inertia weight, c1 and c2 are the learning factors, r1 and r2 are random numbers, α t o,i is the historical optimal position of particle i, and α g o is the global optimal position; When the change rate of the optimal fitness of the continuous iteration times is less than 0.1%, the iteration stops, and finally the optimal penalty factor α and the modal parameter K are obtained.

[0041] VMD decomposes the signal x i (t) into K modal components and solves the constrained variational problem. Its mathematical model is: The constraint conditions are: where u k (t) is the k-th modal component, ω k is the center frequency of the k-th mode, and α t is the penalty factor.

[0042] Step S2: Use Spearman correlation to select the modal component with the largest correlation as the optimal modal signal basis for noise addition processing.

[0043] For step S2, it includes the following sub-steps.

[0044] Step S21: Calculate the Spearman correlation coefficient between each modal component and the original signal, specifically: where d iis the rank difference of the corresponding data points, and n is the data length.

[0045] Step S22, screen out the mode with the greatest correlation with the original signal as the signal basis.

[0046] Step S3, construct a noise feature library, generate noise that conforms to the actual working conditions through dynamic noise injection, and expand the training data set while retaining the effective features.

[0047] For step S3, more specifically, it includes the following sub-steps.

[0048] Step S31, in order to ensure the time-frequency synchronization between the noise-added signal and the original signal, perform time-frequency analysis on the noise-added signal, calculate the time-frequency matrix through the Wigner distribution, introduce the Hamming window function to suppress the cross-term interference, and at the same time maintain the time-frequency resolution of the signal. The specific formula is: Among them, z(t) is the analytic signal, which is obtained by performing analytic processing on the original signal; h(τ) is the window function, and the definition of h(τ) is as follows: Among them, τ is the continuous time variable, T is the total length of the window; α and b are the window function coefficients, where α can take 0.54 and b can take 0.46 here. By weighting the instantaneous autocorrelation function, the influence of cross terms is suppressed.

[0049] Calculate the time-frequency signal-to-noise ratio of the original signal, and the formula is as follows: Among them, W z (t,f) is the time-frequency energy of the original signal, and W s (f,t) is the time-frequency energy of the noise.

[0050] S32, construct a noise feature library, generate composite noise that conforms to the actual working conditions through dynamic noise injection, establish an adaptive noise-adding mechanism based on the dynamic signal-to-noise ratio, and dynamically adjust the mixing noise ratio according to the time-frequency characteristics of the signal.

[0051] Specifically, the noise feature library includes three types of measured noise samples: road surface excitation noise, environmental electromagnetic noise, and sensor drift noise.

[0052] Establish a noise time-frequency matrix, as shown in the following formula: W n (t,f) = α(t,f)W nr (t,f) + β(t,f)W ne (t,f) + γ(t,f)W nd (t,f) Among them, α(t,f) + β(t,f) + γ(t,f) = 1, and the weight coefficients are controlled by the time-frequency signal-to-noise ratio. W nr (t,f) is the signal-to-noise ratio of road surface excitation noise, and W ne (t,f) is the signal-to-noise ratio of environmental electromagnetic noise, and W nd (t,f) is the signal-to-noise ratio of sensor drift noise.

[0053] Step S33: Dynamically adjust the noise energy according to the time-frequency signal-to-noise ratio of the original signal, which is specifically expressed as: The weight update expression is:

[0054] Step S34: Define the signal-to-noise ratio threshold. In the low signal-to-noise ratio region, when the signal-to-noise ratio is less than the lowest threshold, reduce the noise injection; in the high signal-to-noise ratio region, when the signal-to-noise ratio is less than the lowest threshold, increase the noise injection; Low signal-to-noise ratio region: SNR(t,f) < μ s -2σ s High signal-to-noise ratio region: SNR(t,f) ≥ μ s +σ s Among them, μ s is the mean value of the time-frequency signal-to-noise ratio of the original signal, and σ s is the standard deviation of the time-frequency signal-to-noise ratio of the original signal.

[0055] Step S35: Reconstruct the time-domain signal after adding noise through the inverse Wigner distribution, and establish the instantaneous autocorrelation function through the inverse Fourier transform, as shown in the following formula: The expression of the reconstructed signal is: Among them, R n (t,0) represents the instantaneous power value at τ = 0, and z n (t) is the time-domain signal after adding noise.

[0056] In this technical solution, in order to ensure the time-frequency synchronization between the signal after adding noise and the original signal, perform time-frequency analysis on the signal, calculate the time-frequency matrix through the Wigner distribution, introduce the Hamming window function to suppress the cross-term interference, and at the same time maintain the time-frequency resolution of the signal.

[0057] Step S4: Construct a prediction model of deep learning, use a double loss function for model training, and use the prediction model to generate a reconstructed load signal.

