Methods, devices and storage media for calibrating automotive airbag control parameters

Abstract

This invention relates to a method, apparatus, and storage medium for calibrating control parameters for automotive airbags. The method includes: Step S1: Obtaining a list of control parameters for the airbag control unit algorithm and the value range of each control parameter, and selecting control parameters to be optimized based on the value range of each control parameter; Step S2: Loading a trained first surrogate model; Step S3: Generating multiple subpopulations and setting the migration direction of individuals between subpopulations, wherein the initial individual generation strategy and evolutionary control parameters of each subpopulation are different; Step S4: Initialization; Step S5: Generating a comprehensive performance score for all individuals based on the first surrogate model as the fitness value of the individual, performing iterative updates, and performing individual migration between subpopulations when necessary; Step S6: Calibrating based on the control parameter value sequence with the highest comprehensive performance score. Compared with the prior art, this invention can effectively obtain the optimal control parameter values ​​within a wider range.

Description

Technical Field

[0001] This invention relates to the field of automotive passive safety technology, and in particular to a method, device, and storage medium for calibrating automotive airbag control parameters. Background Technology

[0002] The parameter calibration of automotive airbag control strategies is one of the most critical and complex tasks in the development of passive safety systems. The airbag control unit (ACU) contains a large number of algorithm parameters that directly affect the airbag's collision recognition and triggering performance. Typical airbag control algorithm parameters include, but are not limited to: acceleration sensor trigger thresholds (e.g., collision acceleration thresholds set in the range of 10–50g), velocity change (ΔV) criterion thresholds, collision recognition time window lengths (e.g., 1–10 milliseconds), weighting coefficients for multi-sensor signal fusion, parameters of the collision type discrimination matrix, airbag deployment delay time, multi-level airbag inflation strategy parameters, occupant detection sensor thresholds, and various suppression logic parameters. With numerous and interconnected parameters, proper calibration to ensure timely airbag deployment and accurate triggering in various collision scenarios is a significant challenge in vehicle safety development.

[0003] During calibration, engineers typically rely on extensive real-vehicle crash test data. Vehicles undergo crash testing in professional crash tests or sled simulations at different speeds (e.g., 15 km / h, 25 km / h, 32 km / h, 50 km / h, 60 km / h), different impact angles (frontal impact, side impact, offset impact, etc.), and different types of collision barriers (rigid barriers, deformable barriers, moving barriers, etc.). High-speed sensors installed throughout the vehicle (such as B-pillar acceleration sensors, vehicle center acceleration sensors, and front compartment sensors) record the acceleration-time curves during the crash (sampling frequency can reach up to 20 kHz), and the velocity change-time curve is calculated by integrating the acceleration signals. These crash waveform data reveal the vehicle's dynamic response characteristics under various crash conditions.

[0004] Traditional airbag parameter calibration relies heavily on manual experience. Experienced safety systems engineers repeatedly adjust algorithm parameters based on extensive crash test data: meticulously analyzing the characteristics of each sensor's output signal for each crash scenario, and then manually adjusting relevant thresholds and logic. For example, in a high-speed, severe frontal collision, appropriate algorithm parameters need to be set so that the algorithm makes decisions within a precise time window to ensure timely airbag deployment; while in non-collision scenarios such as low-speed minor collisions, vehicle bumps, or violent door slamming, airbag triggering must be avoided, requiring sufficiently robust parameter settings. During parameter tuning, engineers must not only balance the impact of each parameter on various scenarios but also consider the interaction and coupling between parameters, as well as the impact of environmental temperature changes, sensor aging, and manufacturing tolerances on system performance. Therefore, manual calibration often requires multiple rounds of algorithm simulation and parameter correction based on the simulation results, lasting for weeks or even longer. Even with significant time and effort invested, manual trial-and-error methods still struggle to guarantee finding the globally optimal parameter combination. As modern automotive safety systems become increasingly complex, the number of parameters requiring calibration has increased significantly. Traditional manual calibration methods are proving increasingly inadequate in terms of development efficiency and quality optimization, making it imperative to introduce intelligent automatic calibration methods to improve efficiency and calibration results.

[0005] In response, although some existing technologies disclose the use of optimization algorithms to calibrate certain control systems of automobiles, for example, Chinese patent CN110458276A discloses a vehicle control parameter calibration method based on multi-objective particle swarm optimization algorithm. It optimizes the controller parameters to be calibrated through particle swarm optimization algorithm. Specifically, this existing technology dynamically changes the values ​​of control parameters through particle swarm optimization algorithm, and obtains the actual output of the actuator through experimental verification for each case. The deviation of the actual output from the target output is used as the fitness function value, and the control parameters of the active braking system are specifically cited as an example.

[0006] However, the aforementioned existing technologies cannot solve the problem of calibrating the control parameters of automotive airbag controllers. This is because the number of control parameters for airbag controllers is extremely large, approximately 1000.

[0007] 1. Particle swarm optimization has limited support for the number of parameters and does not support parameter optimization of this magnitude.

[0008] 2. If randomness is added to the optimization algorithm for verification, the cost of each experimental verification is extremely high, and it has almost no practical application value. Therefore, even if simulation is used for verification, a single simulation evaluation usually takes more than 10 minutes, which is impossible to achieve with a huge number of 1,000 parameters. Therefore, if those skilled in the art want to improve the optimization speed, they can only use a preset evolution direction, but this will lead to insufficient diversity and ultimately reduce the calibration effect of the control parameters. Summary of the Invention

[0009] The purpose of this invention is to provide a method, device, and storage medium for calibrating automotive airbag control parameters in order to overcome the deficiencies of the prior art.

[0010] The objective of this invention can be achieved through the following technical solutions:

[0011] A method for calibrating automotive airbag control parameters includes:

[0012] Step S1: Obtain the list of control parameters for the airbag control unit algorithm and the value range of each control parameter, and filter out the control parameters to be optimized based on the value range of each control parameter;

[0013] Step S2: Load the trained first surrogate model, wherein the first surrogate model takes the sequence of control parameter values ​​as input and the comprehensive performance score as output, and the comprehensive performance score is a dimensionless value of 0-1;

[0014] Step S3: Generate multiple subpopulations and set the migration direction of individuals between subpopulations. Each individual represents a sequence of control parameter values. The initial individual generation strategy of each subpopulation, as well as the evolutionary control parameters used to control crossover probability, mutation probability, mutation magnitude system, elite individual retention ratio and local search probability, are different.

[0015] Step S4: Initialize each subpopulation according to its own initial individual generation strategy, wherein each individual in any subpopulation corresponds to a set of control parameter value sequences;

[0016] Step S5: Generate the comprehensive performance score of all individuals based on the first agent model as the fitness value of the individual, and update it iteratively based on the evolutionary control parameters. If the diversity of any subpopulation decreases or the fitness value does not improve after multiple iterations, select one or more individuals from the corresponding other subpopulations to migrate in based on the migration direction.

