Vehicle collision strength determination method and device, electronic equipment and storage medium

Abstract

The present application relates to a kind of vehicle crash intensity determination method, device, electronic equipment and storage medium, wherein vehicle crash intensity determination method includes: the acceleration sequence data collected in the process of vehicle collision is acquired;Energy domain feature and time domain feature in the process of vehicle collision are determined based on acceleration sequence data;Energy domain feature and time domain feature are fused, and fusion feature is obtained;Vehicle crash intensity information is determined based on fusion feature.The application embodiment determines energy domain feature and time domain feature based on the acceleration sequence data collected in the process of vehicle collision, energy domain feature and time domain feature are fused, and fusion feature is obtained, vehicle crash intensity information is determined based on the fusion feature, the macroscopic release mode in the process of vehicle collision is embodied by energy domain feature, the reserved microscopic detail feature in the process of vehicle collision is embodied by time domain feature, and the accuracy of vehicle crash intensity information determination is improved.

Description

Technical Field

[0001] This application relates to the field of automotive passive safety technology, and in particular to a method, device, electronic device and storage medium for determining vehicle collision intensity. Background Technology

[0002] With the rapid development of automotive safety technology, vehicle restraint systems are evolving from traditional single-stage fixed parameters to multi-stage adaptive parameters. The widespread adoption of intelligent hardware such as variable stiffness airbags, variable force limit seat belts, and adaptive seats has laid the foundation for personalized and precise occupant protection. However, existing restraint system control strategies still primarily employ a "single ignition + fixed parameters" model, which cannot dynamically adjust according to real-time changes in collision intensity.

[0003] To address this issue, the industry has proposed a multiple ignition technique, which involves adjusting the ignition after the initial ignition based on the predicted remaining collision intensity. However, existing collision intensity prediction technologies face the following key technical challenges:

[0004] First, the duration of the first ignition ( The issue of differences. Under different collision conditions, Significant differences exist: frontal collisions with hard barriers typically take 10-15 milliseconds, while frontal collisions with soft barriers take 15-25 milliseconds, and offset collisions... The duration difference is longer. This variation in duration leads to inconsistent input signal lengths, which traditional fixed-length signal processing methods struggle to handle effectively.

[0005] Secondly, there is the problem of lack of physical consistency. Purely data-driven prediction methods lack physical constraints and are prone to producing prediction results that do not conform to the laws of physics. For example, the combination of predicted peak acceleration and arrival time may violate the law of conservation of energy, which lacks credibility in the field of security.

[0006] Finally, there's the issue of stringent real-time requirements. The in-vehicle environment demands... The time window for triggering the secondary ignition decision is usually only about 5 milliseconds, and the algorithm must complete high-precision prediction under the limited computing resources of the vehicle ECU.

[0007] Existing solutions mainly include threshold triggering methods, traditional signal processing methods, machine learning methods, and deep learning methods, but all have significant shortcomings: threshold methods have limited accuracy, can only make qualitative judgments, and can only identify a limited number of collision intensity levels, typically used to solve standard working conditions; traditional signal processing methods are sensitive to noise and have difficulty handling complex nonlinear relationships; machine learning methods require complex feature engineering; and while deep learning methods have strong feature extraction capabilities, they are limited in their application to... It has poor adaptability to duration differences and lacks physical constraints.

[0008] Therefore, there is an urgent need for a new technical solution to systematically solve the above technical problems and achieve high-precision, high-real-time, and physically reasonable residual collision strength prediction. SUMMARY

[0009] To solve the above technical problems or at least partially solve the above technical problems, the present application provides a vehicle collision strength determination method and device, electronic equipment and storage medium.

[0010] In a first aspect, the present application provides a vehicle collision strength determination method, comprising:

[0011] obtaining acceleration sequence data collected during a vehicle collision process;

[0012] determining energy domain features and time domain features during the vehicle collision process based on the acceleration sequence data;

[0013] performing feature fusion on the energy domain features and the time domain features to obtain fusion features;

[0014] determining vehicle collision strength information based on the fusion features.

[0015] Optionally, obtaining acceleration sequence data collected during a vehicle collision process comprises:

[0016] obtaining original acceleration data collected at multiple time points;

[0017] determining time points at which a first preset number of consecutive composite accelerations exceed a preset acceleration threshold as a collision start time point based on the original acceleration data;

[0018] obtaining a firing time point at which an airbag controller is fired for the first time;

[0019] determining acceleration data of multiple time points between the collision start time point and the firing time point in the original acceleration data as the acceleration sequence data.

[0020] Optionally, determining energy domain features and time domain features during the vehicle collision process based on the acceleration sequence data comprises:

[0021] performing reference offset correction processing on the acceleration sequence data to obtain corrected acceleration sequence data;

[0022] determining cumulative collision energy based on the corrected acceleration sequence data;

[0023] performing equal-energy-interval resampling on the corrected acceleration sequence data based on the cumulative collision energy to obtain a resampled acceleration vector;

[0024] Energy domain features in the resampled acceleration vector are extracted using an energy domain feature extraction model.

[0025] The time-domain features in the corrected acceleration sequence data are extracted using a time-domain feature extraction model.

[0026] Optionally, determining the cumulative collision energy based on the corrected acceleration sequence data includes:

[0027] Obtain the equivalent collision mass of the vehicle;

[0028] Based on the corrected acceleration sequence data, determine the triaxial velocity sequence data;

[0029] The instantaneous power at multiple moments is calculated based on the equivalent collision mass, the triaxial velocity sequence data, and the corrected acceleration sequence data.

[0030] The collision energy sequence is determined based on the instantaneous power at multiple moments;

[0031] The cumulative collision energy is determined by summing multiple collision energy values ​​in the collision energy sequence.

