Automatic driving automatic lane changing parameter generation method and system based on dil-hil linkage and driving behavior large model

Abstract

The application discloses an automatic driving automatic lane-changing parameter generation method and system based on DIL-HIL linkage and driving behavior large model, first, human driving data of multiple scenes and multiple driving styles is collected through a driver-in-the-loop platform, and a standardized data set is constructed after cleaning, normalization and feature extraction; then, a CNN-LSTM hybrid model is used to train a driving behavior large model to learn human lane-changing decision and operation rules; the model generates a parameter candidate set in combination with scene and style requirements, carries out hardware-level simulation test through a hardware-in-the-loop platform, completes quantitative evaluation relying on indexes such as lane-changing success rate, TTC and jerk, and finally outputs optimal lane-changing parameters by optimizing parameters through DIL-HIL closed-loop iteration. The application solves the problems of low efficiency, harsh parameters, and difficulty in balancing safety and comfort of traditional manual calibration, realizes parameter personification and scene self-adaptation, shortens the development cycle, and can be widely applied to lane-changing parameter calibration of various automatic driving vehicles.

Description

Technical Field

[0001] This invention belongs to the technical fields of intelligent driving control, driver behavior analysis, multi-in-the-loop simulation testing, and large-scale artificial intelligence models. Specifically, it relates to a method, system, and computer-readable storage medium for generating automatic lane change parameters for autonomous driving by combining driver-in-the-loop (DIL), hardware-in-the-loop (HIL) linkage simulation and a large-scale driving behavior model. Background Technology

[0002] Automatic lane changing is a core auxiliary function of advanced autonomous driving systems. The rationality of lane changing-related control parameters directly determines the driving safety, ride comfort, scenario adaptability, and road traffic efficiency of autonomous vehicles, and is a key calibration link in the development process of autonomous driving products.

[0003] Currently, the mainstream methods for generating automatic lane-changing parameters for autonomous driving fall into two categories: manual calibration and fixed-rule generation. Manual calibration relies on the professional experience of technicians, repeatedly debugging parameters such as lane-changing trigger timing, driving acceleration, and steering wheel angle on real vehicles or hardware-in-the-loop platforms. The calibration cycle for a single set of parameters is usually several weeks to several months, resulting in high labor and real-vehicle testing costs. Furthermore, the calibration results are easily affected by subjective human factors, making it difficult to comprehensively cover complex traffic scenarios and edge cases. Fixed-rule parameter generation methods pre-set rigid lane-changing logic and use uniform parameters to complete lane-changing control. This fails to differentiate between different drivers' driving styles, ultimately leading to stiff autonomous lane-changing actions that differ significantly from human driving behavior, thus reducing the user experience.

[0004] Existing technologies generally suffer from the following drawbacks:

[0005] 1) The parameters lack humanization and scenario adaptability. Traditional parameters are mostly uniform and cannot match diverse driving styles such as stable, aggressive, and conservative driving styles. When faced with complex road conditions such as traffic flow changes and vehicle dynamic interaction, problems such as unreasonable lane change timing and excessive actions are likely to occur, increasing driving safety hazards.

[0006] 2) Low development and iteration efficiency. The manual calibration process is cumbersome. When new traffic scenarios are added or driving requirements are adjusted, the entire calibration process needs to be repeated. There is a lack of automated iteration mechanisms, making it difficult to quickly respond to product upgrade needs.

[0007] 3) Safety, comfort, and traffic efficiency are difficult to balance. To mitigate risks, manually calibrated parameters are often conservative, leading to decreased traffic efficiency; conversely, prioritizing traffic efficiency comes at the expense of passenger comfort. Furthermore, traditional parameters lack support from large-scale real-world driving data, resulting in insufficient operational stability under extreme conditions and high-risk scenarios.

[0008] At its root, existing technologies have not deeply explored the massive amounts of human driving behavior data, making it impossible to effectively learn the driver's lane-changing decision-making logic and continuous operation patterns; a data-driven automated parameter generation process has not been established, and the overall work relies too heavily on manual labor; driver-in-the-loop simulation, algorithm models, and hardware-in-the-loop verification are disconnected from each other, lacking a closed-loop data feedback mechanism, and parameters that perform well in the simulation environment are prone to performance degradation after being deployed to real controllers, resulting in a disconnect between simulation and engineering implementation.

[0009] With the development of vehicle simulation technology and artificial intelligence large model technology, the industry urgently needs an automated parameter generation solution that integrates DIL-HIL linkage testing and driving behavior large model to solve the technical problems of low efficiency, insufficient anthropomorphism and inability to take into account multiple performance indicators in traditional calibration methods. Summary of the Invention

[0010] This invention aims to overcome the shortcomings of existing autonomous driving automatic lane change parameter calibration technologies, such as insufficient parameter anthropomorphism and personalization, weak scenario adaptability, long development cycles, high iteration costs, and difficulty in balancing safety and comfort. It provides a method, system, and computer-readable storage medium for generating autonomous driving automatic lane change parameters based on DIL-HIL linkage and a large-scale driving behavior model. This invention relies on a driver-in-the-loop platform to collect high-fidelity human driving data, combines it with a CNN-LSTM hybrid driving behavior model to learn human lane change patterns, and completes hardware-level simulation verification and closed-loop iteration through a hardware-in-the-loop platform to achieve fully automated generation of automatic lane change parameters.