[0017] Step S6: Calibrate based on the sequence of control parameter values ​​with the highest overall performance score.

[0018] Step S1 includes:

[0019] Step S1-1: Obtain the list of control parameters for the airbag control unit algorithm and the value range of each control parameter;

[0020] Step S1-2: Based on the value range of each control parameter, eliminate the parts whose value range is a fixed value;

[0021] Step S1-3: Further eliminate the relatively certain portions selected by the calibration engineer;

[0022] Step S1-4: Take all remaining control parameters as parameters to be optimized, and take the extreme values ​​of the range of each control parameter as its own boundary constraints.

[0023] The overall performance score is obtained by weighted summation of normalized airbag correct triggering rate, airbag false triggering rate, deployment timing error, and adaptability score to passenger conditions.

[0024] The first proxy model is trained based on simulation data and / or experimental data. The simulation data is obtained based on a first simulation platform. The first simulation platform completely ports the core ignition algorithm and control parameters in the airbag controller to the PC environment and completes the assignment of control parameters based on the control parameter value sequence. During the simulation, preset collision scenario data is read, the control parameter value sequence to be evaluated is loaded into the core ignition algorithm, and the complete collision recognition, decision-making and ignition control process is executed. After the simulation is completed, the first simulation platform automatically calculates the correct airbag trigger rate, the false airbag trigger rate, and the deployment timing error, and obtains a comprehensive performance score by combining the externally provided adaptability score for passenger conditions.

[0025] The initial individual generation strategy includes:

[0026] Uniform stratified sampling: Divide the value range of each control parameter into multiple intervals, uniformly sample multiple initial individuals in each interval, and additionally sample some individuals at the boundary of the control parameter value;

[0027] Multi-peak random sampling: Set multiple distribution peak values ​​in different regions for the control parameters, sample some initial individuals in the mixed distribution, sample some individuals based on the logarithmic scale, sample some individuals uniformly in the parameter space, and sample some individuals in the edge region;

[0028] Localized concentrated sampling: More than 50% of individuals are sampled near the global median;

[0029] Stratified Beta sampling: The normalized range of the control parameters is divided into multiple sub-intervals, and a certain number of individuals are sampled in each interval using a biased Beta distribution;

[0030] Hybrid Latin hypercube sampling: Individuals are generated by combining Latin hypercube sampling and boundary reinforcement sampling. The Latin hypercube sampling method ensures uniform distribution of projections across all control parameter dimensions.

[0031] The evolutionary control parameters are used to characterize the evolutionary strategy, and the types of evolutionary strategies include:

[0032] Standard strategy;

[0033] Exploration strategy: The initial crossover probability, elite retention ratio, and local search probability are lower than those of the standard strategy population; the initial mutation probability, variable asynchronous initial coefficient, and minimum step size are higher than those of the standard strategy population; and the step size decay factor is lower than 1.

[0034] The strategy is as follows: the initial crossover probability, elite retention ratio, and local search probability are greater than those of the standard strategy population; the initial mutation probability, variable asynchronous length initial coefficient, and minimum step size are lower than those of the standard strategy population; and the step size decay factor is greater than 1.

[0035] The balancing strategy is as follows: the initial crossover probability is greater than that of the standard strategy population and less than that of the exploitation strategy population; the initial mutation probability is less than that of the standard strategy population and greater than that of the exploitation strategy population; the local search probability is greater than that of the standard strategy population and less than that of the exploitation strategy population; the initial coefficient of the variable asynchronous length is less than that of the exploration strategy population and greater than that of the exploitation strategy population; and the minimum step size is greater than that of the exploitation strategy population and less than that of the exploration strategy population. During the iteration process, the crossover probability and mutation probability are periodically adjusted according to a fixed generation period, so that they alternate between a high crossover and low mutation state and a low crossover and high mutation state.

[0036] An adaptive strategy is employed: when diversity is detected to be below the first diversity threshold, the mutation probability is increased and the crossover probability is decreased; conversely, when diversity is detected to be above the threshold, the mutation probability is decreased and the crossover probability is increased.

[0037] The migration direction is circular.

[0038] The migration direction is:

[0039] Superior individuals from a population known for exploration are transferred to a population known for development, and vice versa. The population known for exploration is one with a crossover probability lower than that of the standard strategy and a mutation probability higher than that of the standard strategy. The population known for development is one with a crossover probability higher than that of the standard strategy and a mutation probability lower than that of the standard strategy.

[0040] In step S5, for each iteration of the first set number of generations, multiple individuals are extracted and input into the first simulation platform to obtain the comprehensive performance score output by the first simulation platform. If the comprehensive performance score output by the first simulation platform for multiple individuals differs too much from the comprehensive performance score output by the first surrogate model, the first surrogate model is supplemented with training based on the comprehensive performance score output by the first simulation platform.

[0041] A device for calibrating control parameters for automotive airbags includes a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the method described above.

[0042] A storage medium having a program stored thereon, which, when executed, implements the method described above.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. For an airbag controller with 1000 control parameters, a first surrogate model is designed to generate a comprehensive performance score for any sequence of control parameter values. This allows for rapid response of the comprehensive performance score without any constraints or limitations on the values ​​of the control parameters. Based on this, by designing a multi-strategy subpopulation, the search can be effectively prevented from getting stuck in local optima, allowing for a more thorough exploration of the high-dimensional parameter space and finding better combinations of control parameter values. Furthermore, the migration of high-quality individuals among the subpopulations can further improve the optimization efficiency in multiple blocks. Thus, through the evolutionary process of integrating multiple mutation and crossover operators, the algorithm has the ability to escape local basins and finely search for global optima. The co-evolution among the subpopulations further improves the global search efficiency and the diversity of solutions. Therefore, the final parameter calibration scheme can perform well in various collision scenarios, thereby improving the calibration effect of the control parameters.

[0045] 2. By fixing the values ​​and solidifying certain control parameters that are manually calibrated, and then using the remaining control parameters as control parameters to be optimized, the dimensionality can be reduced, thereby improving the calibration speed.

[0046] 3. The comprehensive performance score is obtained by weighted summation of the normalized airbag correct trigger rate, airbag false trigger rate, deployment timing error, and adaptability score to passenger conditions. The core ignition algorithm and control parameters in the airbag controller of the first simulation platform are the same as those in reality, which can improve the accuracy of the sample set and help provide sufficient training samples to train the first proxy model, thereby improving the accuracy of the first proxy model.

[0047] 4. By designing five initial individual generation strategies and evolutionary strategies, corresponding to five subpopulations respectively, it is possible to search and solve problems in a larger search space while reducing the reliance on human experience and repeated trial and error. This reduces the risk of deviation and error caused by human factors. Engineers only need to provide the parameter range and performance evaluation criteria, and the algorithm can automatically complete the parameter tuning and optimization work. This not only reduces the workload of engineers, but also makes the calibration results more reproducible and avoids the subjective uncertainty caused by human adjustment. It can meet the calibration of airbag controllers with 1000 control parameters.