[0032] Optionally, the acceleration sequence data is resampled at equal energy intervals based on the accumulated collision energy to obtain resampled sequence data, including:

[0033] The accumulated collision energy is divided equally according to a second preset number to obtain a sequence of equally spaced energies;

[0034] In the cumulative energy curve corresponding to the collision energy sequence, find multiple adjacent collision energies that are greater than the equal interval energy and less than or equal to the equal interval energy for each group of equally spaced energies.

[0035] Obtain multiple adjacent moments of collision energy;

[0036] The acceleration sequence data is resampled according to multiple sets of adjacent time points to obtain resampled sequence data.

[0037] Optionally, the energy domain features and the time domain features are fused to obtain fused features, including:

[0038] The energy domain features and the time domain features are concatenated to obtain the concatenated features;

[0039] Attention weights are determined based on the splicing features;

[0040] Calculate the complementary weights of the attention weights;

[0041] The fusion feature is determined based on the attention weight, the energy domain feature, the complementary weight, and the time domain feature.

[0042] Optionally, determining vehicle collision intensity information based on the fused features includes:

[0043] The fused features are input into the peak acceleration prediction head to obtain the remaining peak acceleration;

[0044] The fused features are input into the peak arrival time prediction head to obtain the peak arrival time;

[0045] The remaining peak acceleration and the peak arrival time are determined as the vehicle collision intensity information.

[0046] Optionally, the method further includes:

[0047] During the training phase, the vehicle's current speed and current acceleration are obtained at the moment of ignition when the airbag controller is first ignited;

[0048] The constraint energy is calculated based on the current speed, current acceleration, vehicle collision intensity information, and a preset constraint energy formula.

[0049] Obtain the true remaining energy calculated from complete crash test data;

[0050] The physical constraint loss is determined based on the constraint energy, the actual remaining energy, and the physical constraint coefficient.

[0051] The loss function is determined based on the physical constraint loss.

[0052] Optionally, after determining the vehicle collision intensity information based on the fused features, the method further includes:

[0053] Obtain the vehicle's current speed and current acceleration at the moment of ignition when the airbag controller first ignites;

[0054] The predicted energy is calculated based on the current speed, current acceleration, vehicle collision intensity information, and a preset constraint energy formula.

[0055] Obtain the true remaining energy calculated from complete crash test data;

[0056] Determine the relative error between the predicted energy and the actual remaining energy;

[0057] If the relative error is greater than a preset error threshold, the vehicle collision intensity information is corrected to obtain corrected collision intensity information.

[0058] Optionally, the vehicle collision intensity information is corrected to obtain corrected collision intensity information, including:

[0059] The correction coefficient is determined based on the relative error and the preset correction coefficient calculation formula;

[0060] The vehicle collision intensity information is corrected based on the correction coefficient to obtain corrected collision intensity information.

[0061] Optionally, the vehicle collision intensity information is corrected based on the correction coefficient to obtain corrected collision intensity information, including:

[0062] If the predicted energy is greater than the actual remaining energy, the remaining peak acceleration in the vehicle collision intensity information is reduced based on the correction coefficient.

[0063] Optionally, the vehicle collision intensity information is corrected based on the correction coefficient to obtain corrected collision intensity information, including:

[0064] If the predicted energy is less than or equal to the actual remaining energy, the peak arrival time in the vehicle collision intensity information is extended based on the correction coefficient.

[0065] Secondly, this application provides a vehicle collision intensity determination device, comprising:

[0066] The acquisition module is used to acquire acceleration sequence data collected during the vehicle collision process;

[0067] The first determining module is used to determine the energy domain characteristics and time domain characteristics of the vehicle collision process based on the acceleration sequence data;

[0068] The fusion module is used to fuse the energy domain features and the time domain features to obtain fused features;

[0069] The second determining module is used to determine vehicle collision intensity information based on the fused features.

[0070] Thirdly, this application provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0071] Memory, used to store computer programs;

[0072] The processor, when executing a program stored in memory, implements the vehicle collision intensity determination method described in any of the first aspects.

[0073] Fourthly, this application provides a computer-readable storage medium storing a program for a vehicle collision intensity determination method, wherein when the program for the vehicle collision intensity determination method is executed by a processor, it implements the steps of the vehicle collision intensity determination method described in any of the first aspects.

[0074] The beneficial effects of this invention are:

[0075] This application embodiment determines the energy domain features and time domain features of a vehicle collision process based on acceleration sequence data collected during the collision process. The energy domain features and time domain features are then fused to obtain fused features. Based on the fused features, the vehicle collision intensity information is determined. The energy domain features reflect the macroscopic release pattern during the vehicle collision process, while the time domain features reflect the microscopic detail features preserved during the vehicle collision process, thereby improving the accuracy of determining the vehicle collision intensity information. Attached Figure Description

[0076] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 A flowchart illustrating a method for determining vehicle collision intensity provided in this application embodiment;

[0079] Figure 2 A flowchart illustrating a method for determining vehicle collision intensity in a practical application, provided as an embodiment of this application;

[0080] Figure 3 This application provides a schematic diagram illustrating the instantaneous power variation trend corresponding to the accumulated collision energy in a short TTF interval.

[0081] Figure 4 This application provides a schematic diagram illustrating the instantaneous power variation trend corresponding to the accumulated collision energy within a long TTF range.

[0082] Figure 5 This application provides a schematic diagram illustrating the variation trend of accumulated collision energy in a short TTF range.

[0083] Figure 6 This application provides a schematic diagram illustrating the variation trend of accumulated collision energy over a long TTF range.

[0084] Figure 7 This is a schematic diagram of the original acceleration variation curve under collision conditions corresponding to a short TTF signal, provided in an embodiment of this application.

[0085] Figure 8 This is a schematic diagram of the original acceleration variation curve under collision conditions corresponding to a long TTF signal, provided in an embodiment of this application.