[0011] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0012] A method for generating automatic lane-changing parameters for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model includes the following steps:

[0013] Step S1: Build a driver-in-the-loop (DIL) simulation platform, configure various traffic driving scenarios, recruit drivers with different driving styles to complete automatic lane changing simulation operations, and simultaneously collect multi-dimensional raw data of human driving behavior.

[0014] Step S2: Clean, time-align, and normalize the raw driving behavior data, and extract lane change intention features, traffic scene features, and vehicle operation features to obtain a standardized training dataset.

[0015] Step S3: Construct a large-scale driving behavior model. Use the standardized training dataset to train the large-scale driving behavior model so that it can learn the lane-changing decision-making logic, lane-changing timing selection, and vehicle operation rules of human drivers under different scenarios and driving styles.

[0016] Step S4: Based on the target application scenario and preset driving style, call the trained driving behavior big model to generate a candidate set of automatic lane changing parameters for autonomous driving.

[0017] Step S5: Load the candidate set of automatic lane change parameters into the automatic driving controller of the hardware-in-the-loop (HIL) test platform, conduct hardware-level real-time simulation testing, and collect simulation operation data.

[0018] Step S6: Quantitatively evaluate the test performance of the parameter candidate set based on preset evaluation indicators to determine whether the current parameter candidate set meets the preset performance requirements.

[0019] If the preset performance requirements are met, the current parameter candidate set is output as the optimal automatic lane change parameters and applied to the autonomous driving system. If the preset performance requirements are not met, the driving behavior model or the automatic lane change parameter candidate set is optimized and adjusted based on the evaluation results. If the deviation is small in the performance of local parameters, the parameter candidate set is directly optimized. If the system fails to meet the requirements due to missing scenario samples, the corresponding abnormal scenario DIL driving data is collected and fed back to step S3 to complete the model fine-tuning. Then, step S4 is returned to perform closed-loop iteration until the optimal automatic lane change parameters that meet the requirements are obtained.

[0020] Further optimization is achieved by including the following raw data on multi-dimensional human driving behavior in step S1: steering wheel angle data, accelerator pedal opening data, brake pedal opening data, vehicle speed, vehicle acceleration, vehicle jerk, relative distance between the vehicle and surrounding vehicles, relative speed, and relative position data; the various traffic driving scenarios include highway scenarios, urban road scenarios, congested road scenarios, and rural road scenarios; and the driving styles include stable driving style, aggressive driving style, and conservative driving style.

[0021] Further optimization, step S2 specifically includes the following sub-steps:

[0022] Step S201: Raw data anomaly cleaning: The 3σ criterion is used to identify and remove one-dimensional abnormal sampling points, and to remove entire segments of samples corresponding to signal packet loss and invalid operations, resulting in the initial screening dataset D. clean The formula for calculating the 3σ criterion is as follows:

[0023]

[0024] If satisfied If the sampled point is an outlier, it will be removed. These are sampled values ​​of the original one-dimensional data. The mean of the data. This represents the standard deviation of the data.

[0025] Step S202, Time Segmentation and Time Alignment: Mark the start and end times of the lane change for each sample, uniformly extract the time segment from 5 seconds before the lane change to 5 seconds after the lane change, and complete the time axis alignment by uniformly sampling the frequency, thus obtaining the aligned dataset D. align .

[0026] Step S203: Data normalization: The min-max normalization algorithm is used to unify the dimensions of the aligned data and map it to the interval [0,1] to obtain the normalized dataset D. norm The minimum-maximum normalization formula is:

[0027] ;

[0028] Where x represents the original data in a single dimension after alignment. , These are the minimum and maximum values ​​of the entire sample for this dimension, respectively. The data is after normalization;

[0029] Step S204, Multi-dimensional Feature Extraction: Extract lane change intention features, traffic scene features, and vehicle operation features from normalized data and fuse them. The formula for calculating traffic flow density features is as follows:

[0030] ;

[0031] in, Traffic density, L represents the number of vehicles within the perception range, and L represents the effective road length; a standardized training dataset is generated after fusing features.

[0032] Further optimization, step S3 specifically includes the following sub-steps:

[0033] S301. Constructing a large-scale CNN-LSTM hybrid driving behavior model: The large-scale CNN-LSTM hybrid driving behavior model sequentially includes convolutional layers, LSTM temporal layers, and a fully connected output layer; the convolutional layer operation formula is as follows:

[0034] ;

[0035] in, Input the feature matrix into the model. The convolution kernel weight matrix is... (*) represents the convolution bias term, (*) represents the convolution operation, and (f()) represents the ReLU activation function. The convolutional layer outputs a feature map;

[0036] The cell state update formula for the LSTM temporal layer is:

[0037] ;

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] in, , , These are the input gate, forget gate, and output gate, respectively. In cellular state, For output of the hidden layer, It is the Sigmoid activation function. This is for element-wise multiplication.

[0044] S302. Initialize the network weights, learning rate, and maximum number of iterations for the CNN-LSTM hybrid driving behavior model, and divide the standardized training dataset into training set, validation set, and test set in a 7:2:1 ratio. During the stratified sampling process, ensure that all three datasets contain sample data corresponding to multiple driving styles and multiple traffic scenarios.

