Collision waveform prediction model construction method, electronic device, and storage medium

Abstract

This application provides a method for constructing a collision waveform prediction model, an electronic device, and a storage medium. By embedding extracted explicit physical features into the learning process, physical constraints are introduced to ensure that the prediction results strictly follow fundamental physical laws, thereby improving the reliability and engineering applicability of the prediction results. By fusing local transient features and explicit physical features, a collaborative learning mechanism is constructed between macroscopic curve morphology, mesoscopic physical features, and microscopic wave patterns, achieving comprehensive and multi-granular accurate prediction of acceleration curves. Through multi-task learning and decoding of global feature vectors, physical regularization and multi-task collaborative training are achieved, significantly improving the stability and adaptability of the prediction process under complex scenarios such as parameter perturbations and changing operating conditions. This ensures both prediction accuracy and the physical rationality of the results, as well as the model's efficient generalization ability.

Description

Technical Field

[0001] This application relates to the field of vehicle collision prediction technology, and in particular to a method for constructing a collision waveform prediction model, an electronic device, and a storage medium. Background Technology

[0002] Collision acceleration curves are crucial data in vehicle crash safety assessments, making accurate prediction or extraction of these curves of paramount technical value. A core challenge in addressing vehicle crash safety development is the high-precision, high-efficiency prediction of collision acceleration curves. Finite element simulation prediction methods are computationally expensive, while purely data-driven machine learning methods suffer from issues such as lack of physical consistency, poor interpretability, and weak generalization ability with small sample sizes. Summary of the Invention

[0003] In view of this, the purpose of this application is to propose a method for constructing a collision waveform prediction model, an electronic device and a storage medium, which deeply embeds the physical knowledge of the collision field into the learning process and constructs a physics-guided multi-task learning framework, thereby ensuring the physical rationality of the results and the efficient generalization ability of the model while ensuring the prediction accuracy.

[0004] To achieve the above objectives, this application provides a method for constructing a collision waveform prediction model, comprising:

[0005] Adaptive data augmentation is performed on the real collision acceleration curve to obtain an enhanced collision acceleration curve, and input features are constructed based on the collision response data corresponding to the real collision acceleration curve. Extract the local transient features of the input features, extract the explicit physical features from the enhanced collision acceleration curve, and fuse the local transient features and the explicit physical features to obtain a physics-guided attention map; The global feature vector is determined based on the attention map and the local transient features. The global feature vector is then decoded using multi-task learning to obtain a prediction model that outputs a predicted collision acceleration curve and predicted physical features.

[0006] Optionally, the method for constructing the collision waveform prediction model also includes: The mean square error loss is determined based on the coordinate values ​​at the same moment between the predicted collision acceleration curve and the actual collision acceleration curve. The dynamic time warping loss is determined based on the morphological similarity between the predicted collision acceleration curve and the actual collision acceleration curve. The physical feature loss is determined based on the value of the same feature element between the predicted physical feature and the explicit physical feature; A loss function is constructed based on preset weighting coefficients, the mean square error loss, the dynamic time warping loss, and the physical feature loss; The prediction model is iteratively optimized based on the loss function, and the target prediction model when the loss function converges is determined as the optimal prediction model.

[0007] Optionally, the adaptive data augmentation of the real collision acceleration curve to obtain an enhanced collision acceleration curve includes: The time axis of the real collision acceleration curve is stretched or compressed according to a preset time distortion factor to obtain a time-torque collision acceleration curve. The time-torque collision acceleration curve is scaled according to a preset amplitude scaling factor to obtain an amplitude-scaled collision acceleration curve. The preset Gaussian white noise is superimposed on the amplitude-scaled collision acceleration curve to obtain the noise collision acceleration curve; The filter window length is determined based on the length of the noise collision acceleration curve, and the noise collision acceleration curve is smoothed based on the filter window length to obtain the enhanced collision acceleration curve.

[0008] Optionally, determining the filter window length based on the length of the noise collision acceleration curve includes: The calculated length is obtained by reducing and rounding the curve length and performing odd number transformation. In response to the calculation length being less than or equal to a preset minimum threshold length, the calculation length is determined as the filter window length; In response to the calculated length being greater than a preset minimum threshold length, the minimum threshold length is determined as the filter window length.

[0009] Optionally, fusing the local transient features and the explicit physical features to obtain a physics-guided attention map includes: Determine the time dimension of the local transient feature, and broadcast the explicit physical feature to the time dimension to obtain the broadcast explicit physical feature; The broadcast explicit physical features and the local transient features are spliced ​​together along the channel dimension to obtain the spliced ​​features; The spliced ​​features are fused and dimensionality reduced across channels using a preset convolutional layer to obtain fused features. Each time step value of the fused features is then mapped to a normalized interval to obtain the attention map.

[0010] Optionally, determining the global feature vector based on the attention map and the local transient features includes: The attention map and the local transient features are multiplied element-wise to obtain the weighted features; By capturing the long-range dependencies of the time series of the weighted features, feature representations are obtained; The feature representation is subjected to global average pooling to obtain the global feature vector.

[0011] Optionally, the prediction model includes a main task decoder and an auxiliary task decoder; the prediction model that performs multi-task learning decoding on the global feature vector to obtain the output predicted collision acceleration curve and predicted physical features includes: The global feature vector is expanded in dimension to obtain extended features; the extended features are subjected to feature sufficiency and nonlinear transformation to obtain enhanced features; the enhanced features are reconstructed into a sequence to obtain the main task decoder that outputs the predicted collision acceleration curve; The global feature vector is continuously compressed in dimension to obtain compressed features; the compressed features are then projected onto physical attributes to obtain the auxiliary task decoder that outputs the predicted physical features.

[0012] Optionally, extracting explicit physical features from the enhanced collision acceleration curve includes: The peak acceleration, peak occurrence time, peak half-width, and peak prominence of the enhanced collision acceleration curve are extracted to obtain the peak characteristics; The total energy and the energy ratio of the first and second halves of the enhanced collision acceleration curve are extracted to obtain the energy characteristics. The mean, standard deviation, skewness, and kurtosis of the enhanced collision acceleration curve are extracted to obtain statistical characteristics; Extract the average gradient and maximum gradient of the enhanced collision acceleration curve to obtain gradient features; Linear fitting is performed on the end of the enhanced collision acceleration curve to obtain the attenuation characteristics; The explicit physical characteristics are obtained by integrating the peak characteristics, energy characteristics, statistical characteristics, gradient characteristics, and attenuation characteristics.