[0048] 5. In addition to exploration and exploitation strategies, balancing and adaptive strategies are also implemented to achieve robust, efficient optimization with minimal human intervention, enhancing the algorithm's robustness and adaptability. The balancing strategy acts as a buffer between exploration and exploitation, reducing performance fluctuations and premature convergence risks during switching. It also maintains population diversity and avoids excessive perturbation through periodic adjustments. The adaptive strategy automatically adjusts the intensity of exploration and exploitation based on real-time status such as optimization progress and diversity, eliminating the need for manual parameter switching and reducing manual parameter tuning costs and mismatch risks. It also proactively optimizes parameters to prevent population collapse when diversity is low and flexibly adjusts the search mode when encountering evaluation noise or proxy bias. The individuals produced by both strategies possess both exploration breadth and exploitation depth, making them easier to adapt across islands in scenarios such as multi-island optimization, effectively improving migration gains and ensuring overall optimization results.

[0049] 6. Superior individuals from populations excelling in exploration migrate to populations primarily focused on development, and vice versa, achieving synergistic optimization between exploration and utilization. Exploratory islands extensively search the solution space using high mutation probabilities and large perturbation scales. After discovering potentially valuable regions, high-performing individuals from these exploratory islands are migrated to utilization islands. Utilization islands, with their high crossover probabilities and refined local search capabilities, further develop and refine these candidate solutions. Simultaneously, highly optimized, high-quality genes from utilization islands migrate back to exploratory islands, providing a better starting point for subsequent global searches and avoiding blind exploration. This bidirectional migration mechanism not only accelerates the discovery and convergence of high-quality solutions but also ensures the effectiveness of migration through a success rate gating mechanism (migration_success_rate > 0.4), preventing the ineffective propagation of low-quality genes. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the main steps of the method of the present invention;

[0051] Figure 2 This is a schematic diagram of multi-strategy evolution;

[0052] Among them: 101, subgroup one; 102, subgroup two; 103, subgroup three; 104, subgroup four; 105, subgroup five. Detailed Implementation

[0053] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0054] In this embodiment, the airbag control unit algorithm of a certain automobile is used as the object, and the method of this application is applied to automatically calibrate and optimize its key parameters. First, all parameters and value ranges related to the airbag control unit of this model are organized and stored in a configuration file (such as JSON format). Each parameter includes its hierarchical path, name, physical meaning, upper and lower limits of value range, and default initial value. For example, it includes parameters such as: "Main acceleration sensor trigger threshold (g)", range 10–50g; "Collision recognition time window (ms)", range 1–10ms; "Side impact sensor weight coefficient", range 0.0–1.0; "Occupant classification threshold", etc.

[0055] A method for calibrating automotive airbag control parameters based on multi-strategy evolutionary optimization, such as... Figure 1 As shown, it includes:

[0056] Step S1: Obtain the list of control parameters for the airbag control unit algorithm and the value range of each control parameter. Based on the value range of each control parameter, filter out the control parameters to be optimized, including:

[0057] Step S1-1: Obtain the list of control parameters for the airbag control unit algorithm and the value range of each control parameter;

[0058] Step S1-2: Based on the value range of each control parameter, eliminate the parts whose value range is a fixed value;

[0059] Step S1-3: Further eliminate the relatively certain portions selected by the calibration engineer;

[0060] Step S1-4: Take all remaining control parameters as parameters to be optimized, and take the extreme values ​​of the range of each control parameter as its own boundary constraints.

[0061] The eliminated control parameters do not participate in the subsequent optimization of the genetic algorithm, thereby reducing the dimensionality of the problem.

[0062] Step S2: Load the trained first surrogate model, wherein the first surrogate model takes the sequence of control parameter values ​​as input and the comprehensive performance score as output, and the comprehensive performance score is a dimensionless value of 0-1;

[0063] In actual calibration, the program to be tested and the control parameters need to be input into a first simulation platform on a PC. This platform integrates an ignition control program ported from the airbag ECU. By running the complete airbag control program, the performance of the control parameter value sequence under various collision scenarios is evaluated, including key indicators such as ignition timing accuracy and system robustness. However, a single simulation evaluation typically takes more than 10 minutes, which cannot meet the needs of genetic algorithms for large-scale evaluations. Therefore, this application uses machine learning methods to pre-train a first surrogate model. By learning the mapping relationship between parameter vectors and evaluation scores in historical simulation data, it achieves rapid prediction of the performance of any control parameter value sequence.

[0064] The overall performance score is obtained by weighted summation of normalized airbag correct triggering rate, airbag false triggering rate, deployment timing error, and adaptability score to passenger conditions;

[0065] The first proxy model is trained based on simulation data and / or experimental data. The simulation data is obtained from the first simulation platform. The first simulation platform completely ports the core ignition algorithm and control parameters in the airbag controller to the PC environment and completes the assignment of control parameters based on the control parameter value sequence. During the simulation, it reads the preset collision scenario data, loads the control parameter value sequence to be evaluated into the core ignition algorithm, and executes the complete collision recognition, decision-making and ignition control process. After the simulation is completed, the first simulation platform automatically calculates the correct airbag trigger rate, the false airbag trigger rate, and the deployment timing error, and obtains a comprehensive performance score by combining the externally provided adaptability score for passenger conditions.

[0066] In this embodiment, the first agent model is trained using the gradient boosting decision tree algorithm LightGBM.

[0067] Step S3: Generate multiple subpopulations and set the migration direction of individuals between subpopulations. Each individual represents a sequence of control parameter values. The initial individual generation strategy of each subpopulation, as well as the evolutionary control parameters used to control crossover probability, mutation probability, mutation magnitude system, elite individual retention ratio and local search probability, are different.

[0068] The population is divided into multiple subpopulations, and different candidate solution initialization strategies are adopted for different subpopulations to cover different regions of the parameter space and increase the diversity of the initial population.

[0069] Typical initial individual generation strategies include:

[0070] Uniform stratified sampling: The value range of each control parameter is divided into multiple intervals, and multiple initial individuals are uniformly sampled in each interval. Additionally, some individuals are sampled at the boundaries of the control parameter values. In this embodiment, specifically, the value range of each control parameter is divided into several intervals, and a certain number of initial individuals are uniformly sampled in each interval. For example, after normalizing the control parameter to the [0,1] interval, several points are uniformly sampled in sub-intervals such as [0.0,0.2], [0.3,0.4], [0.6,0.7], and [0.8,1.0], and additionally, some individuals are sampled near the upper and lower boundaries of the control parameter, thus ensuring coverage of both the central region and boundary extreme points.