[0045] S303. A large-scale CNN-LSTM hybrid driving behavior model is iteratively trained based on the training set. The Adam optimizer is used for backpropagation to update the model parameters, and the mean squared error is used as the loss function to calculate the deviation between the predicted parameters and the actual driving parameters. The formula for the mean squared error loss function is:

[0046] Where K is the number of samples in a single batch, To predict driving parameters for the model, These are parameters based on real human driving.

[0047] S304. Combine the validation set with the model convergence determination. When the validation set loss no longer decreases after a preset number of consecutive iterations, or when the number of iterations reaches the preset maximum number of iterations, terminate the training and save the model weight file with the lowest validation set loss to obtain the trained large driving behavior model. When the subsequent HIL hardware-in-the-loop test determines that the parameters do not meet the performance requirements, supplement the corresponding abnormal scenario DIL driving data back to this step to fine-tune and retrain the large driving behavior model.

[0048] Further optimization, step S4 specifically includes the following sub-steps:

[0049] S401, Input Condition Encoding: Convert the target application scenario and preset driving style into feature vectors that the model can recognize, and concatenate them to obtain the comprehensive input vector I.

[0050] S402, Forward Inference of the Model: The comprehensive input vector is input into the trained driving behavior model for forward inference to obtain the normalized original parameter output vector. The calculation formula is as follows:

[0051] ;in, To complete the training of the large-scale driving behavior model, This is the original output parameter vector of the model.

[0052] S403. Parameter Inverse Normalization: Perform inverse normalization on the model output parameters to restore them to true physical dimensional parameters. The inverse normalization formula is:

[0053] ;

[0054] in, To normalize the output values ​​of the model, , y represents the extreme value of the corresponding parameter, and y represents the actual physical parameter after restoration.

[0055] S404, Parameter Structured Encapsulation: Combine the restored lane change trigger conditions, acceleration, steering wheel angle, and lane change end judgment parameters to generate an automatic lane change parameter candidate set.

[0056] Further optimization, step S5 specifically includes the following sub-steps:

[0057] S501, Parameter Burning and Hardware Configuration: The candidate set of automatic lane change parameters is loaded into the autonomous driving domain controller of the HIL platform through the vehicle bus, and the vehicle dynamics model and traffic scene model are loaded at the same time to complete the hardware and simulation condition configuration.

[0058] S502. Real-time simulation operation: Start the HIL real-time simulation system, set a fixed simulation step size, reproduce the target traffic scenario, and trigger the automatic lane change function multiple times to complete the loop test.

[0059] S503. Simulation Data Acquisition: During the simulation process, vehicle motion state data, vehicle-road environment interaction data, and controller and actuator operation data are collected synchronously.

[0060] S504, Data Cache Storage: Store all simulation test data according to test samples to generate HIL simulation dataset for subsequent performance evaluation.

[0061] Further optimization, step S6 specifically includes the following sub-steps:

[0062] S601. Quantitative Calculation of Individual Evaluation Indicators: Based on the HIL simulation dataset, calculate four indicators: lane change success rate, collision time TTC, jerk acceleration, and lane change time. The formula for lane change success rate is as follows:

[0063] ;in, The number of successful lane changes. Total number of tests; Collision time calculation formula:

[0064] ;in, This refers to the relative distance between vehicles. The relative speed between vehicles; the formula for calculating jerk:

[0065] Where a is the vehicle acceleration and t is time.

[0066] S602. Multi-indicator weighted comprehensive score: The four individual indicators are weighted according to preset weights to obtain the comprehensive score. The comprehensive score formula is as follows:

[0067] Where w1, w2, w3, and w4 are the weights of each indicator, and the sum of the weights is 1. , These are the collision time and the mean of the acceleration, respectively. The design logic is as follows: the higher the lane change success rate and average collision time, the better the performance, so they are directly weighted; the higher the jerk and lane change time, the worse the comfort and traffic efficiency, so the reciprocal is used in the weighted calculation.

[0068] S603, Single index threshold comparison: Compare each of the four single indices with the preset safety threshold, comfort threshold, and efficiency threshold one by one;

[0069] S604. Overall performance judgment: When all individual indicators meet the corresponding threshold requirements and the comprehensive score is greater than or equal to the preset qualified score, the current parameter candidate set is judged to meet the preset performance requirements; otherwise, it is judged to not meet the preset performance requirements.

[0070] An automatic lane-changing parameter generation system for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model is used to implement the above-mentioned automatic lane-changing parameter generation method for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model. It includes a DIL data acquisition module, a data preprocessing module, a model training module, a parameter generation module, a HIL verification and testing module, a performance evaluation and iterative optimization module, and a parameter output application module that are connected in sequence.

[0071] The DIL data acquisition module is used to build a DIL simulation platform, simulate various traffic scenarios, and collect raw data on the driving behavior of different drivers.

[0072] The data preprocessing module is used to clean, align, and normalize the raw driving behavior data, and extract multidimensional features to generate a standardized training dataset.

[0073] The model training module is used to build and train a large model of driving behavior and learn the patterns of human lane changing and driving.

[0074] The parameter generation module is used to generate a candidate set of automatic lane change parameters by calling the trained large driving behavior model according to the target scenario and driving style.

[0075] The HIL verification and testing module is used to load the parameter candidate set into the autonomous driving controller, complete the hardware-in-the-loop real-time simulation test, and collect simulation data.

[0076] The performance evaluation and iterative optimization module is used to quantify the test results according to the evaluation indicators, determine whether the parameters meet the standards, and perform closed-loop iterative optimization of the parameters or models.