[0086] Figure 9 A different embodiment provided for this application A diagram illustrating how forced alignment disrupts physical timing characteristics;

[0087] Figure 10 A schematic diagram illustrating the principle of an equal-energy interval resampling interval provided in this application embodiment;

[0088] Figure 11 This application provides a schematic diagram illustrating an example of energy interpolation using equal-energy-interval resampling.

[0089] Figure 12 This is a schematic diagram of the acceleration interpolation result corresponding to an example of energy interpolation with equal energy interval resampling provided in an embodiment of this application;

[0090] Figure 13 A schematic diagram of resampling results with equal energy intervals provided in an embodiment of this application;

[0091] Figure 14 An algorithm framework diagram of a dual-domain architecture is provided for an embodiment of this application;

[0092] Figure 15 A flowchart illustrating a method for determining vehicle collision intensity provided in this application embodiment;

[0093] Figure 16 A schematic diagram comparing training loss convergence curves with and without physical constraints is provided for an embodiment of this application.

[0094] Figure 17 This application provides a schematic diagram comparing the verification accuracy variation curves with and without physical constraints, as part of an embodiment of the present application.

[0095] Figure 18 A flowchart illustrating a method for determining vehicle collision intensity provided in this application embodiment;

[0096] Figure 19 This application provides a schematic diagram comparing the physical consistency excellence rate in a comparison of physical constraint improvement effects.

[0097] Figure 20This application provides a schematic diagram comparing the physical consistency pass rate in a comparison of physical constraint improvement effects.

[0098] Figure 21 This application provides a schematic diagram illustrating the comparison of average relative errors in the evaluation of physical constraint improvement effects.

[0099] Figure 22 This application provides a schematic diagram illustrating the extreme condition adaptability comparison in a comparison of physical constraint improvement effects.

[0100] Figure 23 This is a structural diagram of a vehicle collision intensity determination device provided in an embodiment of this application;

[0101] Figure 24 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0102] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0103] First, before detailing the technical solution of this application, a brief overview of the background of this technology is provided. Automotive passive safety technology has evolved from fixed parameters to adaptive parameters, but existing collision intensity prediction technologies still have significant shortcomings. The main challenge lies in: under different collision conditions... Significant differences in duration, unreasonable prediction results due to a lack of physical constraints, and stringent real-time requirements of the in-vehicle environment. Based on these factors, the applicant, through long-term research and experimental verification, proposes a method, device, electronic device, and storage medium for determining vehicle collision intensity in this application. It should be noted that the algorithm architecture in this application is the result of extensive creative work by the applicant, and it not only handles multi-source time-series data well but can also serve as a general paradigm for neural network models that fuse sequential data and physical constraints.

[0104] This application provides a method for determining vehicle collision intensity, such as... Figure 1 As shown, it includes:

[0105] Step S11: Obtain acceleration sequence data collected during the vehicle collision process;

[0106] In this embodiment of the application, acceleration sequence data may refer to acceleration sequence data collected by a triaxial accelerometer at multiple moments.

[0107] In one embodiment of this application, step S11, acquiring acceleration sequence data collected during a vehicle collision, includes: acquiring raw acceleration data collected at multiple moments; determining, based on the raw acceleration data, moments when the preceding first preset number of composite accelerations all exceed a preset acceleration threshold, as the collision start moment; acquiring the ignition moment of the first ignition of the airbag controller; and determining the acceleration data at multiple moments in the raw acceleration data whose acquisition time is located between the collision start moment and the ignition moment as the acceleration sequence data.

[0108] In practical applications, such as Figure 2 Step 101: Obtain the triaxial acceleration sequence data during the vehicle collision process. This can be achieved using an onboard triaxial accelerometer at a sampling frequency. For example, collision signals can be acquired at a sampling frequency of 2kHz to obtain acceleration data from the moment of collision initiation to the moment of first ignition.

[0109]

[0110] in, The first Acceleration components in the X, Y, and Z directions at each sampling time. This represents the total number of sampling points, expressed in g (gravitational acceleration).

[0111] The collision initiation time detection uses a threshold algorithm:

[0112]

[0113] When the triaxial composite acceleration exceeds 5 consecutive sampling points When the threshold is reached, a collision is considered to have begun. The timing is determined by the ignition algorithm of the airbag ECU, typically based on a comprehensive criterion including acceleration integral, peak value judgment, and multi-sensor fusion. The technical advantages of this detection method include: adaptive threshold design to suit different vehicle characteristics; effective filtering of transient interference to meet duration requirements; and a detection delay of less than 0.5 milliseconds to meet real-time requirements. In implementation, acceleration data is directly acquired from sensors; collision detection parameters can be calibrated according to different vehicle models; signal quality verification includes signal-to-noise ratio detection (required ≥20dB), saturation detection, and outlier handling.

[0114] Step S12: Determine the energy domain characteristics and time domain characteristics of the vehicle collision process based on the acceleration sequence data;

[0115] In one embodiment of this application, step S12, which determines the energy domain characteristics and time domain characteristics of the vehicle collision process based on the acceleration sequence data, includes:

[0116] 1. Perform reference offset correction processing on the acceleration sequence data to obtain corrected acceleration sequence data;

[0117] In practical applications, to eliminate the effects of sensor zero drift, one can, as follows: Figure 2 As shown in step S102, the acceleration signal is preprocessed and velocity integrated using a reference offset correction method:

[0118]

[0119]

[0120] in, This is the corrected acceleration signal, i.e., the corrected acceleration sequence data. The baseline offset vector is obtained by statistically analyzing the acceleration data of N sampling points before the collision, adapting to different sensor characteristics and installation orientations.