[0013] Based on the same inventive concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0014] Based on the same inventive concept, this application also provides a non-transitory computer-readable storage medium that stores computer instructions for causing a computer to perform the method described above.

[0015] As can be seen from the above, the collision waveform prediction model construction method, electronic device, and storage medium provided in this application adaptively augment the real collision acceleration curve to obtain an enhanced collision acceleration curve, and construct input features based on the collision response data corresponding to the real collision acceleration curve; extract local transient features from the input features, extract explicit physical features from the enhanced collision acceleration curve, and fuse the local transient features and explicit physical features to obtain a physics-guided attention map; determine the global feature vector based on the attention map and local transient features, and perform multi-task learning decoding on the global feature vector to obtain the predicted collision acceleration curve and predicted physical features. By embedding the extracted explicit physical features into the learning process, physical constraints are introduced to ensure that the prediction results strictly follow basic physical laws such as energy conservation and momentum theorem, thereby improving the reliability and engineering applicability of the prediction results. By fusing local transient features and explicit physical features, a collaborative learning mechanism is constructed between macroscopic curve morphology, mesoscopic physical features, and microscopic wave patterns to achieve comprehensive and multi-granular accurate prediction of acceleration curves. By using global feature vectors for multi-task learning and decoding, physical regularization and multi-task collaborative training are achieved, significantly improving the stability and adaptability of the prediction process under complex scenarios such as parameter perturbations and changing operating conditions. Domain physics knowledge is deeply embedded into the machine learning model used for collision acceleration curves, constructing a physics-guided multi-task learning framework. This ensures both prediction accuracy and the physical plausibility of the results, as well as the model's efficient generalization ability. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the method for constructing a collision waveform prediction model according to an embodiment of this application; Figure 2 This is a schematic diagram of the multi-task learning framework according to an embodiment of this application; Figure 3 This is a schematic diagram illustrating the composition of the comprehensive loss function in an embodiment of this application; Figure 4 This is a schematic diagram of the apparatus for constructing a collision waveform prediction model according to an embodiment of this application; Figure 5 This is a schematic diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] In this article, it is important to understand that any number of elements in the accompanying figures is for illustrative purposes and not for limitation, and any naming is for distinction only and has no limiting meaning.

[0021] Based on the above background description, the following situations also exist in the related technologies: The relevant technical solutions mainly rely on the following two technical paths to predict collision acceleration curves: numerical simulation method based on finite element analysis and data-driven method based on traditional machine learning.

[0022] Among them, numerical simulation methods based on finite element analysis have significant technical limitations: It consumes huge amounts of computing resources, and a single complete simulation usually takes several hours to several days; it is highly sensitive to parameters, and small changes in input can lead to significant deviations in output results; moreover, it has limited ability to model complex nonlinear systems and it is difficult to accurately capture multi-scale coupling effects in actual physical processes.

[0023] Without CAD data, it is impossible to obtain the collision acceleration curve.

[0024] While data-driven methods based on traditional machine learning have alleviated the computational efficiency problem to some extent, they still face many technical bottlenecks: The monitoring signal is singular, relying solely on numerical fitting of the curve shape, and lacks explicit modeling of the underlying physical laws; Insufficient feature representation makes it difficult to extract sufficiently rich physical information from limited input parameters; excessive data dependence makes it prone to overfitting under small sample conditions, which seriously affects the model's generalization ability and engineering applicability.

[0025] It can be seen that in solving the problem of vehicle collision safety development, there are three core technical issues in predicting collision acceleration curves: ① Technical risks of lack of physical consistency: Purely data-driven black box models may produce prediction results that violate basic physical laws, such as energy non-conservation and momentum anomalies, making the prediction results physically unreliable and severely limiting their practical application in safety-critical fields.

[0026] ② Technical gap in multi-scale feature modeling: The methods in related technologies lack the ability to collaboratively model the features of different physical scales in the collision acceleration curve, and cannot simultaneously capture the intrinsic relationship between macroscopic curve morphology, mesoscopic energy distribution features and microscopic fluctuation details.

[0027] ③ Technical limitations due to insufficient domain adaptability: When system parameters or working conditions change, the prediction models in related technologies often need to collect a large amount of data again for full training. Their transferability and adaptability are poor, and they cannot meet the needs of rapid iterative engineering development.

[0028] In summary, how to achieve high-precision, physically consistent, and highly interpretable intelligent prediction of acceleration curves under small sample conditions is a core technical challenge that urgently needs to be solved in the field of vehicle collision safety development.

[0029] The collision waveform prediction model construction method, electronic device, and storage medium provided in this application embodiment adaptively augment the real collision acceleration curve to obtain an enhanced collision acceleration curve, and construct input features based on the collision response data corresponding to the real collision acceleration curve; extract local transient features from the input features, extract explicit physical features from the enhanced collision acceleration curve, and fuse the local transient features and explicit physical features to obtain a physics-guided attention map; determine a global feature vector based on the attention map and local transient features, and perform multi-task learning decoding on the global feature vector to obtain the predicted collision acceleration curve and predicted physical features. By embedding the extracted explicit physical features into the learning process, physical constraints are introduced to ensure that the prediction results strictly follow basic physical laws such as energy conservation and momentum theorem, thereby improving the reliability and engineering applicability of the prediction results. By fusing local transient features and explicit physical features, a collaborative learning mechanism is constructed between macroscopic curve morphology, mesoscopic physical features, and microscopic wave patterns to achieve comprehensive and multi-granular accurate prediction of acceleration curves. By using global feature vectors for multi-task learning and decoding, physical regularization and multi-task collaborative training are achieved, significantly improving the stability and adaptability of the prediction process under complex scenarios such as parameter perturbations and changing operating conditions. Domain physics knowledge is deeply embedded into the machine learning model used for collision acceleration curves, constructing a physics-guided multi-task learning framework. This ensures both prediction accuracy and the physical plausibility of the results, as well as the model's efficient generalization ability.

[0030] The following describes in detail, with reference to the accompanying drawings, the method for constructing the collision waveform prediction model provided by the embodiments of this application.

[0031] In some embodiments, such as Figure 1 As shown, the method for constructing the collision waveform prediction model includes steps 101-103.