[0058] Reference Figure 3, Step S4 mainly includes the following sub-steps.

[0059] Step S41, divide the signal into time series windows, construct a time series data set, use the response signal in the denoised training set obtained in Step S3 as the input, and the excitation signal as the output.

[0060] Step S42, adopt a loop structure to perform multi-scale feature extraction and feature fusion on the depthwise separable convolution module, Transformer module, and fusion attention mechanism module.

[0061] Specifically, introduce a depthwise separable convolution module, combine depth convolution and pointwise convolution to extract local features of the signal and reduce the number of parameters.

[0062] Specifically, the results after per-channel convolution and pointwise convolution are: Among them, W i is the per-channel convolution kernel, V is the pointwise convolution kernel, k is the convolution kernel size, and i is the number of channels.

[0063] Introduce the Transformer module. The multi-head attention mechanism simultaneously captures multiple long-term dependencies in the signal, enhances the long-term memory ability, and the multi-heads share the projection matrix, and each head calculates the attention weights separately.

[0064] Specifically, map the input data to queries (Q), keys (K), and values (V) respectively, where: Q = XW Q , K = XW K , V = XW V Among them, W Q , W K , W V are the weights learned by each key. Calculate the attention scores through dot product:

[0065] Introduce a fused attention mechanism module to dynamically adjust the importance of features from two dimensions: feature channels and time steps, and enhance the model's attention to key features.

[0066] Specifically, combine channel attention and spatial attention, and its mathematical model is: Among them, x(t) is the excitation signal and payload data after dividing the time series window.

[0067] Step S43, perform training using a dual loss function; the dual loss function combines the mean squared error (MSE) and the mean absolute error (MAE). Step S44, optimize the training process using the early stopping method and the learning rate decay strategy.

[0068] In this technical solution, the prediction model includes a depthwise separable convolution module, a Transformer module, and a fusion attention mechanism module. Using the structural response of the training data obtained in S3 as the input and the impact load time history as the output, the deep learning prediction model is supervised and trained, and a dual loss function is used to improve the performance of the model.

[0069] For step S4, construct time series data with a time step of T c , a feature dimension of D, and divide the impact signals measured by the force sensor and the strain gauge and the corresponding structural response data into a training set and a test set according to a ratio. The data of the strain gauge is used as the feature input into the created deep learning model, and the impact load is used as the output. In the depthwise separable convolution module, the depth convolution assigns an independent convolution kernel to each input channel, with a size of k c ×1, where k c is the convolution kernel size in the time dimension. Each channel uses an independent convolution kernel to avoid cross-channel parameter sharing and reduce the amount of computation. The pointwise convolution uses M 1×1 convolution kernels to perform a linear transformation on the channels, map the multi-channel features to the target dimension, and achieve feature fusion in the channel dimension to enhance the feature expression ability. Compared with the standard convolution, the number of parameters is reduced from k c ×D×M to k c ×D×M, significantly reducing the computational complexity, separating the operations of the time and channel dimensions, and constructing a lightweight network.

[0070] In the Transformer module, the input data is mapped into three groups of matrices, namely query (Q), key (K), and value (V), through a linear transformation. The mapping formula is as follows: Q = XW Q , K = XW K , V = XW V where X is the input matrix, and W Q , W K , W V are learnable weights.

[0071] Divide Q, K, and V evenly into h heads along the feature dimension. The dimension of each head is (T, d k / h), and d k is the projection dimension of Q, K, and V. The input data is projected into multiple subspaces through a linear transformation, and each subspace focuses on different feature interaction patterns. The attention scores are calculated through the dot product, as shown in the following formula:

[0072] After residual connection and layer normalization, the problem of gradient disappearance in deep networks is alleviated. The feedforward neural network has two fully connected layers. The first layer maps the input data to a high-dimensional space, and the second layer maps the high-dimensional features back to the original dimension to capture complex features that the self-attention mechanism fails to capture. The swish non-linear activation function is adopted to further extract the non-linear features in the time series. Through the improved attention module, the model's attention to key features is enhanced, combining channel attention and spatial attention. Its mathematical model is as follows:

[0073] After three cycles of the deep convolutional module, Transformer, and the fusion attention mechanism, feature fusion is performed. After passing through the global pooling layer, the data dimension becomes the target dimension, and the reconstructed impact load is output.

[0074] During the training process, a training strategy with a dual loss function is adopted, combining MAE and MSE. The mathematical expressions are as follows: y i is the predicted value, is the actual value. The Adam optimizer is used, with an initial learning rate of 10 -3 , and the Dropout rate is 0.2. Early stopping is triggered when the validation loss has not decreased for 10 consecutive times.

[0075] The test set is input into the trained model, and the predicted load is output. The MRE and PRE of the predicted load and the actual load are calculated, and the prediction error is at a relatively low level. Experiments show that the present invention can still achieve high-precision load reconstruction under a sparse sensor array layout.