[0121] 2. Determine the cumulative collision energy based on the aforementioned corrected acceleration sequence data;

[0122] The determination of cumulative collision energy based on the corrected acceleration sequence data includes:

[0123] (1) Obtain the equivalent collision mass of the vehicle, which can be calculated based on existing technology;

[0124] (2) Determine the triaxial velocity sequence data based on the corrected acceleration sequence data;

[0125] For the corrected acceleration sequence data, an improved accumulation method is used to calculate the triaxial velocities, resulting in the triaxial velocity sequence data. :

[0126]

[0127] in, The sampling interval is... This is the corrected acceleration signal, i.e., the corrected acceleration sequence data.

[0128] (3) Calculate the instantaneous power at multiple moments based on the equivalent collision mass, the triaxial velocity sequence data, and the corrected acceleration sequence data;

[0129] It is possible Figure 2 As shown in step 103, the instantaneous power is calculated before calculating the cumulative collision energy. :

[0130]

[0131] Where m is the equivalent collision mass. Represents the L2 norm, For the first Instantaneous power at each sampling time, This is a three-axis velocity sequence data.

[0132] (4) Determine the collision energy sequence based on the instantaneous power at multiple moments;

[0133] Arrange the instantaneous power at multiple moments in chronological order to obtain the collision energy sequence.

[0134] (5) The cumulative collision energy is determined by summing up multiple collision energy values ​​in the collision energy sequence.

[0135] It is possible Figure 2 As shown in step 103, the cumulative collision energy is calculated and normalized coordinates are constructed.

[0136] The cumulative collision energy is calculated using the instantaneous power formula, such as... Figure 3 , Figure 4 , Figure 5 and Figure 6 As shown:

[0137]

[0138] Where m is the equivalent collision mass. Represents the L2 norm, For the first Instantaneous power at each sampling time, To accumulate collision energy.

[0139] This power calculation method is based on rigorous physical principles. Instantaneous power reflects the rate at which the collision force does work, while accumulated energy characterizes the loss of kinetic energy of the vehicle and is directly related to the degree of structural deformation.

[0140] 3. Based on the accumulated collision energy, the corrected acceleration sequence data is resampled at equal energy intervals to obtain a resampled acceleration vector;

[0141] Based on the accumulated collision energy, the acceleration sequence data is resampled at equal energy intervals to obtain resampled sequence data, including:

[0142] (1) Divide the accumulated collision energy into equal portions according to a second preset number to obtain a sequence of equally spaced energies;

[0143] It is possible Figure 2 As shown in step S104, the equal-energy-interval resampling algorithm is first executed.Figure 7 and Figure 8 As shown, the timing of the first ignition differs under different collision conditions. If alignment is achieved in the time domain through supplementary methods (…),… Figure 9 As shown, this will destroy the physical characteristics of the acceleration curve, leading to changes and loss of information in the finite-length acceleration curve. Therefore, this technique proposes an equal-energy-interval resampling method.

[0144] like Figure 10 As shown, the original acceleration signal is resampled into a fixed-length sequence through equal-energy-interval interpolation. For the first ignition moment Corresponding total energy Divide into M equal parts (128 in this example) to obtain a sequence of equally spaced energies.

[0145] (2) In the cumulative energy curve corresponding to the collision energy sequence, find multiple sets of adjacent collision energies that are greater than the equal interval energy and less than or equal to the equal interval energy for each set of equal interval energy;

[0146] Based on equally spaced energy sequences, construct equally spaced target energy sequences:

[0147]

[0148] in, For the first ignition moment The corresponding energy, For the j-th target energy value, The length of the resampled sequence is chosen based on the following considerations: maintaining sufficient temporal resolution, balancing information preservation and computational efficiency, and adapting to the typical input size of neural networks.

[0149] For each target energy In the original cumulative energy curve We search for adjacent moments that satisfy the conditions to obtain multiple sets of adjacent collision energies:

[0150] ≤ <

[0151] (3) Obtain multiple adjacent times of multiple adjacent collision energies, that is, obtain two time points of each adjacent collision energy.

[0152] (4) The acceleration sequence data is resampled according to multiple adjacent time points to obtain resampled sequence data.

[0153] like Figure 11 , Figure 12 , Figure 13As shown, the resampled acceleration value is calculated using linear interpolation:

[0154]

[0155]

[0156]

[0157] in, For interpolation weights, This is the resampled acceleration vector.

[0158] These energy resampling methods ensure the uniformity of energy distribution, the preservation of signal characteristics, and the comparability under different operating conditions. By converting time-domain signals into energy-domain signals, they solve the problems under different collision conditions. The problem of inconsistent input signal length caused by time difference.

[0159] IV. Extract the energy domain features from the resampled acceleration vector using the energy domain feature extraction model;

[0160] like Figure 2 Step S105 in the process, such as Figure 14 As shown, a dual-path deep neural network is constructed for feature extraction. The dual-path network includes an energy domain path and a time domain path. The convolutional neural network in the energy domain path employs a multi-scale receptive field design, capable of simultaneously capturing local and global energy release pattern features. The recurrent neural network in the time domain path uses a bidirectional processing mechanism, utilizing both forward and backward temporal information for feature extraction, enabling parallel processing of information from both domains to achieve feature complementarity. The energy domain path focuses on macroscopic physical pattern recognition, while the time domain path preserves microscopic temporal details. Parallel processing improves efficiency, and feature fusion achieves information complementarity. It features parallel processing to improve computational efficiency by leveraging the advantages of multi-core processors; feature complementarity fully utilizes multi-domain information to improve prediction accuracy; modular design facilitates optimization and debugging, supporting incremental improvements; and a flexible architecture supports different network configurations and parameter adjustments.

[0161] Energy domain path design: The energy domain path uses multi-layer dilated convolution to process the resampled signal. Layer 1: Dilation coefficient = 1, kernel size = 3, number of channels = 64; Layer 2: Dilation coefficient = 2, kernel size = 3, number of channels = 128; Layer 3: Dilation coefficient = 4, kernel size = 3, number of channels = 256; Layer 4: Dilation coefficient = 8, kernel size = 3, number of channels = 512. Each layer is followed by layer normalization, GELU activation, and dropout of 0.1. By expanding the receptive field through dilated convolutions without increasing the number of parameters, energy release patterns at different scales can be captured.