[0032] Step 101: Perform adaptive data augmentation on the real collision acceleration curve to obtain the augmented collision acceleration curve, and construct input features based on the collision response data corresponding to the real collision acceleration curve.

[0033] In practice, historical collision data or collision simulation data are collected, and training samples are obtained from the historical collision data or collision simulation data. Each training sample includes the real collision acceleration curve Y(t) and input features X.

[0034] The input feature vector X consists of collision response data (key structured parameters that affect the collision response). The input feature vector X has 94 dimensions. Its specific composition and some example data are shown in Table 1.

[0035] Table 1 Examples of Input Feature Vector Construction

[0036] It can be seen that the input feature vector includes: Operating conditions: Collision type (e.g., 100% frontal overlap rigid barrier, 40% frontal overlap deformable barrier, side collision, etc.), collision speed (km / h); Basic vehicle information: curb weight (kg), powertrain type (gasoline / pure electric / hybrid), drive type (front-wheel drive / rear-wheel drive / four-wheel drive), body type (unibody / body-on-frame); Energy absorption space information: distance from the front bumper to the front end of the anti-collision beam (mm), distance from the rear end of the anti-collision beam to the cooling module (mm), deformation space of the front energy absorption structure (DFF) (mm); Key component information: Energy-absorbing box material grade, thickness (mm), and cross-sectional shape code; anti-collision beam material, thickness, and cross-section; longitudinal beam material, thickness, and cross-section; subframe material, thickness, and cross-section; shotgun material, thickness, and cross-section, etc. Powertrain information: powertrain layout (transverse / longitudinal), external dimensions (length × width × height, mm), number of mounting points and failure conditions.

[0037] The actual collision acceleration curve Y(t) is a signal of acceleration over time collected at a specific location on the vehicle (such as the bottom of the B-pillar, the sill beam, etc.) under a specific collision condition. In this embodiment, the curve is represented in discrete time series form, with a sampling frequency f_s of 10kHz and a total sampling duration T_total of 120ms. Therefore, each curve contains T = f_s × T_total = 1200 data points, denoted as Y = [y_1, y_2, ..., y_T].

[0038] In some embodiments, adaptive data augmentation is performed on the real collision acceleration curve to obtain an enhanced collision acceleration curve, including: The time axis of the real collision acceleration curve is stretched or compressed according to the preset time warp factor to obtain the time torque collision acceleration curve. The amplitude of the time-torque collision acceleration curve is scaled according to a preset amplitude scaling factor to obtain an amplitude-scaled collision acceleration curve. The preset Gaussian white noise is superimposed on the amplitude-scaled collision acceleration curve to obtain the noise collision acceleration curve; The length of the filter window is determined based on the length of the noise collision acceleration curve, and the noise collision acceleration curve is smoothed according to the length of the filter window to obtain the enhanced collision acceleration curve.

[0039] The process of determining the filter window length based on the length of the noise collision acceleration curve includes: The calculated length is obtained by reducing and rounding the curve length and performing odd-number transformation. In response to the calculation length being less than or equal to the preset minimum threshold length, the calculation length is determined as the filter window length; In response to a calculated length greater than a preset minimum threshold length, the minimum threshold length is determined as the filter window length.

[0040] In practice, to improve the model's robustness and generalization ability under small sample conditions, adaptive data augmentation is required on the actual collision acceleration curves. For each original curve Y(t), the following operations are performed sequentially to generate K (e.g., K=3) augmented curves: S1. Adaptive Time Warp: Generate a time warp factor λ_t following a uniform distribution U(0.97, 1.03). Map the original time axis t to a new time axis t' = λ_t × t, and obtain the acceleration values ​​at the new time points through linear interpolation, generating the curve Y_1(t'). This operation simulates small fluctuations in the time scale during physical processes.

[0041] S2. Adaptive Amplitude Scaling: Generate an amplitude scaling factor λ_a that follows a uniform distribution U(0.98, 1.02). Scaling the curve Y_1 by amplitude yields Y_2(t) = λ_a × Y_1(t'). This operation simulates the response amplitude variation under different collision energy levels.

[0042] S3. Adaptive Noise Injection: Generate white noise ε following a Gaussian distribution N(0, 0.003) and superimpose it onto curve Y_2 to obtain the noisy curve Y_3(t) = Y_2(t) + ε. This operation simulates sensor measurement errors or environmental interference.

[0043] S4. Adaptive Intelligent Smoothing: A Savitzky-Golay filter is used to smooth Y_3, resulting in the enhanced collision acceleration curve Y_aug(t). The filter window length L is adaptively determined based on the curve length T: L = min(15, floor(T / 10)×2 + 1), ensuring that the window length is an odd number and within a reasonable range. This effectively removes high-frequency noise while preserving key physical characteristics of the curve, such as peaks and valleys.

[0044] For example, to enhance 50 original real collision acceleration curves in the database, taking the first real collision acceleration curve Y1(t) as an example, the specific process of generating 3 enhanced curves is as follows: A random λ_t=0.98 is generated, and adaptive time warping is performed to obtain the time torque collision acceleration curve Y1(t').

[0045] A random λ_a=1.01 is generated, and adaptive amplitude scaling is performed to obtain the amplitude-scaled collision acceleration curve Y_2(t)= λ_a × Y_1(t').

[0046] Adding Gaussian white noise ε with a mean of 0 and a standard deviation of 0.003, we obtain the noise collision acceleration curve Y_3(t) = Y_2(t) +ε.

[0047] The final enhanced collision acceleration curve Y_aug(t) is obtained by smoothing the surface using a Savitzky-Golay filter with a window length of L=11 (L = min(15, floor(1200 / 10)×2+1)=11 when T=1200).

[0048] Repeat the above process three times, generating three enhanced curves for each original curve. The final dataset is expanded to 50 × 4 = 200 curves (original + 3 enhanced curves).

[0049] In the adaptive intelligent smoothing process, when determining the Savitzky-Golay filter window length L based on the curve length T of the noise collision acceleration curve, T is first reduced, rounded, and transformed into an odd number to obtain the calculated length L_calc = floor(T / 10) × 2 + 1. Then, it is determined whether L_calc is less than or equal to a preset minimum threshold length L_min (e.g., 15). If L_calc ≤ L_min, then L_calc is determined as the filter window length L; if L_calc > L_min, then L_min is determined as the filter window length L. The adaptive rule L = min(15, floor(T / 10)×2+1)=11 ensures that the window length is always odd and not too large, effectively preserving the key physical characteristics of the curve while smoothing noise.