[0077] The parameter output application module is used to output the final optimal automatic lane change parameters and deploy the parameters to the autonomous driving system.

[0078] Further optimizations are made to the HIL verification test module, which includes an autonomous driving domain controller, a vehicle dynamics simulation unit, an onboard bus simulation unit, and a real-time monitoring unit; the autonomous driving domain controller is used to load and run the candidate set of automatic lane change parameters.

[0079] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the above-mentioned method for generating automatic lane-changing parameters for autonomous driving based on the DIL-HIL linkage and a large-scale driving behavior model.

[0080] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0081] 1. This invention collects various driving style data based on the DIL platform, and uses deep learning of human lane-changing decision-making and operation patterns through a large driving behavior model to generate differentiated lane-changing parameters. It abandons the traditional "one-size-fits-all" parameter mode, and the autonomous driving lane-changing behavior conforms to human driving habits, effectively improving ride comfort and user acceptance.

[0082] 2. The DIL platform can reproduce various driving conditions such as regular roads, dense traffic flow, and close-range overtaking; the CNN-LSTM hybrid model learns driving strategies in all scenarios simultaneously, and the generated parameters can be adapted to complex traffic environments to reduce safety risks.

[0083] 3. This invention constructs a fully automated workflow, shortening the traditional parameter calibration cycle of several months to several weeks, reducing labor costs and real vehicle testing costs; when facing new scenarios and new requirements, only a small amount of data needs to be added to complete the model fine-tuning, and the iteration response speed is fast.

[0084] 4. This invention uses collision time, acceleration, lane change success rate, and lane change time as quantitative indicators, and relies on HIL hardware-level simulation to complete multiple rounds of iterative optimization, balancing ride comfort and road traffic efficiency while ensuring driving safety.

[0085] 5. This invention constructs a DIL-HIL data closed loop, which feeds back the hardware-in-the-loop test results to the data acquisition and model training stages, solving the industry problem of good simulation results but hardware failure during implementation, and improving the engineering adaptability of parameters and models.

[0086] 6. The CNN-LSTM spatiotemporal hybrid architecture is adopted to extract traffic spatial features and model the timing operation rules of lane changing. Compared with a single network, it has higher accuracy in replicating human driving behavior. Attached Figure Description

[0087] Figure 1 This is an overall flowchart of the automatic lane change parameter generation method for autonomous driving based on DIL-HIL linkage and a large driving behavior model of the present invention.

[0088] Figure 2 This is a module architecture diagram of the automatic lane change parameter generation system for autonomous driving based on the DIL-HIL linkage and driving behavior big model of the present invention. Detailed Implementation

[0089] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0090] Example 1:

[0091] In this embodiment, the test conditions and evaluation criteria are as follows: Road and vehicle conditions: The test road is set as a one-way three-lane highway with a dry surface and good visibility. The vehicle is in automatic cruise control mode with a base speed of 100 km / h; there is a slow-moving vehicle in the same lane ahead at 60 km / h, requiring the vehicle to perform a left lane change and overtaking maneuver; the target driving style is set to a smooth driving style.

[0092] Performance qualification thresholds: Set the lower limit for each performance indicator: lane change success rate not less than 98%, mean collision time... Average jerk not less than 3.0s Not higher than 8.0 m / s3 The time required for a single lane change ranges from 2.0s to 4.0s.

[0093] like Figure 1 As shown, the method for generating automatic lane-changing parameters for autonomous driving based on the DIL-HIL linkage and a large-scale driving behavior model includes the following steps:

[0094] Step S1: DIL Driving Behavior Raw Data Collection: Build a DIL driver-in-the-loop simulation platform to fully reproduce the above-mentioned highway overtaking conditions. Recruit 10 licensed drivers, including 5 stable drivers and 5 aggressive drivers; each driver independently completes 20 standardized left lane change operations, and a total of 200 sets of raw driving samples are collected.

[0095] The data sampling frequency is set to 10Hz, and the synchronously collected data includes: steering wheel angle, accelerator pedal opening, brake pedal opening, vehicle speed, vehicle acceleration, vehicle jerk, relative distance, relative speed and relative position between the vehicle and surrounding vehicles.

[0096] Step S2, data preprocessing and feature extraction, specifically includes:

[0097] 1) Abnormal data cleaning: The 3σ criterion is used to identify and remove one-dimensional abnormal sampling points, and to remove entire segments of samples corresponding to signal packet loss and invalid operations, resulting in the initial screening dataset D. clean The formula for calculating the 3σ criterion is as follows:

[0098]

[0099] In the formula, These are sampled values ​​of the original one-dimensional data. The mean of the data. This represents the standard deviation of the data in the corresponding dimension. If it satisfies... The sampling point was identified as an outlier and removed; simultaneously, invalid samples with signal loss duration exceeding 200ms and those where the driver did not complete a full lane change were also removed, ultimately retaining 192 valid samples, resulting in the initial screening dataset D. clean .

[0100] 2) Timing Segmentation and Time Alignment: A continuous time segment from 5 seconds before the lane change starts to 5 seconds after the lane change ends is uniformly extracted from each sample group. All samples are resampled to 10Hz to complete time axis alignment, resulting in the aligned dataset D. align .