[0162] The energy domain path employs multi-layer dilated convolutions with dilation coefficients of 1, 2, 4, and 8 to capture energy release patterns at different time scales. Each dilated convolution layer is followed by layer normalization, GELU activation function, and dropout regularization to improve training stability.

[0163] 5. Use a time-domain feature extraction model to extract the time-domain features from the corrected acceleration sequence data.

[0164] Time domain path design: Bidirectional LSTM is used to process the raw time-series signals. Number of hidden units: 256. Bidirectional LSTM can utilize both forward and backward information simultaneously to effectively capture the temporal dependencies of the collision process.

[0165] The time-domain path employs a right-aligned padding strategy for input sequences of different lengths, maintaining signal end alignment and padding the beginning with 0 values. Bidirectional LSTM can simultaneously utilize forward and backward timing information, effectively capturing high-frequency timing details.

[0166] Step S13: Perform feature fusion on the energy domain features and the time domain features to obtain fused features;

[0167] In this step, dual-path features can be integrated through a multi-layer feature fusion mechanism, including feature concatenation, nonlinear transformation, and attention weight calculation, to achieve effective fusion of information from the two domains.

[0168] In one embodiment of this application, feature fusion of the energy domain features and the time domain features to obtain fused features includes: feature splicing of the energy domain features and the time domain features to obtain spliced ​​features; determining attention weights based on the spliced ​​features; calculating complementary weights of the attention weights; and determining the fused features based on the attention weights, the energy domain features, the complementary weights, and the time domain features.

[0169] like Figure 2 In step S106, the multi-layer feature fusion mechanism is implemented. Assume the energy domain feature is... The time domain features are The fusion process is as follows:

[0170]

[0171]

[0172]

[0173] in, Indicates feature splicing, For the sigmoid function, and For learnable parameters, This represents element-wise multiplication. Attention weights. The adaptive adjustment of the contribution of features from both domains enhances model adaptability. The adaptive trade-off attention mechanism automatically adjusts the importance of the two domains based on the input. Information complementarity fully utilizes the complementary characteristics of the energy and time domains; the energy domain provides a macroscopic model, while the time domain retains detailed information. Interpretable attention weights provide transparency to the decision-making process, facilitating engineering analysis and debugging.

[0174] Step S14: Determine vehicle collision intensity information based on the fused features.

[0175] In one embodiment of this application, step S14, which determines vehicle collision intensity information based on the fused features, includes: inputting the fused features into a peak acceleration prediction head to obtain the remaining peak acceleration; inputting the fused features into a peak arrival time prediction head to obtain the peak arrival time; and determining the remaining peak acceleration and the peak arrival time as the vehicle collision intensity information.

[0176] like Figure 2 In step S107, multi-task joint prediction is performed. Based on fused features... The remaining peak acceleration is output through two prediction heads. and peak arrival time .

[0177] After determining the vehicle collision intensity information, restraint system control commands can be generated and communicated with the airbag ECU and seat belt controller.

[0178] This application embodiment determines the energy domain and time domain features of a vehicle collision based on acceleration sequence data collected during the collision process. It then fuses these energy domain and time domain features to obtain fused features, and determines the vehicle collision intensity information based on these fused features. The energy domain features reflect the macroscopic release pattern during the collision, while the time domain features reflect the retention of microscopic details, improving the accuracy of collision intensity determination. Furthermore, the duration of collision intensity determination in this application meets the real-time requirements of in-vehicle systems. The time from triggering the output prediction result is no more than 3 milliseconds.

[0179] In yet another embodiment of this application, as Figure 15 As shown, the method further includes:

[0180] Step S21: During the training phase, obtain the vehicle's current speed and current acceleration at the moment of ignition when the airbag controller is first ignited.

[0181] Step S22: Calculate the constraint energy based on the current speed, current acceleration, vehicle collision intensity information, and a preset constraint energy formula;

[0182] Step S23: Obtain the actual remaining energy calculated from the complete collision test data;

[0183] Step S24: Determine the physical constraint loss based on the constraint energy, the actual remaining energy, and the physical constraint coefficient;

[0184] Step S25: Determine the loss function based on the physical constraint loss.

[0185] Purely data-driven prediction methods are prone to producing results that violate the laws of physics. This method ensures the rationality of the prediction results through a physical constraint mechanism based on the change of residual kinetic energy. The theoretical basis of the physical constraints is: establishing the kinetic energy relationship between the current state and the predicted state based on the kinetic energy theorem; constraining the rationality of the prediction parameters through the change of residual kinetic energy; and using a phased constraint strategy to adapt to different needs of training and inference.

[0186] At the moment of first ignition The vehicle already has its current speed. and acceleration The remaining collision process begins from the current state, with the predicted peak acceleration... In time It will be reached later. Assume that during the remaining collision process, the acceleration will decrease from the current value. Change to peak If simplified to a linear process, then the acceleration at each moment... :

[0187]

[0188] Since acceleration changes linearly, therefore in time... Average acceleration within :

[0189]

[0190] In time Inside, the change in velocity is:

[0191]

[0192] Therefore, the change in energy:

[0193]

[0194] Therefore, the constraint energy formula is:

[0195]

[0196] The first term represents the contribution of the product of current momentum and average acceleration, and the second term represents the contribution of the change in acceleration itself. This constraint is based on a rigorous derivation of the kinetic energy theorem and has a clear physical meaning.