[0050] Step 102: Extract local transient features from the input features, extract explicit physical features from the enhanced collision acceleration curve, and fuse the local transient features and explicit physical features to obtain a physics-guided attention map.

[0051] In practice, the 94-dimensional input feature X is embedded through a fully connected layer to transform it into a high-dimensional feature space. Simultaneously, the enhanced acceleration curve Y_aug(t) is used as another input.

[0052] Local transient feature extraction: The acceleration sequence Y_aug is input into a one-dimensional convolutional neural network encoder. This encoder uses convolutional kernels of different sizes (e.g., sizes 3 and 5) to extract local transient features under different receptive fields, such as peaks and abrupt change points. The final output is a feature map F_cnn of dimension T×D to represent the local transient features, where D is the feature dimension (e.g., 64).

[0053] In some embodiments, extracting explicit physical features from the enhanced collision acceleration curve includes: Peak features are obtained by extracting the peak acceleration, peak occurrence time, peak half-width, and peak prominence from the enhanced collision acceleration curve; The total energy and the energy ratio of the first and second halves of the enhanced collision acceleration curve are extracted to obtain energy characteristics. The mean, standard deviation, skewness, and kurtosis of the enhanced collision acceleration curves are extracted to obtain statistical characteristics; Extract the average gradient and maximum gradient from the enhanced collision acceleration curve to obtain gradient features; Linear fitting is performed on the end of the enhanced collision acceleration curve to obtain the attenuation characteristics; By integrating peak characteristics, energy characteristics, statistical characteristics, gradient characteristics, and decay characteristics, explicit physical characteristics are obtained.

[0054] In specific implementation, explicit physical feature extraction involves directly calculating a set of explicit physical features Φ_raw from the enhanced collision acceleration curve Y_aug, which characterizes its physical essence. For example, the explicit physical features Φ_raw are 18-dimensional features, including: Peak characteristics (4 dimensions): peak acceleration A_max, peak occurrence time t_max, peak half-width W, peak prominence P.

[0055] Energy characteristics (2D): Total energy E_total (obtained by numerical integration of the square of acceleration), energy ratio of the first and second halves R_energy.

[0056] Statistical characteristics (4 dimensions): mean μ, standard deviation σ, skewness S, kurtosis K.

[0057] Gradient features (2-dimensional): mean gradient G_mean, maximum gradient G_max.

[0058] Attenuation characteristics (6 dimensions): Linear fitting is performed on the end of the curve (such as the last 20% of the time period) to obtain parameters such as slope k_tail, intercept and fitting error.

[0059] Where A_max = max(Y_aug); t_max = argmax(Y_aug) × Δt (converted to time); W: Peak half-width, which is the width at half the peak height (calculated by interpolation). P: Peak prominence, which is the difference between the peak value and the average value of the adjacent valley values; E_total = Σ(Y_aug[i]^2)×Δt, Δt=0.0001s; R_energy = E_front / E_rear, where E_front is the energy in the first half of the time period and E_rear is the energy in the second half of the time period; μ = mean(Y_aug); σ = std(Y_aug); S = skewness(Y_aug); K = kurtosis(Y_aug); G_mean = mean(|Y_aug[i+1]-Y_aug[i]|) / Δt; G_max = max(|Y_aug[i+1]-Y_aug[i]|) / Δt; Attenuation characteristics: The acceleration data of the last 20% of time points (i.e., t∈[96ms,120ms]) were linearly fitted to y= k_tail × t + b, and a total of 6 parameters were obtained, including the slope k_tail and the fitting determination coefficient R².

[0060] The local transient features F_cnn extracted by the CNN are fused with the explicit physical features Φ_raw. First, Φ_raw is broadcast along the time dimension, expanding its dimension to T×18, consistent with F_cnn, denoted as Φ_raw_expanded. Then, F_cnn and Φ_raw_expanded are concatenated along the feature dimension to obtain a fused feature. Next, a 1×1 convolutional layer is used to perform cross-channel information fusion and dimensionality reduction on this fused feature, compressing the number of channels to 1. Finally, a sigmoid activation function is used to map the value at each time step to the [0,1] interval, generating a physics-guided attention map A ∈ R^(T×1). This attention map quantifies the importance of different time points on the acceleration curve, enabling the model to focus on physically significant regions such as peaks and energy concentration areas.

[0061] Step 103: Determine the global feature vector based on the attention map and local transient features, perform multi-task learning decoding on the global feature vector, and obtain the prediction model of the output predicted collision acceleration curve and predicted physical features.

[0062] In practice, after obtaining the attention map A, it is multiplied element-wise with the local transient feature F_cnn to obtain the weighted feature F_attended = F_cnn ⊙ A, which is weighted by physical importance. The weighted feature strengthens the feature representation of key physical moments.

[0063] Then, the weighted feature F_attended is input into a Transformer encoder. The Transformer utilizes its multi-head self-attention mechanism to capture long-range dependencies throughout the acceleration time series, outputting a feature representation F_transformer ∈ R^(T×D) containing global contextual information. This feature representation is then subjected to global average pooling along the time dimension T, resulting in a fixed-length global feature vector F_global ∈ R^D that integrates multi-scale local information and global dependencies.

[0064] After obtaining the global feature vector F_global, the model simultaneously outputs two prediction results through a multi-task learning decoding architecture: Main Task Decoding (Acceleration Curve Regression): A main task decoder is constructed to decode the global feature vector F_global. The main task decoder consists of several fully connected layers. Specifically, firstly, a fully connected layer maps the 128-dimensional F_global to a 256-dimensional space and performs ReLU activation to obtain the extended feature H1. Then, a 256-dimensional fully connected layer and ReLU activation are used for feature recombination and nonlinear transformation to obtain the enhanced feature H2. Finally, an output layer containing 1200 neurons (without an activation function) maps H2 to the predicted collision acceleration curve. .