[0101] 3) Data normalization: The min-max normalization algorithm is used to unify the dimensions of the aligned data and map it to the interval [0,1] to obtain the normalized dataset D. norm The minimum-maximum normalization formula is:

[0102] ;

[0103] Where x represents the original data in a single dimension after alignment. , These are the minimum and maximum values ​​of the entire sample for this dimension, respectively. This is the normalized data. 4) Feature extraction and fusion: Extract lane change intention features, traffic scene features, and vehicle operation features, and complete feature fusion. The formula for calculating traffic flow density is:

[0104] ;

[0105] in, Traffic density, The number of vehicles within the sensing range is represented by L, and the effective road length is represented by L. A standardized training dataset is generated after feature fusion. In this embodiment, the effective road length L = 100m, the number of vehicles within the sensing range Ncar = 3, and the traffic density is calculated. =0.03 vehicles / m. After feature fusion is completed, a standardized training dataset is generated.

[0106] Step S3: Large Model Training of Driving Behavior

[0107] This embodiment uses a CNN-LSTM hybrid network as the large model of driving behavior. The network consists of convolutional layers, LSTM temporal layers, and fully connected output layers connected in series.

[0108] 1) The network calculation formula is as follows:

[0109] Convolutional layer operation formula: ;in, Input the feature matrix into the model. The convolution kernel weight matrix is... (*) represents the convolution bias term, (*) represents the convolution operation, and (f()) represents the ReLU activation function. The convolutional layer outputs a feature map;

[0110] The cell state update formula for the LSTM temporal layer is:

[0111] ;

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] in, , , These are the input gate, forget gate, and output gate, respectively. In cellular state, For output of the hidden layer, It is the Sigmoid activation function. This is for element-wise multiplication.

[0118] 2) Dataset partitioning: The 192 standardized training datasets were stratified and divided into 134 training sets, 38 validation sets, and 20 test sets according to a ratio of 7:2:1. Each set of samples contained data from different driving styles and working conditions.

[0119] 3) Iterative model training: Use the Adam optimizer and set the initial learning rate. =0.001, single batch sample size K=16, maximum number of iterations Epoch max =200. The mean squared error is used as the loss function, and the calculation formula is as follows: Where K is the number of samples in a single batch, To predict driving parameters for the model, These are parameters based on real human driving.

[0120] 4) Convergence Determination and Model Saving: An early stopping mechanism is implemented; training is terminated when the validation set loss fails to decrease for 10 consecutive iterations. This model converged after the 86th iteration, and the current optimal model weight file was saved.

[0121] Step S4, automatic lane change parameter candidate set generation, specifically includes:

[0122] 1) Convert the target application scenario and preset driving style into feature vectors that the model can recognize, and concatenate them to obtain the comprehensive input vector I;

[0123] 2) Input the comprehensive input vector into the trained model for forward inference: This yields the normalized parameter vector;

[0124] 3) Perform an inverse normalization operation to restore the actual physical parameters, as shown in the following formula:

[0125] ;

[0126] in, To normalize the output values ​​of the model, , y represents the extreme value of the corresponding parameter, and y represents the actual physical parameter after restoration.

[0127] 4) The parameters are encapsulated in a structured manner to obtain a parameter candidate set: the lane change triggering condition is that the distance between the self-vehicle and the vehicle in front is less than 80m, and the distance between the self-vehicle and the vehicle behind is greater than 60m; the longitudinal acceleration for lane change is 0.8m / s². 2 The peak steering wheel angle is 15°, and the rate of change of the angle is uniform; the condition for completing a lane change is that the vehicle is fully in the target lane and the lateral offset of the vehicle body is less than 0.3m.

[0128] Step S5: HIL Hardware-in-the-Loop Simulation Test: Load the parameter candidate set into the autonomous driving domain controller of the HIL platform via the vehicle CAN bus, and simultaneously load the vehicle dynamics model and the highway scene model, setting the simulation step size to 10ms. Conduct 100 consecutive cyclic lane change tests, simultaneously collecting data such as vehicle motion state, vehicle-to-infrastructure interaction, and controller output to generate the HIL simulation dataset.

[0129] Step S6, performance evaluation and compliance determination, specifically includes:

[0130] 1) Calculation of each indicator, specifically:

[0131] Formula for calculating lane change success rate: ;

[0132] in, The number of successful lane changes. This represents the total number of tests. This test... =100、 =100, the calculated lane change success rate is 100%.

[0133] Collision time calculation formula: ;

[0134] in, This refers to the relative distance between vehicles. The relative speeds between vehicles; the average collision time in this test. =4.5.

[0135] Formula for calculating jerk: ;

[0136] Where a is the vehicle acceleration, t is time, and the average jerk in this test is... =6.0m / s 3 .

[0137] According to statistics, the average lane change time was 2.8 seconds.

[0138] 2) Weighted Comprehensive Score: Set the indicator weights w1=0.4, w2=0.3, w3=0.2, and w4=0.1, with the sum of the weights being 1. The comprehensive score formula is as follows: In the formula, This represents the time taken for a single lane change. The calculated overall score is higher than the preset passing score.

[0139] 3) Threshold comparison: If all indicators meet the preset qualified threshold, the current parameter candidate set is determined to meet the performance requirements.

[0140] Step S7: Output the parameters. Output the parameters in this set as the optimal automatic lane change parameters and deploy them to the autonomous driving system.