[0197] During the training phase, physical constraints are introduced into the loss function:

[0198]

[0199]

[0200] in, To accelerate loss prediction, To predict losses over time, For the actual remaining energy, For physical consistency loss, The physical constraint coefficient is a weighting coefficient that is dynamically adjusted using a dynamic adjustment strategy. This represents the loss due to physical constraints.

[0201]

[0202] , To achieve a gradual training cycle, dynamic weights avoid convergence difficulties caused by excessive constraints in the early training phase, while strengthening physical consistency requirements in the later training phase; smooth weight changes prevent training instability. For example... Figure 16 and Figure 17 The diagram shows the loss function during the training phase. Physical constraints can help the model converge faster and reduce the loss.

[0203] In another embodiment of this application, after determining the vehicle collision intensity information based on the fused features in step S14, as follows: Figure 18 As shown, the method further includes:

[0204] Step S31: Obtain the vehicle's current speed and current acceleration at the moment of ignition when the airbag controller is first ignited;

[0205] Step S32: Calculate the predicted energy based on the current speed, current acceleration, vehicle collision intensity information, and a preset constraint energy formula;

[0206] Step S33: Obtain the actual remaining energy calculated from the complete collision test data;

[0207] Step S34: Determine the relative error between the predicted energy and the actual remaining energy;

[0208] Step S35: If the relative error is greater than a preset error threshold, the vehicle collision intensity information is corrected to obtain corrected collision intensity information.

[0209] In one embodiment of this application, step S35 corrects the vehicle collision intensity information to obtain corrected collision intensity information, including: determining a correction coefficient based on the relative error and a preset correction coefficient calculation formula; and correcting the vehicle collision intensity information based on the correction coefficient to obtain corrected collision intensity information.

[0210] The process of correcting the vehicle collision intensity information based on the correction coefficient to obtain corrected collision intensity information includes:

[0211] If the predicted energy is greater than the actual remaining energy, the remaining peak acceleration in the vehicle collision intensity information is reduced based on the correction coefficient.

[0212] If the predicted energy is less than or equal to the actual remaining energy, the peak arrival time in the vehicle collision intensity information is extended based on the correction coefficient.

[0213] In practical applications, the prediction results can be physically consistent with changes in residual kinetic energy. At the moment of first ignition... The vehicle has a current speed and acceleration Calculate the constraint energy corresponding to the prediction parameters:

[0214]

[0215] in, The speed at the moment of first ignition. For the acceleration at the moment of first ignition, To predict peak acceleration, To predict the peak arrival time, the training phase will... Compared with the actual remaining energy calculated from complete crash test data The comparison is performed, and a correction mechanism is triggered during the inference phase when the relative error exceeds 20%.

[0216] It is worth noting that, in practical operation, this application can predict the evolution of collision intensity at each moment within a future time window. The advantage of doing so is that detailed timing information of the collision process can be obtained, providing rich control inputs for multi-level constraint systems.

[0217] When the physical consistency error of the prediction results exceeds a threshold, a soft correction mechanism is triggered:

[0218]

[0219] if Then the correction factor is:

[0220]

[0221] The initial value is 0.88, meaning the model will still make some corrections to the predictions even in the absence of error. This setting ensures the conservatism of the correction, preventing the model from over-adjusting the acceleration and causing unnecessary risks. As the error increases, It will change accordingly; when the error is less than 0.20, There will be some increase, allowing for more adjustments. Conversely, a larger error will guide the model to adjust more carefully. If the predicted energy is higher than the actual value, the peak acceleration calculated after the update will tend to be smaller, in order to reduce overly high energy predictions; if the predicted energy is lower than the actual value, the model will increase the time prediction to give the vehicle more time to reach the target state.

[0222] Predicted energy too high ( Prioritize reducing peak acceleration:

[0223]

[0224] Otherwise, if the predicted energy is too low, the arrival time should be extended first.

[0225]

[0226] To evaluate the effectiveness of the improved physical constraints, this invention introduces a quantitative evaluation index system. These indexes include the following:

[0227] First, the consistency excellence rate This metric measures the proportion of predictions with an error ratio less than 0.08. It reflects the model's effective predictive ability within a small error range, and a high consistency rate indicates the model's reliability. .

[0228] Secondly, the consistency pass rate This indicates the proportion of predictions with an error ratio less than 0.20. This metric is used to evaluate the model's ability to provide reasonable predictions in most cases; a high consistency pass rate means that the model still has good predictive performance over a wide error range. .

[0229] In addition, the average relative error It provides the average relative error between the predicted energy and the actual energy, reflecting the overall accuracy of the model. A smaller average relative error means the model is more accurate in energy estimation. .

[0230] Finally, constraint validity This measures the rate of change of the average relative error after introducing physical constraints. A positive value indicates an improvement in model performance after introducing constraints, thereby enhancing credibility. .

[0231] By comprehensively evaluating the above indicators, the actual application effect of the improved physical constraint mechanism can be effectively determined.

[0232] In practical implementation, assume the following parameters for a collision scenario: current velocity 3.2 m / s, current acceleration 15.6g, predicted peak acceleration 28.4g, predicted arrival time 4.2 ms, and vehicle mass 1600 kg. In this scenario, the calculated constraint energy is 5654 J. This numerical example demonstrates that the new constraint formula can provide a reasonable energy estimate.

[0233] In terms of algorithm performance evaluation, the improved physical constraints showed a significant performance improvement on 12,880 test data points, such as... Figure 19 , Figure 20 , Figure 21 and Figure 22 As shown, the improved constraint method has significantly improved in all indicators, especially in terms of physical consistency and adaptability to extreme conditions.

[0234] Through the above modifications, this invention successfully resolves the logical problems of the original physical constraint formula, providing a more rigorous and accurate physical consistency constraint mechanism. The constraint formula considers the current collision state, is derived based on a rigorous kinetic energy theorem, and has clear physical meaning, thus providing a reliable physical guarantee for predicting the remaining collision intensity of the vehicle.