[0065] Auxiliary task decoding (physical feature prediction): An auxiliary task decoder is constructed to decode the global feature vector F_global in parallel with the main task decoder. The auxiliary task decoder first maps the 128-dimensional F_global to a 64-dimensional space through a fully connected layer and then performs ReLU activation to obtain the compressed feature Φ1. Then, an output layer containing 18 neurons maps the compressed feature Φ1 to an 18-dimensional predicted physical feature Φ_pred, where each dimension corresponds one-to-one with the explicit physical feature Φ_raw calculated from the curve.

[0066] By using a parallel decoding design that shares global feature vectors, the auxiliary task forces the encoder to generate F_global containing sufficient physical information, thus providing a physical consistency constraint for the curve prediction of the main task, enabling the prediction model to achieve collaborative learning of physical laws and data-driven approaches.

[0067] In summary, by embedding extracted explicit physical features into the learning process, physical constraints are introduced to ensure that the prediction results strictly adhere to fundamental physical laws such as energy conservation and momentum theorem, thereby improving the reliability and engineering applicability of the prediction results. By fusing local transient features and explicit physical features, a collaborative learning mechanism is constructed between macroscopic curve morphology, mesoscopic physical features, and microscopic wave patterns, achieving comprehensive and multi-granular accurate prediction of acceleration curves. Multi-task learning decoding using global feature vectors enables physical regularization and multi-task collaborative training, significantly improving the stability and adaptability of the prediction process under complex scenarios such as parameter perturbations and changing operating conditions. Deeply embedding domain physics knowledge into the machine learning model used for collision acceleration curves constructs a physics-guided multi-task learning framework, thus ensuring both prediction accuracy and the physical rationality of the results, as well as the model's efficient generalization ability.

[0068] In some embodiments, such as Figure 2 As shown, the method for constructing the collision waveform prediction model also includes: The mean square error loss is determined based on the coordinate values ​​at the same moment between the predicted and actual collision acceleration curves.

[0069] In practice, after decoding the global feature vector, to ensure that the collision waveform prediction model based on physics-guided multi-task learning possesses high numerical accuracy, correct waveform morphology, and strict physical consistency, a comprehensive loss function needs to be constructed to optimize the model parameters end-to-end. This comprehensive loss function is not a single error metric, but a supervisory signal system composed of a weighted combination of multiple sub-loss terms with clear physical or mathematical meanings, designed to guide and constrain the model's learning process from different dimensions.

[0070] The overall loss function consists of a weighted sum of three core sub-loss components: mean squared error loss (L_MSE), dynamic time warping loss (L_DTW), and physical feature loss (L_phy). The definitions and calculation methods of each component are as follows: Mean squared error loss L_MSE: Calculates the predicted acceleration curve The mean of the sum of squares of the differences between the point-by-point coordinates of y(t) and the true curve Y(t), i.e., L_MSE = (1 / T) × Σ(y_i - _i) 2 This ensures the numerical accuracy of the prediction results.

[0071] The dynamic time warping loss is determined based on the morphological similarity between the predicted and actual collision acceleration curves.

[0072] In practical implementation, the Dynamic Time Warping Loss (L_DTW) focuses on measuring the similarity between the predicted and actual curves in terms of overall waveform shape. Due to the complex nonlinearity of the collision process, there may be a slight time phase shift between the predicted and actual collision acceleration curves (i.e., a slight misalignment in the timing of peaks or troughs). L_MSE is very sensitive to such shifts and imposes a large penalty on them.

[0073] The dynamic time warping algorithm automatically finds an optimal nonlinear alignment path between the predicted and actual collision acceleration curves and calculates the cumulative distance under this optimal alignment. The dynamic time warping loss is insensitive to small scaling and offsets on the time axis, and can more effectively evaluate whether the two curves are consistent in "shape," thereby guiding the model to learn the correct collision waveform pattern, rather than just a precise point correspondence. That is, calculating the predicted collision acceleration curve... The dynamic time warping distance between the actual collision acceleration curve Y(t) and the actual collision acceleration curve Y(t). The dynamic time warping loss L_DTW measures the similarity of the overall shape of the two curves by finding the optimal nonlinear alignment path. It is insensitive to small phase shifts on the time axis, thus guiding the model to focus on the correctness of the waveform shape.

[0074] The physical feature loss is determined based on the value of the same feature element between the predicted physical feature and the explicit physical feature.

[0075] Physical Feature Loss L_phy: The physical feature loss L_phy is a key regularization term that introduces physical consistency. Its computational model calculates the error between the physical feature Φ_pred predicted by the auxiliary task decoder and the explicit physical feature vector Φ_raw obtained directly from the real acceleration curve. The explicit physical feature vector Φ_raw contains 18 dimensions of features across 5 categories: peak, energy, statistics, gradient, and decay.

[0076] Calculate the mean square error of each corresponding element between the predicted physical feature Φ_pred and the explicit physical feature vector Φ_raw, i.e., L_phy = (1 / 18) × Σ(Φ_pred_j - Φ_raw_j) 2 This approach uses physical laws as a supervisory signal, forcing the model to learn representations that accurately reflect key physical properties.

[0077] A loss function is constructed based on preset weighting coefficients, mean square error loss, dynamic time warping loss, and physical feature loss; In specific implementation, such as Figure 3As shown, the total loss function is defined as: L_total = α × L_MSE + β × L_DTW + γ × L_phy, where α, β, and γ are balancing weights, optionally set to 0.6, 0.2, and 0.2 respectively. Using the original and augmented datasets, the Adam optimizer is used to perform end-to-end joint training on the model until the loss function converges. The model obtained at this point is the optimal prediction model.

[0078] The prediction model is iteratively optimized based on the loss function, and the target prediction model when the loss function converges is determined as the optimal prediction model.

[0079] In practice, during model training, the optimizer (such as Adam) aims to minimize the comprehensive loss function L_total, backpropagates gradients, and synchronously updates all parameters in the neural network feature extraction module, attention module, and main / auxiliary task decoder.

[0080] The collaborative optimization mechanism of this process is reflected in: Master-Auxiliary Task Collaboration: The gradient of L_phy is backpropagated through the auxiliary task decoder, forcing the shared global feature vector F_global to encode rich physical information. This makes the master task decoder naturally subject to physical constraints when using F_global to predict curves.

[0081] Multi-objective collaboration: L_MSE and L_DTW jointly ensure that the predicted curve approximates reality in both "numerical" and "morphological" aspects; L_phy ensures that this approximation is within the limits allowed by physical laws. The three complement each other, guiding the model to learn a prediction function that is not only highly fitted to the data but also physically interpretable and reliable.