[0141] This embodiment took 3 weeks to complete; 100% lane change success rate in 100 simulation tests, and all safety, comfort, and efficiency indicators were better than the preset thresholds. The generated lane change parameters closely match human driving habits for smooth and stable driving, with smooth and seamless movements, and the parameters operate stably in the hardware environment.

[0142] Example 2:

[0143] In this embodiment, the test road is an urban road with dense traffic, the average driving speed of the vehicle is 30km / h, and the target driving style is conservative.

[0144] Performance qualification thresholds: Lane change success rate not less than 98%, mean time to collision (MTBF) Average jerk not less than 2.5s Not higher than 7.0 m / s 3 The time required for a single lane change ranges from 1.5s to 3.5s.

[0145] The specific implementation steps are as follows:

[0146] 1. Follow steps S1 to S4 to complete DIL data acquisition, data preprocessing, model training, and parameter generation to obtain an initial parameter candidate set;

[0147] 2. Load the initial parameters into the HIL platform and complete 100 cycles of simulation testing;

[0148] 3. Performance Indicators: Lane change success rate 97%, average acceleration... =7.8m / s 3 Two indicators failed to reach the qualified threshold, and the judgment parameters did not meet the requirements;

[0149] 4. Closed-loop optimization: Analysis revealed that the original dataset lacked sufficient samples of congested close-range interaction scenarios. Fifty additional conservative driving samples from urban congested road sections were added to the DIL platform, and this new data was fed back into the model training stage for fine-tuning.

[0150] 5. Based on the fine-tuned model, regenerate the parameter candidate set and conduct 100 more HIL simulation tests.

[0151] 6. Secondary indicator statistics: Lane change success rate 99%, average collision time =3.2s, average jerk =6.5m / s 3 The average lane change time was 2.2 seconds, and all indicators met the standards.

[0152] 7. Output optimal automatic lane change parameters.

[0153] This closed-loop iteration took a total of 4 days. This embodiment verifies the DIL-HIL linkage closed-loop optimization mechanism of the present invention, which can supplement data and fine-tune the model for weak scenarios, quickly complete parameter optimization, and has high iteration efficiency.

[0154] Example 3:

[0155] In this embodiment, the same highway overtaking scenario as in Embodiment 1 is used, and two senior autonomous driving engineers calibrate the automatic lane change parameters using traditional manual methods.

[0156] The implementation process is as follows: Engineers repeatedly debugged parameters using the HIL platform, with an overall calibration cycle of 60 days. The final manually calibrated parameters were: lane change triggering conditions were a distance of less than 70m between the vehicle and the vehicle in front, and a distance of more than 50m between the vehicle and the vehicle behind; lane change acceleration was 0.6m / s². 2 The peak steering wheel angle is 12°.

[0157] 100 H-simulation tests were conducted on this set of parameters. Statistical results: lane change success rate 99%, average collision time... =3.8s average jerk =8.5m / s 3 The average lane change time is 3.6 seconds.

[0158] Table 1 shows a comparison of the experimental parameters and test results between this embodiment and Embodiment 1.

[0159] Table 1. Experimental parameters and test results

[0160]

[0161] The comparison results show that, compared with the traditional manual calibration method, the present invention significantly shortens the development cycle, while having obvious advantages in safety and comfort, and supports driving style customization and rapid iteration.

[0162] Example 4:

[0163] like Figure 2As shown, an automatic lane-changing parameter generation system for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model is used to implement the above-mentioned automatic lane-changing parameter generation method for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model. It includes a DIL data acquisition module, a data preprocessing module, a model training module, a parameter generation module, a HIL verification and testing module, a performance evaluation and iterative optimization module, and a parameter output application module that are connected in sequence.

[0164] The DIL data acquisition module is used to build a DIL simulation platform, simulate various traffic scenarios, and collect raw data on the driving behavior of different drivers.

[0165] The data preprocessing module is used to clean, align, and normalize the raw driving behavior data, and extract multidimensional features to generate a standardized training dataset.

[0166] The model training module is used to build and train a large model of driving behavior and learn the patterns of human lane changing and driving.

[0167] The parameter generation module is used to generate a candidate set of automatic lane change parameters by calling the trained large driving behavior model according to the target scenario and driving style.

[0168] The HIL verification and testing module is used to load the parameter candidate set into the autonomous driving controller, complete the hardware-in-the-loop real-time simulation test, and collect simulation data.

[0169] The performance evaluation and iterative optimization module is used to quantify the test results according to the evaluation indicators, determine whether the parameters meet the standards, and perform closed-loop iterative optimization of the parameters or models.

[0170] The parameter output application module is used to output the final optimal automatic lane change parameters and deploy the parameters to the autonomous driving system.

[0171] The HIL verification test module includes an autonomous driving domain controller, a vehicle dynamics simulation unit, an on-board bus simulation unit, and a real-time monitoring unit; the autonomous driving domain controller is used to load and run the candidate set of automatic lane change parameters.

[0172] Example 5:

[0173] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the above-mentioned method for generating automatic lane-changing parameters for autonomous driving based on the DIL-HIL linkage and a large-scale driving behavior model.