[0071] Multi-peak random sampling: Multiple distribution peaks are set for the control parameters in different regions. Initial individuals are sampled from a mixed distribution, then some individuals are sampled based on a logarithmic scale, and some individuals are sampled uniformly within the parameter space, with some individuals also sampled in marginal regions. In this embodiment, specifically, multi-peak random sampling of control parameter values ​​is used to achieve broad exploration. For example, multiple distribution peaks are set for a certain control parameter in different regions (e.g., a mixture of three Gaussian distributions with means at 0.2, 0.5, and 0.8, and standard deviations of 0.15). Approximately 40% of the initial individuals are sampled from this mixed distribution. Alternatively, approximately 30% of the individuals are sampled using a logarithmic scale to favor the smaller value regions in the control parameter space; the remaining approximately 30% of the individuals are sampled uniformly within the control parameter space, but with emphasis on marginal regions (e.g., half in the high range [0.8, 1.0], and the other half in the low range [0.0, 0.2]. By combining multiple probability distributions for sampling, high diversity in the initial population is ensured globally.

[0072] Local concentrated sampling: More than 50% of the individuals are sampled near the global median. In this embodiment, specifically, dense sampling is performed in the empirically likely good solution region. For example, the control parameter values ​​of about 60% of the initial individuals are limited to the vicinity of the global median (normalized value of about 0.5, with a small normal perturbation, such as a standard deviation of 0.15), about 20% of the individuals take values ​​near the quartiles of the control parameters (such as around 0.25 and 0.75, each with a small perturbation of 0.1), and the remaining 20% ​​are sampled completely randomly, thereby strengthening the exploration of the central region of the control parameter space and the statistical empirical region.

[0073] Stratified Beta sampling: The normalized range of the control parameters is divided into multiple sub-intervals. Within each interval, a certain number of individuals are sampled using a biased Beta distribution. Specifically, in this embodiment, the normalized range of the control parameters is divided into multiple sub-intervals, and a certain number of points are sampled using a biased Beta distribution within each interval. Preferably, a Beta distribution with shape parameters α=2 and β=5 (biased towards smaller values) can be used in the lower intervals, and α=5 and β=2 (biased towards larger values) can be used in the higher intervals, thereby covering the values ​​from the lower limit to the upper limit of the control parameters and ensuring balanced coverage of the initial population.

[0074] Hybrid Latin Hypercube Sampling: Individuals are generated by combining Latin hypercube sampling and boundary reinforcement sampling. Latin hypercube sampling ensures a uniform distribution of projections across all control parameter dimensions. In this embodiment, specifically, Latin hypercube sampling (LHS) and boundary reinforcement sampling are combined. Some individuals are generated using the LHS method to ensure a uniform distribution of projections across all control parameter dimensions; some individuals are specifically sampled in regions close to the lower or upper limits of the control parameters to increase the proportion of boundary schemes; the remaining individuals can be generated using a skewed distribution (e.g., a left-skewed or right-skewed Beta distribution) to emphasize one end of the control parameter space. Through this hybrid approach, an initial solution set with a broad distribution and consideration of local priorities is obtained.

[0075] By employing one or more of the above initialization strategies in different subpopulations, this application obtained diverse candidate solutions in the initial generation, providing a good starting point for subsequent evolution that emphasizes both global exploration and local development. After initialization is complete.

[0076] Furthermore, this application sets different initial evolutionary control parameters and their variation strategies for each subpopulation to reflect different optimization tendencies. The evolutionary control parameters include, but are not limited to, crossover probability, mutation probability, mutation magnitude (step size) coefficient, elite individual retention ratio, local search probability, and adaptive adjustment rules for these parameters during the evolutionary iteration process. Based on this, the evolutionary control parameters are used to characterize the evolutionary strategy, and the types of evolutionary strategies include:

[0077] In this embodiment, the standard strategy employs balanced parameter settings: an initial crossover probability of approximately 0.8, an initial mutation probability of approximately 0.6, an elite retention rate of approximately 20%, a local search probability of approximately 0.1, an initial coefficient of variable step size of 1.0, a minimum step size of 0.05, and a step size that decreases linearly during the evolutionary process.

[0078] Exploration strategy: The initial crossover probability, elite retention rate, and local search probability are lower than those of the standard strategy population; the initial mutation probability, variable asynchronous length initial coefficient, and minimum step size are higher than those of the standard strategy population, and the step size decay factor is lower than 1, emphasizing global exploration capability. In this embodiment, the initial crossover probability is low (e.g., 0.6), the initial mutation probability is high (e.g., 0.9), the elite retention rate is low (e.g., 10–15%, preferably about 15%), and the local search probability is about 0.05. The variable asynchronous length initial coefficient is set to 1.5 (large perturbation range), the minimum step size is 0.1, and the step size decay factor is lower than 1 (e.g., 0.8) to maintain a large mutation range in the early stages of evolution.

[0079] The strategy employed is as follows: The initial crossover probability, elite retention rate, and local search probability are higher than those of the standard strategy population; the initial mutation probability, variable-asynchronous length initial coefficient, and minimum step size are lower than those of the standard strategy population, and the step size decay factor is greater than 1. This emphasizes local exploitation. In this embodiment, the initial crossover probability is relatively high (e.g., 0.95), the initial mutation probability is relatively low (e.g., 0.4), the elite retention rate is relatively high (e.g., 25%), and the local search probability is approximately 0.2. The variable-asynchronous length initial coefficient is set to 0.8 (for smaller initial perturbations), the minimum step size is 0.02, and the step size decay factor is greater than 1 (e.g., 1.2), allowing the mutation amplitude to converge and decrease more quickly.

[0080] The balancing strategy is as follows: the initial crossover probability is greater than that of the standard strategy population and less than that of the exploitation strategy population; the initial mutation probability is less than that of the standard strategy population and greater than that of the exploitation strategy population; the local search probability is greater than that of the standard strategy population and less than that of the exploitation strategy population; the variable asynchronous length initial coefficient is less than that of the exploration strategy population and greater than that of the exploitation strategy population; and the minimum step size is greater than that of the exploitation strategy population and less than that of the exploration strategy population. During iteration, the crossover and mutation probabilities are adjusted periodically using a fixed generational period, alternating between a high crossover and low mutation state and a low crossover and high mutation state. A moderately weighted parameter and a periodic adjustment mechanism are used. In this embodiment, the initial crossover probability is approximately 0.85, the mutation probability is approximately 0.5, and the local search probability is approximately 0.15. The variable asynchronous length initial coefficient is 1.0, and the minimum step size is 0.05. During evolution, the crossover and mutation probabilities are adjusted periodically using a fixed generational period (e.g., every 50 generations), alternating between a high crossover / low mutation state and a low crossover / high mutation state, maintaining a dynamic balance between exploration and convergence.

[0081] An adaptive strategy is employed. When diversity is detected to be below a first diversity threshold, the mutation probability is increased and the crossover probability is decreased; conversely, when diversity is detected to be above a threshold, the mutation probability is decreased and the crossover probability is increased. Population diversity is monitored in real time, and evolutionary control parameters are dynamically adjusted. In this embodiment, the initial crossover probability is approximately 0.8, the mutation probability is approximately 0.7, and the local search probability is approximately 0.1. The initial step size coefficient is 1.2, the minimum step size is 0.08, and the step size decay factor is approximately 0.9. The population adjusts the crossover / mutation probabilities in real time based on the diversity level: when diversity is detected to be low, the mutation probability is increased and the crossover probability is decreased; when diversity is too high, the opposite adjustment is made, ensuring the algorithm maintains a balance between exploration and convergence. Furthermore, a stagnation generation threshold is set. When fitness fails to improve for several consecutive generations and exceeds the threshold, a mechanism is triggered to temporarily increase the mutation magnitude and decrease the crossover probability in an attempt to escape the local optimum trap.