[0235] The physical consistency constraints include constraint loss during the training phase and soft correction during the inference phase. During the training phase, constraint loss is established by comparing the constraint energy of the predicted parameters with the actual measured residual energy. During the inference phase, an adjustment strategy is selected based on the direction of energy deviation: when the predicted energy is too high, the peak acceleration is preferentially reduced; when the predicted energy is too low, the peak time is preferentially extended.

[0236] Clearly, the residual collision intensity is mainly affected by factors such as vehicle structural deformation, material properties, and collision conditions, all of which are strongly correlated with the collected acceleration sequences. Therefore, the second alternative prediction method is a staged prediction approach: first, the energy-normalized sequence data is input into a pre-trained collision pattern recognition model to obtain the collision type classification result; then, the classification result, the resampled sequence, and the original time-series signal are input together into a pre-trained intensity prediction model to achieve a more accurate residual intensity prediction. Compared with the end-to-end method, this method can explicitly obtain collision pattern information, providing better interpretability, but it involves two models connected in series during the inference process, which increases computational overhead.

[0237] Through the detailed implementation described above, this invention successfully solves the key technical problems faced by traditional methods. The new physical constraint mechanism takes into account the current collision state, is based on a rigorous derivation of the kinetic energy theorem, establishes constraint loss during the training phase by comparing it with actual measurement data, and ensures the physical rationality of the prediction results during the inference phase through energy balance estimation and soft correction.

[0238] It should be understood that although the steps in the flowchart are executed sequentially as indicated by the arrows, these steps can be optimized through parallel processing. The energy and time domain branches of the dual-path network can be computed in parallel, and feature fusion and multi-task prediction can be pipelined, thereby further improving the real-time performance of the algorithm.

[0239] The vehicle residual collision intensity determination method provided in this application addresses the TTF duration difference problem through energy normalization technology, fully utilizes multi-domain information through a dual-path network, ensures the rationality of the prediction results through physical constraints based on the current state, and achieves automotive-grade deployment requirements through optimization techniques. This method is not only applicable to vehicle collision intensity prediction but can also be extended to other engineering prediction problems that require the fusion of multi-source time-series data and physical constraints, possessing significant theoretical and practical value.

[0240] In another embodiment of this application, a vehicle collision intensity determination device is also provided, such as... Figure 23 As shown, it includes:

[0241] The acquisition module 111 is used to acquire acceleration sequence data collected during the vehicle collision process;

[0242] The first determining module 112 is used to determine the energy domain characteristics and time domain characteristics of the vehicle collision process based on the acceleration sequence data.

[0243] The fusion module 113 is used to fuse the energy domain features and the time domain features to obtain fused features;

[0244] The second determining module 114 is used to determine vehicle collision intensity information based on the fused features.

[0245] In another embodiment of this application, an electronic device is also provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0246] Memory, used to store computer programs;

[0247] When the processor executes the program stored in the memory, it implements the vehicle collision intensity determination method described in any of the foregoing method embodiments.

[0248] The processor is an automotive-grade system-on-a-chip (SoC) with over 16 GFLOPS of floating-point performance. It supports multi-core parallel processing and a dedicated AI acceleration unit, providing the computational power and AI acceleration capabilities required for real-time prediction. The memory includes system memory and non-volatile memory, storing network weight parameters, normalization parameters, calibration data, and system software. The communication interface supports vehicle network protocols such as CAN-FD, enabling real-time data exchange with other ECUs.

[0249] The electronic device provided in this embodiment of the invention, through its processor executing a program stored in its memory, addresses the issue of energy normalization technology effectively solving the problem of predicting residual collision intensity in the aforementioned vehicle collision intensity prediction method. The duration difference issue enables unified processing of signals from different operating conditions. Furthermore, the dual-path network architecture fully utilizes complementary information from the energy and time domains; the energy domain captures macroscopic release patterns, while the time domain preserves microscopic details. Physical consistency constraints ensure that the prediction results conform to fundamental physical laws, significantly improving the system's engineering reliability. Traditional linear models and simple neural networks struggle to handle the complex nonlinear relationships of multi-source heterogeneous data. The method provided in this invention achieves high-precision intensity prediction and offers a general model framework for fusing sequential and scalar data, which can be extended to similar engineering prediction problems.

[0250] The communication bus 1140 mentioned in the above-mentioned electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 24 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0251] The communication interface 1120 is used for communication between the above-mentioned electronic device and other devices.

[0252] The memory 1130 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0253] The processor 1110 mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0254] In another embodiment of this application, a computer-readable storage medium is provided, on which a program for a vehicle collision intensity determination method is stored. When the program for the vehicle collision intensity determination method is executed by a processor, it implements the steps of the vehicle collision intensity determination method described in any of the foregoing method embodiments.

[0255] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0256] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for determining the intensity of a vehicle collision, characterized in that, include: Acquire acceleration sequence data collected during a vehicle collision; Based on the acceleration sequence data, determine the energy domain characteristics and time domain characteristics of the vehicle collision process; Based on the acceleration sequence data, the energy domain characteristics and time domain characteristics of the vehicle collision process are determined, including: The acceleration sequence data is subjected to reference offset correction processing to obtain corrected acceleration sequence data; The cumulative collision energy is determined based on the corrected acceleration sequence data; Based on the accumulated collision energy, the corrected acceleration sequence data is resampled at equal energy intervals to obtain a resampled acceleration vector. Energy domain features in the resampled acceleration vector are extracted using an energy domain feature extraction model. The time-domain features in the corrected acceleration sequence data are extracted using a time-domain feature extraction model. The energy domain features and the time domain features are fused to obtain fused features; The energy domain features and the time domain features are fused to obtain fused features, including: The energy domain features and the time domain features are concatenated to obtain the concatenated features; Attention weights are determined based on the splicing features; Calculate the complementary weights of the attention weights; The fusion feature is determined based on the attention weight, the energy domain feature, the complementary weight, and the time domain feature; Vehicle collision intensity information is determined based on the fused features.