[0076] After step S4, step S5 may further be included to calculate the average relative error and peak relative error between the reconstructed load signal and the actual result to evaluate the performance of the prediction model.

[0077] Specifically, the average relative error MRE between the reconstructed result and the actual result is calculated, and the expression is as follows: where F represents the actual result and F' represents the reconstructed result.

[0078] The peak relative error PRE between the reconstructed result and the actual result is calculated, and the expression is as follows:

[0079] In this technical solution, the average relative error MRE and the peak relative error PRE are respectively calculated, and then the performance of the model is evaluated, which is convenient for subsequent improvement.

[0080] A collision load reconstruction method in this embodiment proposes a signal decomposition mechanism that combines PSO-VMD and correlation analysis. Compared with traditional EMD / VMD decomposition methods, it adaptively determines decomposition parameters through the particle swarm optimization algorithm to solve the mode mixing problem and improve accuracy. Secondly, based on the mode screening strategy of Spearman correlation coefficient, by quantitatively evaluating the correlation degree between each IMF component and the original signal, the key signal carriers are accurately locked. Compared with the traditional energy entropy screening method, the accuracy of target mode recognition is improved. A multi-dimensional noise injection system is constructed, and a dynamic noise injection mechanism is designed for the signal characteristics of composite materials, breaking through the limitation of single noise enhancement and improving the data enhancement efficiency. In addition, a heterogeneous fusion architecture of depthwise separable convolution and Transformer is designed, and a multi-scale attention collaboration mechanism is deployed. Channel attention and temporal attention modules are respectively deployed in the spatial domain (depthwise convolution branch) and the temporal domain (Transformer branch), and a residual cross-connection structure is developed to achieve deep fusion of local features and global dependencies, making the correlation coefficient between the reconstructed signal and the real excitation higher, further demonstrating that this method can accurately identify the impact load acting on the mechanical structure.

[0081] Further, in the signal decomposition mechanism of PSO-VMD, a two-dimensional solution space composed of the penalty factor α and the number of modal layers K is established. The optimized permutation entropy is used as the fitness function, and an inertia weight adaptive adjustment algorithm is introduced. When the particle swarm falls into a local optimum, the parameter search range is automatically relaxed to ensure the global optimization ability.

[0082] Further, for the K IMF components generated by decomposition, calculate their Spearman rank correlation coefficients with the original signal x(t), and select the mode with the largest correlation with the original signal as the signal basis. To ensure the time-frequency synchronization between the noisy signal and the original signal, perform time-frequency analysis on the signal, add noise signals that conform to the actual working conditions to the single signal, simulate the non-stationary noise characteristics in the actual working conditions, improve the sensitivity of the model to non-linear components, and maintain the time-frequency characteristics of the impact load during the data enhancement process.

[0083] Further, a deep learning model is established, the Adam optimizer is used as the optimization strategy, and the mean square error and mean absolute error between the reconstructed load and the actual load are selected as the double loss functions. The load reconstruction accuracy of the deep learning model is verified using the test set. To quantitatively evaluate the accuracy and generalization of the model, two physical quantities, the mean relative error and the peak relative error, are introduced.

[0084] Furthermore, a deep learning network integrating models is proposed for reconstructing the impact load of a mechanical structure. The model consists of a depthwise separable convolution module, a Transformer module, and a fusion attention mechanism. The model includes channel-wise convolution and pointwise convolution, which can capture local spatial features and fuse information between channels. A custom multi-head self-attention mechanism is adopted to calculate the global dependencies of adaptive time steps. A feed-forward neural network is added to extract non-linear features, and residual connections are used to alleviate the problem of gradient disappearance. Channel attention and spatial attention are combined to generate comprehensive attention weights, highlighting important features and suppressing noise. An early stopping strategy and a Dropout layer are introduced to randomly discard some neurons to avoid overfitting and enhance generalization ability.

[0085] In summary, the present invention proposes a method for reconstructing the lateral collision load of a vehicle body based on variational mode decomposition and deep learning to address the ill-posedness problem in load reconstruction; a verification platform for the load reconstruction method is built; the vibration response data is downsampled and filtered, followed by variational mode decomposition optimized by PSO, and the correlation of the decomposed modes is verified to find the mode with the largest correlation as the signal basis, and then noise is added to expand the training set of the deep learning model, and the mapping relationship between the impact load and the response is established in supervised learning; the results show that this method is more suitable for reconstructing the lateral collision load of a vehicle body with a sparse sensor array, and the accuracy of load reconstruction is high.

[0086] Embodiment 2 Based on the above Embodiment 1, this embodiment also proposes a collision load reconstruction system, which mainly includes an acquisition module, a processing module, a training module, a reconstruction module, and an evaluation module. Among them, the acquisition module is connected to the processing module, the processing module is connected to the training module, the training module is connected to the reconstruction module, and the reconstruction module is connected to the evaluation module.