[0082] Different evolutionary strategies can be combined with different initial individual generation strategies and applied to different subpopulations. In this embodiment, for example... Figure 2 As shown, the five subpopulations are subpopulation 1 (101), subpopulation 2 (102), subpopulation 3 (103), subpopulation 4 (104), and subpopulation 5 (105). In this embodiment, among the five subpopulations:

[0083] The subpopulation 101 follows a standard strategy with a crossover probability Pc = 0.8, a mutation probability Pm = 0.6, an elite retention rate of 20%, and a local search probability of 0.1. The variable-asynchronous length has an initial coefficient of 1.0 and a minimum coefficient of 0.05 (linearly decaying to this value within the maximum number of generations). Each control parameter range is divided into several specified intervals, and a certain number of values ​​are uniformly sampled within each interval to ensure coverage of different segments. For example, several intervals such as [0.0, 0.2], [0.3, 0.4], [0.6, 0.7], and [0.8, 1.0] are selected, and the same number of samples are generated to cover low, medium, and high value ranges. Additionally, some individuals with control parameters at extreme boundary positions are generated, for example, by directly setting some samples to 0 or 1 (corresponding to the lower or upper limit of the control parameter), to ensure that the initial population includes extreme value schemes.

[0084] Subpopulation 2, 102, employs an exploration strategy with a crossover probability Pc = 0.6, a mutation probability Pm = 0.9, a low elite retention rate of approximately 15%, and a local search probability of 0.05. The initial step size coefficient is 1.5, with a minimum of 0.1, and a step size decay factor of 0.8 (to slow down the decay rate and maintain a larger mutation amplitude over a longer period). Since the exploration strategy focuses on broadly covering the parameter space, the values ​​of each parameter are generated through a mixture of multiple distributions: approximately 40% of the initial individuals are sampled using a multi-peak Gaussian distribution for each parameter dimension; for example, the normalized value of the parameter is set to have multiple Gaussian peaks around 0.2, 0.5, and 0.8 (with a standard deviation of 0.15), and random sampling is performed according to the Gaussian mixture model; approximately 30% of the individuals are sampled using a log-uniform distribution: by taking the logarithm of the interval [0.1, 1.0] and then uniformly sampling, values ​​biased towards smaller values ​​are generated; the remaining approximately 30% of the individuals are sampled using a uniform distribution but biased towards the boundary region: specifically, random numbers are first uniformly generated in the interval [0, 1], and then half of the values ​​are forcibly mapped to the higher end of the interval (e.g., values ​​greater than 0.5 are re-uniformly mapped to [0.8, 1.0]), and the other half is mapped to the lower end of the interval (values ​​less than or equal to 0.5 are mapped to [0.0, 0.2]).

[0085] Subpopulation 3 (103) employs a utilization strategy with a crossover probability Pc = 0.95, a mutation probability Pm = 0.4, a high elite retention rate of approximately 25%, and a local search probability of 0.2. The initial coefficient of the variable-asynchronous length is 0.8, with a minimum of 0.02, and a step size decay factor of 1.2 (to accelerate decay and rapidly reduce mutation amplitude, emphasizing convergence). Approximately 60% of the initial individuals have their control parameter values ​​concentrated near the global center (normalized value ~0.5), with a small normalized random perturbation (e.g., standard deviation 0.15). Approximately 20% of the individuals have their control parameter values ​​near the first and third quartiles: for example, small perturbations (standard deviation approximately 0.1) are added around 0.25 and 0.75 respectively. The remaining approximately 20% of the individuals have their control parameter values ​​uniformly and randomly sampled across the entire range [0,1] to ensure no possibility in the entire space is missed.

[0086] Subpopulation 4 (104) employs a balancing strategy with a crossover probability Pc = 0.85, a mutation probability Pm = 0.5, and a local search probability of 0.15. The initial coefficient of the variable-length coefficient is 1.0, with a minimum of 0.05. Every fixed number of generations (50 generations), the crossover and mutation probabilities are periodically adjusted: for example, the crossover probability Pc is increased and the mutation probability Pm is decreased for several generations, then the adjustment is reversed, alternating between exploration and convergence in a periodic manner. The [0,1] interval is divided into 8 equal sub-segments, each generating an equal number of samples. Within each sub-segment, a Beta distribution with different shape parameters is used for sampling: a left-skewed distribution (e.g., α = 2, β = 5) is used for the half of the sub-segment closest to 0, and a right-skewed distribution (e.g., α = 5, β = 2) is used for the half of the sub-segment closest to 0, so that smaller values ​​are generated more in the low-value sub-segments and larger values ​​are generated more in the high-value sub-segments. If the total number of individuals cannot be divided evenly by the number of sub-segments, the remaining small number of samples are then uniformly sampled and assigned to random sub-segments.

[0087] Subpopulation 5 (105) employs an adaptive strategy with a crossover probability Pc = 0.8, a mutation probability Pm = 0.7, and a local search probability of 0.1. The variable-asynchronous length initial coefficient is 1.2, with a minimum of 0.08, and a step size decay factor of 0.9. The population adjusts the crossover probability Pc and mutation probability Pm in real time based on its diversity: when the normalized variance (diversity index) of individuals in the population is below a threshold (indicating that the population is converging), the mutation probability Pm is temporarily increased and the crossover probability Pc is decreased; when the diversity is above the upper limit, the opposite adjustment is made to control the diffusion range of the solution. About 35% of the individuals are generated through Latin hypercube sampling (LHS) to ensure the uniformity of the projection distribution of each control parameter dimension; about 35% of the individuals are generated through hierarchical Beta distribution: half are sampled using left-skewed Beta (2,5) in the region near 0, and the other half are sampled using right-skewed Beta (5,2) in the region near 1; the remaining about 30% of the individuals are specifically reinforced for boundary values: for the values ​​randomly and uniformly sampled in [0,1], if >0.5, they are remapped to random values ​​in the interval [0.9,1.0], and if ≤0.5, they are mapped to the interval [0.0,0.1].

[0088] After generating initial control parameter matrices for each subpopulation according to the above strategy, this embodiment further processes the entire initial population to improve diversity and the coverage of solution distribution:

[0089] 1. Random perturbation: Apply small Gaussian noise with zero mean (standard deviation of about 0.02) to the control parameter values ​​of all generated individuals, where the noise amplitude is doubled for the components of control parameter values ​​in the middle region (e.g., between 0.2 and 0.8) to increase the spread of solutions in the middle region.