[0174] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating automatic lane-changing parameters for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model, characterized in that, Includes the following steps: Step S1: Build a driver-in-the-loop (DIL) simulation platform, configure various traffic driving scenarios, recruit drivers with different driving styles to complete automatic lane changing simulation operations, and simultaneously collect multi-dimensional raw data of human driving behavior. Step S2: Clean, time-align, and normalize the raw driving behavior data, and extract lane change intention features, traffic scene features, and vehicle operation features to obtain a standardized training dataset. Step S3: Construct a large-scale driving behavior model. Use the standardized training dataset to train the large-scale driving behavior model so that it can learn the lane-changing decision-making logic, lane-changing timing selection, and vehicle operation rules of human drivers under different scenarios and driving styles. Step S4: Based on the target application scenario and preset driving style, call the trained driving behavior big model to generate a candidate set of automatic lane changing parameters for autonomous driving. Step S5: Load the candidate set of automatic lane change parameters into the automatic driving controller of the hardware-in-the-loop (HIL) test platform, conduct hardware-level real-time simulation testing, and collect simulation operation data. Step S6: Quantitatively evaluate the test performance of the parameter candidate set based on preset evaluation indicators to determine whether the current parameter candidate set meets the preset performance requirements. If the preset performance requirements are met, the current parameter candidate set will be output as the optimal automatic lane change parameters and applied to the autonomous driving system; if the preset performance requirements are not met, the driving behavior model or the automatic lane change parameter candidate set will be optimized and adjusted based on the evaluation results. If the performance of local parameters is slightly deviated, the parameter candidate set will be directly optimized. If the system fails to meet the standard due to missing scenario samples, DIL driving data for the corresponding abnormal scenario is collected and fed back to step S3 to complete model fine-tuning, and then returned to step S4 to perform closed-loop iteration until the optimal automatic lane change parameters that meet the requirements are obtained.

2. The method for generating automatic lane-changing parameters for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model according to claim 1, characterized in that, The multi-dimensional raw data of human driving behavior mentioned in step S1 includes steering wheel angle data, accelerator pedal opening data, brake pedal opening data, vehicle speed, vehicle acceleration, vehicle jerk, relative distance between the vehicle and surrounding vehicles, relative speed, and relative position data; the various traffic driving scenarios include highway scenarios, urban road scenarios, congested road scenarios, and rural road scenarios; the driving styles include smooth driving style, aggressive driving style, and conservative driving style.

3. The method for generating automatic lane-changing parameters for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model according to claim 1, characterized in that, Step S2 specifically includes the following sub-steps: Step S201: Raw data anomaly cleaning: The 3σ criterion is used to identify and remove one-dimensional abnormal sampling points, and to remove entire segments of samples corresponding to signal packet loss and invalid operations, resulting in the initial screening dataset D. clean The formula for calculating the 3σ criterion is as follows: If satisfied If the sampled point is an outlier, it will be removed. These are sampled values ​​of the original one-dimensional data. The mean of the data. The standard deviation of the data; Step S202, Time Segmentation and Time Alignment: Mark the start and end times of the lane change for each sample, uniformly extract the time segment from 5 seconds before the lane change to 5 seconds after the lane change, and complete the time axis alignment by uniformly sampling the frequency, thus obtaining the aligned dataset D. align ; Step S203: Data normalization: The min-max normalization algorithm is used to unify the dimensions of the aligned data and map it to the interval [0,1] to obtain the normalized dataset D. norm The minimum-maximum normalization formula is: ； Where x represents the original data in a single dimension after alignment. , These are the minimum and maximum values ​​of the entire sample for this dimension, respectively. The data is after normalization; Step S204, Multi-dimensional Feature Extraction: Extract lane change intention features, traffic scene features, and vehicle operation features from normalized data and fuse them. The formula for calculating traffic flow density features is as follows: ； in, Traffic density, L represents the number of vehicles within the perception range, and L represents the effective road length; a standardized training dataset is generated after fusing features.

4. The method for generating automatic lane-changing parameters for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: S301. Constructing a large-scale CNN-LSTM hybrid driving behavior model: The large-scale CNN-LSTM hybrid driving behavior model sequentially includes convolutional layers, LSTM temporal layers, and a fully connected output layer; the convolutional layer operation formula is as follows: ； in, Input the feature matrix into the model. The convolution kernel weight matrix is... (*) represents the convolution bias term, (*) represents the convolution operation, and (f()) represents the ReLU activation function. The convolutional layer outputs a feature map; The cell state update formula for the LSTM temporal layer is: ； ； ； ； ； ； in, , , These are the input gate, forget gate, and output gate, respectively. In cellular state, For output of the hidden layer, It is the Sigmoid activation function. For element-wise multiplication; S302. Initialize the network weights, learning rate, and maximum number of iterations of the CNN-LSTM hybrid driving behavior model, and divide the standardized training dataset into training set, validation set, and test set in a 7:2:1 ratio. During the stratified sampling process, ensure that all three datasets contain sample data corresponding to multiple driving styles and multiple traffic scenarios. S303. A large-scale CNN-LSTM hybrid driving behavior model is iteratively trained based on the training set. The Adam optimizer is used for backpropagation to update the model parameters, and the mean squared error is used as the loss function to calculate the deviation between the predicted parameters and the actual driving parameters. The formula for the mean squared error loss function is: Where K is the number of samples in a single batch, To predict driving parameters for the model, These are parameters for real human driving. S304. Combine the validation set to determine model convergence. When the validation set loss no longer decreases after a preset number of iterations, or when the number of iterations reaches the preset maximum number of iterations, terminate training and save the model weight file with the lowest validation set loss to obtain the trained large driving behavior model. When the subsequent HIL hardware-in-the-loop test determines that the parameters do not meet the performance requirements, collect and supplement the corresponding abnormal scenario DIL driving data and return it to this step to fine-tune and retrain the large driving behavior model.