2. The method for determining vehicle collision intensity according to claim 1, characterized in that, Acquire acceleration sequence data collected during the vehicle collision, including: Acquire raw acceleration data collected at multiple time points; Based on the original acceleration data, the moment when the first preset number of consecutive composite accelerations all exceed the preset acceleration threshold is determined as the collision start moment; Obtain the ignition moment of the first ignition of the airbag controller; The acceleration data collected at multiple times between the collision initiation time and the ignition time from the original acceleration data are determined as the acceleration sequence data.

3. The method for determining vehicle collision intensity according to claim 1, characterized in that, Determining the cumulative collision energy based on the corrected acceleration sequence data includes: Obtain the equivalent collision mass of the vehicle; Based on the corrected acceleration sequence data, determine the triaxial velocity sequence data; The instantaneous power at multiple moments is calculated based on the equivalent collision mass, the triaxial velocity sequence data, and the corrected acceleration sequence data. The collision energy sequence is determined based on the instantaneous power at multiple moments; The cumulative collision energy is determined by summing multiple collision energy values ​​in the collision energy sequence.

4. The method for determining vehicle collision intensity according to claim 1, characterized in that, Based on the accumulated collision energy, the acceleration sequence data is resampled at equal energy intervals to obtain resampled sequence data, including: The accumulated collision energy is divided equally according to a second preset number to obtain a sequence of equally spaced energies; In the cumulative energy curve corresponding to the collision energy sequence, find multiple adjacent collision energies that are greater than the equal interval energy and less than or equal to the equal interval energy for each group of equally spaced energies. Obtain multiple adjacent moments of collision energy; The acceleration sequence data is resampled according to multiple sets of adjacent time points to obtain resampled sequence data.

5. The method for determining vehicle collision intensity according to claim 1, characterized in that, Based on the fused features, vehicle collision intensity information is determined, including: The fused features are input into the peak acceleration prediction head to obtain the remaining peak acceleration; The fused features are input into the peak arrival time prediction head to obtain the peak arrival time; The remaining peak acceleration and the peak arrival time are determined as the vehicle collision intensity information.

6. The method for determining vehicle collision intensity according to claim 1, characterized in that, The method further includes: During the training phase, the vehicle's current speed and current acceleration are obtained at the moment of ignition when the airbag controller is first ignited; The constraint energy is calculated based on the current speed, current acceleration, vehicle collision intensity information, and a preset constraint energy formula. Obtain the true remaining energy calculated from complete crash test data; The physical constraint loss is determined based on the constraint energy, the actual remaining energy, and the physical constraint coefficient. The loss function is determined based on the physical constraint loss.

7. The method for determining vehicle collision intensity according to claim 1, characterized in that, After determining the vehicle collision intensity information based on the fused features, the method further includes: Obtain the vehicle's current speed and current acceleration at the moment of ignition when the airbag controller first ignites; The predicted energy is calculated based on the current speed, current acceleration, vehicle collision intensity information, and a preset constraint energy formula. Obtain the true remaining energy calculated from complete crash test data; Determine the relative error between the predicted energy and the actual remaining energy; If the relative error is greater than a preset error threshold, the vehicle collision intensity information is corrected to obtain corrected collision intensity information.

8. The method for determining vehicle collision intensity according to claim 7, characterized in that, The vehicle collision intensity information is corrected to obtain corrected collision intensity information, including: The correction coefficient is determined based on the relative error and the preset correction coefficient calculation formula; The vehicle collision intensity information is corrected based on the correction coefficient to obtain corrected collision intensity information.

9. The method for determining vehicle collision intensity according to claim 8, characterized in that, The vehicle collision intensity information is corrected based on the correction coefficient to obtain corrected collision intensity information, including: If the predicted energy is greater than the actual remaining energy, the remaining peak acceleration in the vehicle collision intensity information is reduced based on the correction coefficient.

10. The method for determining vehicle collision intensity according to claim 8, characterized in that, The vehicle collision intensity information is corrected based on the correction coefficient to obtain corrected collision intensity information, including: If the predicted energy is less than or equal to the actual remaining energy, the peak arrival time in the vehicle collision intensity information is extended based on the correction coefficient.

11. A vehicle collision intensity determination device, characterized in that, include: The acquisition module is used to acquire acceleration sequence data collected during the vehicle collision process; The first determining module is used to determine the energy domain characteristics and time domain characteristics of the vehicle collision process based on the acceleration sequence data; Determining energy domain and time domain features of a vehicle collision process based on the acceleration sequence data includes: performing reference offset correction processing on the acceleration sequence data to obtain corrected acceleration sequence data; determining the cumulative collision energy based on the corrected acceleration sequence data; resampling the corrected acceleration sequence data at equal energy intervals based on the cumulative collision energy to obtain a resampled acceleration vector; extracting energy domain features from the resampled acceleration vector using an energy domain feature extraction model; and extracting time domain features from the corrected acceleration sequence data using a time domain feature extraction model. A fusion module is used to fuse the energy domain features and the time domain features to obtain fused features; fusing the energy domain features and the time domain features to obtain fused features includes: concatenating the energy domain features and the time domain features to obtain concatenated features; determining attention weights based on the concatenated features; calculating complementary weights of the attention weights; and determining the fused features based on the attention weights, the energy domain features, the complementary weights, and the time domain features. The second determining module is used to determine vehicle collision intensity information based on the fused features.

12. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the vehicle collision intensity determination method according to any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for a method of determining vehicle collision intensity, which, when executed by a processor, implements the steps of the method for determining vehicle collision intensity according to any one of claims 1-10.

Images