[0082] The target prediction model at which the loss function converges is determined as the optimal prediction model. By fusing the three supervision signals of numerical error, morphological similarity and physical consistency and weighting and balancing them, the technical problem of lack of physical consistency in pure data-driven models is avoided. This is the core of achieving high-precision and physically reliable intelligent prediction of collision waveforms.

[0083] In some embodiments, local transient features and explicit physical features are fused to obtain a physics-guided attention map, including: Determine the time dimension of the local transient features, and broadcast the explicit physical features to the time dimension to obtain the broadcast explicit physical features; By splicing broadcast explicit physical features and local transient features along the channel dimension, spliced ​​features are obtained. Based on the preset convolutional layers, cross-channel information fusion and dimensionality reduction are performed on the spliced ​​features to obtain fused features. Each time step value of the fused features is mapped to a normalized interval to obtain an attention map.

[0084] In specific implementation, the temporal dimension T of the local transient feature F_cnn extracted by the CNN is first determined. Next, the explicit physical feature Φ_raw, with a dimension of 18, is broadcast along the temporal dimension and copied T times, resulting in a broadcast explicit physical feature Φ_raw_expanded aligned with the temporal dimension of F_cnn, with a dimension of [T×18]. Then, Φ_raw_expanded is concatenated with F_cnn (with a dimension of [T×D], where D is the feature dimension, e.g., 64) along the channel dimension, resulting in a concatenated feature that integrates local details and global physical properties, with a dimension of [T×(D+18)]. Subsequently, a pre-defined 1×1 convolutional layer is used to perform cross-channel information fusion and dimensionality reduction on this concatenated feature, compressing the number of channels from (D+18) to 1, resulting in a fused feature with a dimension of [T×1]. Finally, the Sigmoid activation function is used to map the value of each time step of this fused feature to a normalized interval of [0,1], thus obtaining the physics-guided attention map A.

[0085] In some embodiments, determining the global feature vector based on the attention map and local transient features includes: The attention map and local transient features are multiplied element-wise to obtain the weighted features; By capturing the long-range dependencies of the time series of weighted features, feature representations are obtained; Global average pooling is performed on the feature representation to obtain the global feature vector.

[0086] In practice, firstly, the physically guided attention map A is multiplied element-wise (Hadamard product) with the local transient features F_cnn to obtain the weighted features F_attended = F_cnn ⊙ A. This operation uses the attention map to weight the importance of the original local features. Next, the weighted features F_attended are input into a Transformer encoder. The Transformer encoder uses its internal multi-head self-attention mechanism to capture the long-range dependencies of the weighted features over time and outputs a feature representation F_transformer that contains global contextual information. Finally, global average pooling is performed on the feature representation F_transformer over the time dimension T, which averages the features across all time steps to obtain a fixed-length global feature vector F_global that summarizes the information of the entire sequence.

[0087] In some embodiments, the prediction model includes a main task decoder and an auxiliary task decoder; the prediction model, which performs multi-task learning decoding on the global feature vector to obtain the output predicted collision acceleration curve and predicted physical features, includes: The global feature vector is expanded in dimension to obtain extended features; the extended features are subjected to feature sufficiency and nonlinear transformations to obtain enhanced features; the enhanced features are then reconstructed into a sequence to obtain the main task decoder that outputs the predicted collision acceleration curve. The global feature vector is continuously compressed in dimension to obtain compressed features; the compressed features are then projected onto physical properties to obtain an auxiliary task decoder that outputs predicted physical features.

[0088] In practice, the main task decoder is responsible for outputting the predicted collision acceleration curves. Its processing flow is as follows: The input global feature vector F_global (e.g., 128-dimensional) is expanded in dimension by passing it through a first fully connected layer to a higher-dimensional (e.g., 256-dimensional) space, resulting in expanded feature H1. Feature recombination and nonlinear transformation are performed on expanded feature H1, and then passed through a second fully connected layer (maintaining 256 dimensions) and the ReLU activation function to obtain enhanced feature H2. Finally, sequence reconstruction is performed on enhanced feature H2, and it is mapped to a complete sequence of predicted acceleration curves through a third fully connected layer (output dimension T=1200, no activation function). (t).

[0089] For example, the main task decoder uses a three-layer fully connected network to perform a step-by-step transformation on F_global, mapping the 128-dimensional global features to a T=1200-dimensional acceleration sequence. The specific processing procedure is as follows: First-level linear transformation and activation processing: First, F_global is input into the first fully connected layer, which contains 256 neurons. The processing procedure is: H1 = ReLU(W1·F_global + b1). Where: W1∈ ^(256×128) is the first-layer weight matrix, which linearly maps the 128-dimensional features to a 256-dimensional space; B1∈ ^256 is the bias term; ReLU(·) = max(0, ·) is the nonlinear activation function, introducing nonlinear transformation capability; H1 ∈ ^256 represents the intermediate feature vector output from the first layer. This expands the 128-dimensional compressed features to 256 dimensions, providing a richer representation space for subsequent curve detail reconstruction.

[0090] Second-layer linear transformation and activation processing: H1 is input into the second fully connected layer, which also contains 256 neurons. The processing procedure is as follows: H2 = ReLU(W2·H1 + b2) Where: W2∈ ^(256×256) is the second-layer weight matrix, preserving a 256-dimensional feature space; b2∈ ^256 is the bias term; H2∈ ^256 represents the intermediate feature vector output from the second layer. Feature recombination and nonlinear transformation are performed within the same dimensional space to enhance the model's ability to fit complex curve shapes.

[0091] The third linear transformation (output layer): H2 is input into the third fully connected layer (output layer), which contains T=1200 neurons, corresponding to each time point of the acceleration curve. The processing procedure is: =W3·H2+b3.

[0092] Where: W3∈ ^(1200×256) is the output layer weight matrix, which maps 256-dimensional features to a 1200-dimensional output space; b3∈ ^1200 is the bias term; = [ 1, 2, ..., _T] ∈ ^T represents the final predicted collision acceleration curve. The output layer does not use an activation function to ensure that the predicted acceleration values ​​cover both positive and negative ranges and are not limited by the activation function's range.

[0093] The dimensionality changes during processing are: F_global (128 dimensions) → H1 (256 dimensions) → H2 (256 dimensions) → (1200 dimensions).