[0087] The system of this technical solution mainly includes an acquisition module, a processing module, a training module, a reconstruction module, and an evaluation module. The acquisition module can collect the structural response and load data during impact; the processing module can perform multi-modal decomposition and noise addition processing; the training module can build a deep learning network model and perform supervised training using the training set; the reconstruction model can input the structural response measured during the impact of the mechanical structure into the prediction model and then output the identified impact load; the evaluation module can evaluate the prediction model according to the average relative error between the reconstructed load signal and the actual result.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A method for reconstructing a collision load, characterized in that It includes the following steps: S1. Collect the excitation signal and response signal when the vehicle body is laterally impacted, and perform multi-modal decomposition on the excitation signal and response signal using the variational mode decomposition method; S2. Select the modal component with the largest correlation as the optimal modal signal basis according to Spearman correlation for noise addition processing; S3. Establish a noise feature library, generate composite noise that conforms to the actual working conditions, and expand the training data set while retaining the effective features; S4. Construct a prediction model for deep learning, use a dual loss function for model training, and generate a reconstructed load signal using the prediction model.

2. The method for reconstructing a collision load according to claim 1, characterized in that, The step S1 includes: S11. Select the lateral bending area of the vehicle body as the structural impact measurement area, perform standardization processing on the collected excitation signal and response signal, remove outliers, and eliminate the trend term; S12. Perform multi-scale decomposition on the signal, and use the particle swarm optimization algorithm to optimize the penalty factor and the number of modal layers in the variational mode decomposition algorithm to obtain several modal components.

3. A method for reconstructing a collision load according to claim 1 or 2, characterized in that The step S2 specifically includes: calculating the Spearman correlation coefficient between each modal component and the original signal; screening out the modal with the largest correlation with the original signal as the signal basis.

4. The method for reconstructing a collision load according to claim 3, wherein The step S3 includes: S31. Perform time-frequency analysis on the noise-added signal, calculate the time-frequency matrix through the Wigner distribution, introduce the Hamming window function, and calculate the time-frequency signal-to-noise ratio of the original signal; S32. Construct a noise feature library, dynamically inject noise to generate composite noise that conforms to the actual working conditions, establish an adaptive noise addition mechanism based on the dynamic signal-to-noise ratio, and dynamically adjust the mixed noise ratio according to the time-frequency characteristics of the signal.

5. A method for reconstructing a collision load according to claim 4, characterized in that The step S3 further includes: S33. Dynamically adjust the noise energy according to the time-frequency signal-to-noise ratio of the original signal; S34. Define the signal-to-noise ratio threshold. In the low signal-to-noise ratio region, when the signal-to-noise ratio is less than the lowest threshold, reduce the noise injection; in the high signal-to-noise ratio region, when the signal-to-noise ratio is less than the lowest threshold, increase the noise injection; S35. Reconstruct the time-domain signal after noise addition through the inverse Wigner distribution.

6. A method for reconstructing a collision load according to claim 1 or 2, characterized in that The step S4 includes: S41. Divide the signal into time series windows, construct a time series data set, use the response signal obtained in S3 as the input, and the excitation signal as the output; S42. Adopt a cyclic structure to perform multi-scale feature extraction and feature fusion on the depthwise separable convolution module, Transformer module, and fusion attention mechanism module; S43. Use a dual loss function for training; S44. Use the early stopping method and learning rate decay strategy to optimize the training process.

7. A method for reconstructing a collision load according to claim 1 or 2, characterized in that It further includes step S5, calculating the average relative error and peak relative error between the reconstructed load signal and the actual result, and evaluating the performance of the prediction model.

8. A method for reconstructing a collision load according to claim 6, characterized in that, The step S42 includes: S421. Introduce a depthwise separable convolution module to extract the local features of the signal; S422. Introduce a Transformer module to capture various long-term dependencies in the signal and enhance the long-term memory ability; S423. Introduce a fusion attention mechanism module to enhance the model's attention to key features.

9. A method for reconstructing a collision load according to claim 1 or 4, characterized in that, The noise feature library contains three types of measured noise samples: road surface excitation noise, environmental electromagnetic noise, and sensor drift noise.

10. A collision load reconstruction system, applicable to a collision load reconstruction method according to any one of claims 1-9, characterized in that, It includes: The acquisition module acquires the excitation signal and the response signal during impact and performs preprocessing; The processing module performs multi-modal decomposition on the signals and selects the effective modes for noise addition processing; The training module trains the constructed deep learning network model to generate a trained prediction model; The reconstruction module outputs the reconstructed load signal using the trained prediction model; The evaluation module evaluates the performance of the prediction model.

Images