[0090] 2. Boundary reinforcement: Adjust some of the control parameter values ​​that meet the extreme conditions to their extreme values. Specifically, for control parameter values ​​that are already very close to the boundary in the initial individuals (e.g., <0.1 or >0.9), truncate them to 0 or 1 at a certain ratio to ensure that the lower and upper limits of each control parameter appear in the initial population.

[0091] 3. Disrupt correlations: For each column of the control parameter matrix (corresponding to each control parameter dimension), randomly select some dimensions to shuffle the value order of all individuals in that column, in order to cut off any possible parameter correlation patterns between individuals and ensure that the value sequence of each control parameter is more random and independent.

[0092] After the above processing, the initial candidate solution sets for each of the five subpopulations are obtained. Then, the performance of all initial individuals is evaluated using the previously trained LightGBM surrogate model, and fitness values ​​are calculated. At this point, each individual is associated with a predicted performance score, providing a starting point for the evolution of the genetic algorithm.

[0093] In some embodiments, the migration direction is circular. Furthermore, in this embodiment, the migration direction is as follows: superior individuals from a population excelling in exploration migrate to a population primarily focused on development, and superior individuals from a population primarily focused on development migrate to a population excelling in exploration. The population excelling in exploration is one with a crossover probability lower than that of the standard strategy and a mutation probability higher than that of the standard strategy; the population primarily focused on development is one with a crossover probability higher than that of the standard strategy and a mutation probability lower than that of the standard strategy. For the standard strategy, superior individuals from any other subpopulation can be used as the migrating individuals.

[0094] Step S4: Initialize each subpopulation according to its own initial individual generation strategy, wherein each individual in any subpopulation corresponds to a set of control parameter value sequences;

[0095] Step S5: Generate the comprehensive performance score of all individuals based on the first agent model as the fitness value of the individual, and update it iteratively based on the evolutionary control parameters. If the diversity of any subpopulation decreases or the fitness value does not improve after multiple iterations, select one or more individuals from the corresponding other subpopulations to migrate in based on the migration direction.

[0096] In each iteration of the first set number of generations, multiple individuals are extracted and input into the first simulation platform to obtain the comprehensive performance score output by the first simulation platform. If the comprehensive performance score output by the first simulation platform for multiple individuals differs too much from the comprehensive performance score output by the first surrogate model, the first surrogate model is supplemented with training based on the comprehensive performance score output by the first simulation platform.

[0097] The algorithm employs a multi-generational genetic algorithm for evolutionary iteration, continuously reproducing, selecting, mutating, and replacing individuals in each subpopulation. In this embodiment, the maximum number of generations G is set to 20,000. The algorithm iterates as many times as possible, within the limits of computational resources, to fully search the high-dimensional space. Each generation of evolution includes the following sub-steps:

[0098] Step S5-1: Selection and Elite Retention: Based on the fitness distribution of the subpopulation, an elite retention strategy is adopted to directly copy the highest-fitting individuals (e.g., 20% of the subpopulation size) from the current generation into the next generation, ensuring that excellent solutions are not destroyed by genetic operations. For the remaining individuals, a ranking-based selection method is used to generate a candidate set of parents: individuals are ranked from highest to lowest fitness, and the selection probability is calculated based on the ranking (higher rankings have higher probabilities, but not simply based on fitness proportions, to avoid complete monopolization by top individuals). The required number of parent individuals are randomly selected according to this probability distribution (repeated selection is allowed) to generate offspring.

[0099] Step S5-2: Crossover Reproduction: Perform crossover operations on the selected parent individuals according to the set crossover probability. If crossover is triggered, select a crossover method for each pair of parent individuals according to the predetermined probability to generate offspring; if crossover does not occur, the offspring directly inherit the genes of one of the parent individuals. The crossover methods include one or more of the following:

[0100] 1. Simulated Binary Crossover (SBX): A crossover method for real-number encoding. A crossover distribution index η is set, and offspring gene values ​​are calculated using a random number generation formula, ensuring that offspring are distributed near their parent gene values. Preferably, the distribution index η is dynamically adjusted according to the current evolutionary progress: a smaller η value is used in the early stages of evolution or when the global optimal fitness is not yet high, to increase the probability of offspring moving away from the parent solution (enhancing exploration); the η value is increased in the later stages of evolution or when a high-fitness solution has been found, making offspring closer to the parent values ​​(enhancing convergence). By adaptively adjusting η, a balance between exploration and convergence is achieved at different stages.

[0101] 2. Arithmetic Crossover: This method generates offspring by linearly combining the control parameters of the parents in a certain proportion. For example, a coefficient α between 0 and 1 is randomly generated. For each parent pair, two offspring are generated: for offspring 1, each control parameter = α × parent 1 control parameter + (1-α) × parent 2 control parameter; for offspring 2, the opposite is true. The offspring obtained by this method lie between the parent solutions, preserving the intermediate characteristics of the parents, which helps to generate robust offspring solutions when the parent solutions are similar.

[0102] 3. Uniform crossover: For each pair of parents, a binary mask string of equal length is randomly generated. The 0 / 1 of the mask determines which parent each control parameter of the offspring inherits from, thus producing offspring with a mixture of two genes. Uniform crossover gives each control parameter an independent inheritance opportunity, maximizing the recombination of parent genes and increasing the diversity of new individuals.

[0103] After the crossover operation is completed, the resulting offspring individuals are subjected to boundary checks. If the control parameter values ​​exceed the allowable range, they are truncated to the boundary values.

[0104] Step S5-3: Mutation Operation: Mutate the offspring individuals after crossover according to the set mutation probability. If an individual triggers a mutation, a random mutation method is selected to modify its genes. Mutation methods include one or more of the following:

[0105] 1. Gaussian perturbation mutation: This involves adding zero-mean random Gaussian noise to a portion of the individual's control parameters (e.g., approximately 30% of the control parameter dimensions). The standard deviation of the Gaussian noise is determined by multiplying the range of values ​​for the control parameter by the current global variation length coefficient. By perturbing only a subset of gene loci, most of the desirable characteristics of the individual solution can be preserved, while exploring local dimensions to find better values ​​in the solution neighborhood.

[0106] 2. Differential Evolution Mutation: Drawing inspiration from the differential evolution algorithm, multiple individuals (e.g., 5) are randomly selected from the current subpopulation. Their genes are weighted and combined according to their fitness to form a differential vector. This vector is then used to replace or perturb the genes of the mutated individuals. By incorporating information from multiple individuals in the population, new candidate solutions are generated, enabling mutation operations that move beyond the current individual's basin and helping to explore entirely new solution regions.

[0107] 3. Local perturbation mutation: Applying small perturbations to all control parameters of an individual (e.g., adding Gaussian noise with a standard deviation much smaller than the typical mutation amplitude to each control parameter), which is equivalent to performing a fine search in the vicinity of the current solution. This mutation method is often used in the later stages of optimization or for exploitative populations, to explore further improvement potential around known good solutions.