[0094] The auxiliary task decoder is responsible for outputting the predicted physical features. Its processing flow is as follows: the same global feature vector F_global is continuously compressed in dimension, and it is mapped to a lower dimension (e.g., 64-dimensional) space through the first fully connected layer to obtain the compressed feature Φ1; then, the compressed feature Φ1 is projected into physical attributes, and it is mapped to the predicted physical feature vector Φ_pred through the second fully connected layer (the output dimension is 18, corresponding to 18-dimensional physical features, with no activation function).

[0095] For example, the auxiliary task decoder uses a two-layer fully connected network to perform independent feature transformations on the same F_global, mapping the 128-dimensional global features to an 18-dimensional physical feature vector. The specific processing is as follows: First-level linear transformation and activation processing: F_global is input into the first fully connected layer of the auxiliary task, which contains 64 neurons. The processing procedure is: Φ1 = ReLU(W4 · F_global + b4).

[0096] Where: W4 ∈ ^(64×128) is the first-layer weight matrix, mapping the 128-dimensional features to 64 dimensions; b4 ∈ ^64 is the bias term; Φ1 ∈ ^64 represents the intermediate feature vector output from the first layer. A shallow network structure (64 dimensions) is used because the physical feature prediction task is relatively simple, and a shallow network is sufficient for modeling and can avoid overfitting.

[0097] The second linear transformation (output layer): Φ1 is input to the second fully connected layer (output layer), which contains 18 neurons, corresponding to each dimension of the 18-dimensional physical feature. The processing procedure is: Φ_pred = W5 · Φ1 + b5. Where: W5 ∈ ^(18×64) is the output layer weight matrix, which maps the 64-dimensional features to the 18-dimensional output space, realizing the projection of physical properties; b5 ∈ ^18 is the bias term; Φ_pred ∈ ^18 represents the predicted 18-dimensional physical feature vector, with each dimension corresponding one-to-one with the explicitly extracted physical feature Φ_raw.

[0098] The dimensional changes during processing are: F_global (128 dimensions) → Φ1 (64 dimensions) → Φ_pred (18 dimensions).

[0099] In practical applications, for a new vehicle design parameter (i.e., a new set of input feature vectors X_new), it is directly input into the trained optimal prediction model. The main task decoder of the optimal prediction model will output the new predicted collision acceleration curve. The new predicted collision acceleration curves can be used for subsequent collision safety performance assessments and structural optimization.

[0100] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0101] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0102] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a device for constructing a collision waveform prediction model.

[0103] refer to Figure 4 The apparatus for constructing the collision waveform prediction model includes: Input building module 10 is configured to: perform adaptive data augmentation on the real collision acceleration curve to obtain the augmented collision acceleration curve, and construct input features based on the collision response data corresponding to the real collision acceleration curve; The feature fusion module 20 is configured to: extract local transient features of the input features, extract explicit physical features from the enhanced collision acceleration curve, and fuse the local transient features and explicit physical features to obtain a physics-guided attention map; The prediction output module 30 is configured to: determine the global feature vector based on the attention map and local transient features, perform multi-task learning decoding on the global feature vector, and obtain the prediction model of the output predicted collision acceleration curve and predicted physical features.

[0104] Optionally, the prediction output module 30 is also configured as follows: The mean square error loss is determined based on the coordinate values ​​at the same moment between the predicted and actual collision acceleration curves. The dynamic time warping loss is determined based on the morphological similarity between the predicted and actual collision acceleration curves. The physical feature loss is determined based on the value of the same feature element between the predicted physical feature and the explicit physical feature; A loss function is constructed based on preset weighting coefficients, mean square error loss, dynamic time warping loss, and physical feature loss; The prediction model is iteratively optimized based on the loss function, and the target prediction model when the loss function converges is determined as the optimal prediction model.

[0105] Optionally, input building module 10 is also configured as follows: The time axis of the real collision acceleration curve is stretched or compressed according to the preset time warp factor to obtain the time torque collision acceleration curve. The amplitude of the time-torque collision acceleration curve is scaled according to a preset amplitude scaling factor to obtain an amplitude-scaled collision acceleration curve. The preset Gaussian white noise is superimposed on the amplitude-scaled collision acceleration curve to obtain the noise collision acceleration curve; The length of the filter window is determined based on the length of the noise collision acceleration curve, and the noise collision acceleration curve is smoothed according to the length of the filter window to obtain the enhanced collision acceleration curve.

[0106] Optionally, input building module 10 is also configured as follows: The calculated length is obtained by reducing and rounding the curve length and performing odd-number transformation. In response to the calculation length being less than or equal to the preset minimum threshold length, the calculation length is determined as the filter window length; In response to a calculated length greater than a preset minimum threshold length, the minimum threshold length is determined as the filter window length.

[0107] Optionally, the feature fusion module 20 is also configured to: Determine the time dimension of the local transient features, and broadcast the explicit physical features to the time dimension to obtain the broadcast explicit physical features; By splicing broadcast explicit physical features and local transient features along the channel dimension, spliced ​​features are obtained. Based on the preset convolutional layers, cross-channel information fusion and dimensionality reduction are performed on the spliced ​​features to obtain fused features. Each time step value of the fused features is mapped to a normalized interval to obtain an attention map.

[0108] Optionally, the prediction output module 30 is also configured as follows: The attention map and local transient features are multiplied element-wise to obtain the weighted features; By capturing the long-range dependencies of the time series of weighted features, feature representations are obtained; Global average pooling is performed on the feature representation to obtain the global feature vector.

[0109] Optionally, the prediction output module 30 is also configured as follows: The global feature vector is expanded in dimension to obtain extended features; the extended features are subjected to feature sufficiency and nonlinear transformations to obtain enhanced features; the enhanced features are then reconstructed into a sequence to obtain the main task decoder that outputs the predicted collision acceleration curve. The global feature vector is continuously compressed in dimension to obtain compressed features; the compressed features are then projected onto physical properties to obtain an auxiliary task decoder that outputs predicted physical features.