5. The method for generating automatic lane-changing parameters for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model according to claim 1, characterized in that, Step S4 specifically includes the following sub-steps: S401, Input Condition Encoding: Convert the target application scenario and preset driving style into feature vectors that the model can recognize, and concatenate them to obtain the comprehensive input vector I; S402, Forward Inference of the Model: The comprehensive input vector is input into the trained driving behavior model for forward inference to obtain the normalized original parameter output vector. The calculation formula is as follows: ;in, To complete the training of the large-scale driving behavior model, This is the original output parameter vector of the model; S403. Parameter Inverse Normalization: Perform inverse normalization on the model output parameters to restore them to true physical dimensional parameters. The inverse normalization formula is: ;in, To normalize the output values ​​of the model, , y represents the extreme value of the corresponding parameter, and y represents the actual physical parameter after restoration. S404, Parameter Structured Encapsulation: Combine the restored lane change trigger conditions, acceleration, steering wheel angle, and lane change end judgment parameters to generate an automatic lane change parameter candidate set.

6. The method for generating automatic lane-changing parameters for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model according to claim 1, characterized in that, Step S5 specifically includes the following sub-steps: S501, Parameter Burning and Hardware Configuration: Load the candidate set of automatic lane change parameters into the autonomous driving domain controller of the HIL platform through the vehicle bus, and load the vehicle dynamics model and traffic scene model at the same time to complete the hardware and simulation condition configuration. S502 Real-time simulation operation: Start the HIL real-time simulation system, set a fixed simulation step size, reproduce the target traffic scenario, and trigger the automatic lane change function multiple times to complete the loop test. S503, Simulation Data Acquisition: During the simulation process, vehicle motion state data, vehicle-road environment interaction data, and controller and actuator operation data are acquired synchronously. S504, Data Cache Storage: Store all simulation test data according to test samples to generate HIL simulation dataset for subsequent performance evaluation.

7. The method for generating automatic lane-changing parameters for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model according to claim 1, characterized in that, Step S6 specifically includes the following sub-steps: S601. Quantitative Calculation of Individual Evaluation Indicators: Based on the HIL simulation dataset, calculate four indicators: lane change success rate, collision time TTC, jerk acceleration, and lane change time. The formula for lane change success rate is as follows: ;in, The number of successful lane changes. Total number of tests; Collision time calculation formula: ;in, This refers to the relative distance between vehicles. The relative speed between vehicles; the formula for calculating jerk: Where a is the vehicle acceleration and t is time; S602. Multi-indicator weighted comprehensive score: The four individual indicators are weighted according to preset weights to obtain the comprehensive score. The comprehensive score formula is as follows: Where w1, w2, w3, and w4 are the weights of each indicator, and the sum of the weights is 1. , These are the collision time and the mean of the acceleration, respectively. This refers to the time consumed by a single lane change. S603, Single index threshold comparison: Compare each of the four single indices with the preset safety threshold, comfort threshold, and efficiency threshold one by one; S604. Overall performance judgment: When all individual indicators meet the corresponding threshold requirements and the comprehensive score is greater than or equal to the preset qualified score, the current parameter candidate set is judged to meet the preset performance requirements; otherwise, it is judged to not meet the preset performance requirements.

8. An automatic lane-changing parameter generation system for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model, characterized in that, The method for generating automatic lane change parameters for autonomous driving based on DIL-HIL linkage and a large driving behavior model as described in any one of claims 1 to 7 includes a DIL data acquisition module, a data preprocessing module, a model training module, a parameter generation module, a HIL verification and testing module, a performance evaluation and iterative optimization module, and a parameter output application module that are connected in sequence. The DIL data acquisition module is used to build a DIL simulation platform, simulate various traffic scenarios, and collect raw data on the driving behavior of different drivers. The data preprocessing module is used to clean, align, and normalize the raw driving behavior data, and extract multidimensional features to generate a standardized training dataset. The model training module is used to build and train a large model of driving behavior and learn the patterns of human lane changing and driving. The parameter generation module is used to generate a candidate set of automatic lane change parameters by calling the trained large driving behavior model according to the target scenario and driving style. The HIL verification and testing module is used to load the parameter candidate set into the autonomous driving controller, complete the hardware-in-the-loop real-time simulation test, and collect simulation data. The performance evaluation and iterative optimization module is used to quantify the test results according to the evaluation indicators, determine whether the parameters meet the standards, and perform closed-loop iterative optimization of the parameters or models. The parameter output application module is used to output the final optimal automatic lane change parameters and deploy the parameters to the autonomous driving system.

9. The automatic lane-changing parameter generation system for autonomous driving based on DIL-HIL linkage and a large-scale driving behavior model according to claim 8, characterized in that, The HIL verification test module includes an autonomous driving domain controller, a vehicle dynamics simulation unit, an on-board bus simulation unit, and a real-time monitoring unit; the autonomous driving domain controller is used to load and run the candidate set of automatic lane change parameters.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the automatic lane change parameter generation method for autonomous driving based on the DIL-HIL linkage and driving behavior big model as described in any one of claims 1 to 7.

Images