[0108] 4. Jump Mutation: This involves applying a large-scale random perturbation (e.g., a standard deviation more than twice the current general variation amplitude) to an individual with a small probability, causing significant changes in some or all of the individual's control parameters, thus allowing it to jump to a position far from the current solution for exploration. Although this mutation may temporarily reduce the individual's fitness, it helps to discover regions where the global optimal solution is located, and is particularly effective in breaking convergence when the subpopulation diversity is too low.

[0109] After the mutation operation is completed, boundary constraint correction is performed again on the mutated individuals to ensure that all control parameters remain within their respective allowable ranges. The above-mentioned mutation operators can be selected by preset probabilities or randomly. This invention preferably increases the probability of generating large perturbations (such as jump mutations) in the early stage to expand the search space, and increases the proportion of local perturbations in the later stage to refine the search for the optimal solution.

[0110] Step S5-4: Fitness Assessment: For newly generated individuals obtained through crossover and / or mutation, calculate their performance score using the surrogate assessment model provided in Step 2, which serves as the individual's fitness value. Since the surrogate model assessment speed is much faster than real physical simulation or collision experiments, new individuals from each subpopulation can be fed into the surrogate model for prediction in batches and in parallel, obtaining fitness on a large scale and in parallel, thereby supporting larger subpopulation sizes and more iterations.

[0111] Step S5-5: Subpopulation Update: The offspring individuals with newly calculated fitness in each generation are combined with the elite individuals of the previous generation to form the next generation subpopulation, replacing the non-elite individuals of the previous generation. Specifically, this update is performed in each subpopulation, recording the best fitness value of each subpopulation, as well as the best fitness and corresponding control parameter value sequence of the total population composed of all subpopulations.

[0112] Step S5-6: Surrogate Model Validation: Periodically validate the accuracy of the surrogate model at preset generational intervals. For example, after a fixed number of generations (e.g., 100 generations), randomly select a small number of individuals (e.g., a few) and simultaneously evaluate their performance using real collision simulations or experiments. Compare the obtained real fitness with the surrogate model's predicted values. Calculate error metrics (e.g., mean squared error (MSE), mean absolute error (MAE), maximum error, etc.) to monitor the surrogate model's prediction accuracy. If the surrogate model's error is found to be large, increase the frequency of real evaluations in subsequent iterations or stop the optimization process to retrain and update the surrogate model, ensuring the reliability of the evaluation stage during optimization.

[0113] In this embodiment, a sampling verification of the surrogate model's accuracy is performed every time a preset verification cycle (e.g., every 100 generations) is reached. Specifically, a small subset of individuals (e.g., 50) is randomly selected from the total population, and their fitness is predicted using the surrogate model. Simultaneously, a high-precision simulation model (or a performance function based on real-vehicle test data) is called to calculate the true fitness. The results are compared, and error statistics such as MSE, MAE, and maximum error are calculated. Typical error levels recorded in this embodiment are: mean squared error (MSE) on the order of 10^-3, and the maximum deviation between the predicted and true values ​​not exceeding 5%. If a significant increase in error is observed, countermeasures are considered, such as increasing the evaluation frequency of real simulations in subsequent iterations, or pausing optimization and retraining the surrogate model with new data to prevent the surrogate model from becoming inaccurate and causing the optimization to deviate from the correct direction. During the verification process, the number of times the surrogate model has been evaluated and the number of times the real model has been evaluated are also recorded to assess the efficiency improvement of this invention compared to pure real evaluation (typically several tens of times or more).

[0114] Step S5-7: Adaptive Adjustment and Migration: Dynamically adjust evolutionary control parameters such as crossover and mutation based on the evolutionary status of each subpopulation. When a significant decrease in individual diversity within a subpopulation is detected, or when fitness has not improved for several consecutive generations (stagnation), increase the mutation probability or mutation magnitude of that subpopulation, decrease the crossover probability, and reintroduce new random perturbations. Simultaneously, after several generations of population evolution, perform inter-population individual migration: dynamically set the migration interval based on global population diversity. In each migration generation, select several individuals with high fitness from each subpopulation as immigrants to replace the worst-fit individuals in the target population. The migration direction can be set according to differences in population strategies. For example, migrate excellent individuals from a population skilled in exploration to a population focused on development, or migrate excellent solutions from a population adept at local optimization to an exploration-oriented population, or use circular migration (island 1 to island 2, ..., island N to island 1) to facilitate information exchange between islands. Through adaptive migration, each subpopulation shares excellent genes while maintaining its own search characteristics, collaboratively approaching the global optimum.

[0115] In this embodiment, population diversity D The calculation formula is:

[0116]

[0117] in: n To control the number of parameters, For the population in the first i Standard deviation across each control parameter dimension For the first i The maximum value of each control parameter dimension. For the first i The minimum value of each control parameter dimension.

[0118] Based on the diversity calculated above, it is determined whether a subpopulation needs to be infiltrated by superior individuals from other subpopulations. In this embodiment, the basic migration cycle is set to 50 generations, with a diversity threshold of 0.1. When global diversity drops below the threshold (indicating that each subpopulation may converge near its local optimum), early migration is triggered, shortening the actual migration interval to approximately 30 generations; if diversity remains acceptable, migration proceeds at the default frequency. If migration is required in the current generation, the top k fittest individuals from each subpopulation are selected as immigrants (in this example, k=5, equivalent to selecting 5 optimal individuals from each population). The genes of these immigrant individuals are copied and sent to the target population. When selecting the migration direction, targeted exchanges are performed based on the strategic positioning of each subpopulation: for example, superior individuals from the exploration population are migrated to the utilization and balancing population to help the latter introduce new diversity and escape potential local optima; after receiving the immigrants, the target population replaces its current k least fittest individuals with these immigrant individuals. The effects of the migrating individuals are then monitored: the fitness changes of the target population after the update are compared. If the overall worst fitness of the target population improves after migration, it indicates that the migration has a beneficial effect, and the migration is counted as successful. If multiple migrations fail to improve the performance of the target population, the frequency of subsequent migrations of the corresponding source population is reduced. Through such dynamic adjustments, the effectiveness of the migration strategy is gradually optimized. In the later stages of this embodiment, the migration success rate between subpopulations tends to stabilize above 70%, indicating that most migration operations have a positive contribution to the performance of the target population.

[0119] Step S6: Calibrate based on the control parameter value sequence with the highest comprehensive performance score. Specifically, the iteration ends when a predetermined termination condition is met, and the currently found optimal control parameter value sequence and its corresponding performance index value are output. The termination condition can be reaching a preset maximum number of iterations, or the improvement in global optimal fitness in the most recent several generations being less than a threshold, etc. The final optimal parameter set can be used to update the calibration database of the airbag control unit and verify its performance.