[0110] Optionally, the feature fusion module 20 is also configured to: Peak features are obtained by extracting the peak acceleration, peak occurrence time, peak half-width, and peak prominence from the enhanced collision acceleration curve; The total energy and the energy ratio of the first and second halves of the enhanced collision acceleration curve are extracted to obtain energy characteristics. The mean, standard deviation, skewness, and kurtosis of the enhanced collision acceleration curves are extracted to obtain statistical characteristics; Extract the average gradient and maximum gradient from the enhanced collision acceleration curve to obtain gradient features; Linear fitting is performed on the end of the enhanced collision acceleration curve to obtain the attenuation characteristics; By integrating peak characteristics, energy characteristics, statistical characteristics, gradient characteristics, and decay characteristics, explicit physical characteristics are obtained.

[0111] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0112] The apparatus of the above embodiments is used to implement the method for constructing the corresponding collision waveform prediction model in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0113] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for constructing the collision waveform prediction model described in any of the above embodiments.

[0114] Figure 5 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0115] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0116] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0117] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0118] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0119] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0120] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0121] The electronic devices described above are used to implement the method for constructing the corresponding collision waveform prediction model in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0122] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the collision waveform prediction model construction method as described in any of the above embodiments.

[0123] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0124] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the collision waveform prediction model construction method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0125] Based on the same concept, corresponding to the methods of any of the above embodiments, this application also provides a computer program product, including computer program instructions, which, when run on a computer, cause the computer to execute the collision waveform prediction model construction method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0126] It is understood that before using the technical solutions of the various embodiments in this application, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0127] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this application's technical solution.

[0128] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0129] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this application. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this application.

[0130] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0131] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0132] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0133] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the claims of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A method for constructing a collision waveform prediction model, characterized in that, include: Adaptive data augmentation is performed on the real collision acceleration curve to obtain an enhanced collision acceleration curve, and input features are constructed based on the collision response data corresponding to the real collision acceleration curve. Extract the local transient features of the input features, extract the explicit physical features from the enhanced collision acceleration curve, and fuse the local transient features and the explicit physical features to obtain a physics-guided attention map; The global feature vector is determined based on the attention map and the local transient features. The global feature vector is then decoded using multi-task learning to obtain a prediction model that outputs a predicted collision acceleration curve and predicted physical features.

2. The method for constructing a collision waveform prediction model according to claim 1, characterized in that, Also includes: The mean square error loss is determined based on the coordinate values ​​at the same moment between the predicted collision acceleration curve and the actual collision acceleration curve. The dynamic time warping loss is determined based on the morphological similarity between the predicted collision acceleration curve and the actual collision acceleration curve. The physical feature loss is determined based on the value of the same feature element between the predicted physical feature and the explicit physical feature; A loss function is constructed based on preset weighting coefficients, the mean square error loss, the dynamic time warping loss, and the physical feature loss; The prediction model is iteratively optimized based on the loss function, and the target prediction model when the loss function converges is determined as the optimal prediction model.

3. The method for constructing the collision waveform prediction model according to claim 1, characterized in that, The adaptive data augmentation of the real collision acceleration curve to obtain the enhanced collision acceleration curve includes: The time axis of the real collision acceleration curve is stretched or compressed according to a preset time distortion factor to obtain a time-torque collision acceleration curve. The time-torque collision acceleration curve is scaled according to a preset amplitude scaling factor to obtain an amplitude-scaled collision acceleration curve. The preset Gaussian white noise is superimposed on the amplitude-scaled collision acceleration curve to obtain the noise collision acceleration curve; The filter window length is determined based on the length of the noise collision acceleration curve, and the noise collision acceleration curve is smoothed based on the filter window length to obtain the enhanced collision acceleration curve.

4. The method for constructing the collision waveform prediction model according to claim 3, characterized in that, Determining the filter window length based on the curve length of the noise collision acceleration curve includes: The calculated length is obtained by reducing and rounding the curve length and performing odd number transformation. In response to the calculation length being less than or equal to a preset minimum threshold length, the calculation length is determined as the filter window length; In response to the calculated length being greater than a preset minimum threshold length, the minimum threshold length is determined as the filter window length.

5. The method for constructing a collision waveform prediction model according to claim 1, characterized in that, The process of fusing the local transient features and the explicit physical features to obtain a physics-guided attention map includes: Determine the time dimension of the local transient feature, and broadcast the explicit physical feature to the time dimension to obtain the broadcast explicit physical feature; The broadcast explicit physical features and the local transient features are spliced ​​together along the channel dimension to obtain the spliced ​​features; The spliced ​​features are fused and dimensionality reduced across channels using a preset convolutional layer to obtain fused features. Each time step value of the fused features is then mapped to a normalized interval to obtain the attention map.

6. The method for constructing a collision waveform prediction model according to claim 1, characterized in that, The step of determining the global feature vector based on the attention map and the local transient features includes: The attention map and the local transient features are multiplied element-wise to obtain the weighted features; By capturing the long-range dependencies of the time series of the weighted features, feature representations are obtained; The feature representation is subjected to global average pooling to obtain the global feature vector.

7. The method for constructing a collision waveform prediction model according to claim 1, characterized in that, The prediction model includes a main task decoder and an auxiliary task decoder; the prediction model that performs multi-task learning and decoding on the global feature vector to obtain the output predicted collision acceleration curve and predicted physical features includes: The global feature vector is expanded in dimension to obtain extended features; the extended features are subjected to feature sufficiency and nonlinear transformation to obtain enhanced features; the enhanced features are reconstructed into a sequence to obtain the main task decoder that outputs the predicted collision acceleration curve; The global feature vector is continuously compressed in dimension to obtain compressed features; the compressed features are then projected onto physical attributes to obtain the auxiliary task decoder that outputs the predicted physical features.

8. The method for constructing a collision waveform prediction model according to claim 1, characterized in that, The extraction of explicit physical features from the enhanced collision acceleration curve includes: The peak acceleration, peak occurrence time, peak half-width, and peak prominence of the enhanced collision acceleration curve are extracted to obtain the peak characteristics; The total energy and the energy ratio of the first and second halves of the enhanced collision acceleration curve are extracted to obtain the energy characteristics. The mean, standard deviation, skewness, and kurtosis of the enhanced collision acceleration curve are extracted to obtain statistical characteristics; Extract the average gradient and maximum gradient of the enhanced collision acceleration curve to obtain gradient features; Linear fitting is performed on the end of the enhanced collision acceleration curve to obtain the attenuation characteristics; The explicit physical characteristics are obtained by integrating the peak characteristics, energy characteristics, statistical characteristics, gradient characteristics, and attenuation characteristics.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 8.

Images