[0120] The optimized control parameter value sequence obtained in this embodiment was used for verification in actual vehicle collision tests. The results show that, compared with the parameter scheme manually calibrated by engineers, the automatically optimized scheme exhibits superior decision-making performance in various collision scenarios—for medium-to-high-speed severe collisions, airbags can accurately identify and deploy at the optimal time; for low-speed minor collisions, shoulder impacts, and severe bumps, which are not safety-threatening situations, the airbag control unit did not falsely trigger. Statistics show that the optimized scheme, while ensuring a 100% trigger rate in serious collisions, reduces the false trigger rate in non-collision scenarios to nearly 0%, significantly improving occupant protection and system robustness. This verifies the effectiveness and superiority of the automatic calibration method of this invention.

[0121] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for calibrating control parameters of an automobile airbag, characterized in that, The method comprises the following steps: Step S1: obtaining a list of control parameters of an airbag control unit algorithm and a value range of each control parameter, and screening out control parameters to be optimized based on the value range of each control parameter; Step S2: loading a trained first agent model, wherein the first agent model takes a control parameter value sequence as input and outputs a comprehensive performance score, which is a dimensionless value between 0 and 1; Step S3: generating a plurality of sub-populations and setting the migration direction of individuals between the sub-populations, wherein the initial individual generation strategy of each sub-population and the evolution control parameters for controlling the crossover probability, mutation probability, mutation amplitude system, elite individual retention ratio and local search probability are different from each other; Step S4: respectively completing the initialization of each sub-population according to the respective initial individual generation strategy, wherein each individual of any sub-population corresponds to a group of control parameter value sequences; Step S5: generating the comprehensive performance score of all individuals based on the first agent model as the fitness value of the individual, and iteratively updating based on the evolution control parameters, wherein if the diversity of any sub-population decreases or the fitness value does not improve for consecutive iterations, one or more individuals are selected from the corresponding other sub-population based on the migration direction and migrated in; Step S6: calibrating based on the control parameter value sequence with the highest comprehensive performance score.

2. The method of calibrating control parameters of an airbag of an automobile according to claim 1, wherein, The step S1 comprises: Step S1-1: obtaining a list of control parameters of an airbag control unit algorithm and a value range of each control parameter; Step S1-2: based on the value range of each control parameter, removing the part with a fixed value range; Step S1-3: further removing the relatively certain part selected by a calibration engineer; Step S1-4: taking all the remaining control parameters as parameters to be optimized, and taking the extreme value of the value range of each control parameter as a boundary constraint condition of itself.

3. The method of calibrating control parameters of an airbag of an automobile according to claim 1, wherein, The comprehensive performance score is obtained by weighted summation of the normalized airbag correct triggering rate, airbag false triggering rate, deployment timing error and occupant condition adaptability score; The first agent model is trained based on simulation data and / or test data, wherein the simulation data is obtained based on a first simulation platform, the first simulation platform completely transplants the core ignition algorithm and control parameters in the airbag controller to a PC environment and completes the assignment of the control parameters based on the control parameter value sequence, in the simulation process, the pre-set collision scene data is read, the control parameter value sequence to be evaluated is loaded into the core ignition algorithm, the complete collision recognition, decision and ignition control process is executed, after the simulation is completed, the first simulation platform automatically calculates the airbag correct triggering rate, airbag false triggering rate and deployment timing error, and combines the externally provided occupant condition adaptability score to obtain the comprehensive performance score.

4. The method of calibrating control parameters of an airbag of an automobile according to claim 1, wherein, The initial individual generation strategy comprises: Uniform stratified sampling: dividing the value range of each control parameter into multiple intervals, uniformly sampling a plurality of initial individuals in each interval, and additionally sampling part of the individuals at the boundary of the control parameter value; Multi-peak random sampling: set multiple different regional distribution peaks for control parameters, sample part of initial individuals in mixed distribution, sample part of individuals based on logarithmic scale, and sample part of individuals uniformly in parameter space, and sample part of individuals in edge region; Local concentrated sampling: use more than 50% of individuals near the global median value; Hierarchical Beta sampling: divide the normalized range of control parameters into multiple subintervals, and sample a certain number of individuals in each interval using a biased Beta distribution; Mixed Latin hypercube sampling: combine Latin hypercube sampling and boundary reinforcement sampling to generate individuals, wherein the Latin hypercube sampling ensures that the projection on each control parameter dimension is uniformly distributed.

5. The method of calibrating control parameters of an airbag of an automobile according to claim 3, wherein, The evolution control parameter is used to characterize the evolution strategy, and the types of evolution strategy include: Standard strategy; Exploration strategy: initial crossover probability, elite retention ratio, local search probability is lower than that of standard strategy population, initial mutation probability, mutation step initial coefficient, minimum step is greater than that of standard strategy population, and step attenuation factor is lower than 1; Utilization strategy: initial crossover probability, elite retention ratio, local search probability is greater than that of standard strategy population, initial mutation probability, mutation step initial coefficient, minimum step is lower than that of standard strategy population, and step attenuation factor is greater than 1; Balanced strategy: initial crossover probability is greater than that of standard strategy population and less than that of utilization strategy population, initial mutation probability is less than that of standard strategy population and greater than that of utilization strategy population, local search probability is greater than that of standard strategy population and less than that of utilization strategy population, mutation step initial coefficient is less than that of exploration strategy population and greater than that of utilization strategy population, minimum step is greater than that of utilization strategy population and less than that of exploration strategy population, and the crossover probability and mutation probability are periodically adjusted according to a periodic function in the iteration process at a fixed generation cycle, so that they alternately change between high crossover and low mutation state and low crossover and high mutation state; Self-adaptive strategy: when the diversity is lower than the first diversity threshold, the mutation probability is increased and the crossover probability is decreased, and vice versa.

6. The method of calibrating control parameters of an airbag of a vehicle according to claim 5, wherein, The migration direction is ring migration.

7. The method of calibrating control parameters of an airbag of a vehicle according to claim 5, wherein The migration direction is: Excellent individuals in the exploration-oriented population are migrated into the development-oriented population, and excellent individuals in the development-oriented population are migrated into the exploration-oriented population, wherein the exploration-oriented population is a population with a lower crossover probability and a higher mutation probability than the standard strategy, and the development-oriented population is a population with a higher crossover probability and a lower mutation probability than the standard strategy.

8. The method of calibrating control parameters of an airbag of a vehicle according to claim 5, wherein In step S5, every first set of generations is iterated, a plurality of individuals are extracted and input into a first simulation platform, the comprehensive performance score output by the first simulation platform is obtained, and if the comprehensive performance score output by the first simulation platform corresponding to the plurality of individuals and the comprehensive performance score output by the first agent model differ too much, the first agent model is supplemented and trained based on the comprehensive performance score output by the first simulation platform.

9. An apparatus for calibrating control parameters of an automobile airbag, comprising a memory, a processor, and a program stored in the memory, wherein, The processor implements the method of any one of claims 1-8 when executing the program.

10. A storage medium having stored thereon a program, characterized in that The program is executed to implement the method of any one of claims 1-8. The program is executed to implement the method of any one of claims 1-8.

Images