Autonomous driving simulation test method combining transfer learning and vehicle configuration search

Abstract

The present invention discloses an automatic driving simulation test method combining transfer learning with vehicle configuration search, comprising: jointly simulating a simulator and an automatic driving system to be tested; selecting a vehicle dynamics model, and building a vehicle test operation scenario for a variable vehicle configuration; setting a metric for evaluating vehicle safety performance; filtering excessively small disturbances, and establishing a multi-objective optimization model; using a multi-layer perceptron architecture, and setting a vehicle safety performance metric prediction model according to the complexity of the metric; using a multi-objective search algorithm to search a first scenario, collect vehicle configuration combinations and labels, and pre-train the model; using a multi-objective search algorithm to search a second scenario, collect vehicle configuration combinations and labels, and using a transfer learning method to adjust the weights of the obtained pre-trained model to provide appropriate safety performance metric indicators in the second scenario; freezing the model that meets the model evaluation indicators during the training process, and generating an unsafe configuration.

Description

Technical Field

[0001] The present invention relates to a safety testing method for an autonomous driving vehicle, and in particular to a lightweight autonomous driving simulation testing method combining transfer learning with vehicle configuration search. Background Art

[0002] Autonomous driving systems represent a breakthrough innovation in the field of transportation, which has the potential to improve road safety by reducing human driving errors and alleviate traffic congestion to a certain extent. However, to realize the vision of autonomous driving, many challenges must be overcome, one of the most important of which is to ensure that the system can operate safely and reliably in a variety of complex situations. Therefore, how to improve the safety and stability of autonomous driving vehicles during driving has become a key research direction for researchers in many scientific research institutions today, and has very significant research value and significance.

[0003] Therefore, it is necessary to conduct comprehensive and effective testing of autonomous vehicles. However, testing autonomous driving systems in the real world requires a lot of time and resources, so most tests are now based on high-simulation simulators, which not only save time but also reproduce the errors of the real world that are difficult to replicate. As the scale and complexity of autonomous driving systems continue to grow, traditional manual testing methods can no longer meet the needs of testing. Therefore, the automation of autonomous driving testing has become an imperative trend, and many automated methods have been implemented, most of which use search-based methods that focus on generating unsafe test scenarios. However, what is often overlooked is the inherent variability in vehicle configurations, such as tire friction and damping rates, where even small perturbations can have a significant impact on system performance. Researchers have begun to recognize this problem and use multi-objective search algorithms to search for configurations that are prone to collisions. However, during the search process, these methods need to be verified using corresponding simulators, resulting in the consumption of a large amount of computing resources.

[0004] Therefore, it is an extremely important and challenging issue to ensure the safety of automobile products from the perspective of the vehicle's own configuration and reduce the consumption of computing resources by the multi-objective search algorithm. A lightweight autonomous driving simulation test method combining transfer learning and vehicle configuration search is proposed. The transfer learning method is used to train the model to evaluate the safety level of the vehicle configuration, and the multi-objective search algorithm is combined to help guide the generation of unsafe vehicle configurations. Summary of the invention

[0005] Purpose of the invention: The purpose of the invention is to use transfer learning combined with a multi-objective search algorithm to help automotive engineers solve the problem of consuming a lot of computing resources and time costs when searching for the impact of key vehicle configurations on the safety of autonomous driving systems in high-fidelity simulators. There is no need to test through high-fidelity simulators in the future, and a customized safety performance metric prediction model is used to analyze the key configurations that affect the safety of autonomous driving systems.

[0006] The present application is a method for autonomous driving simulation testing that combines transfer learning with vehicle configuration search, and the testing method comprises the following steps:

[0007] Step 1: Select a high-fidelity simulator and the autonomous driving system to be tested for joint simulation;

[0008] Step 2: Select a vehicle dynamics model, set the initial value and variation range for each vehicle configuration parameter for the selected variable vehicle configuration, and build at least two vehicle test operation scenarios;

[0009] Step 3: Set up vehicle safety performance evaluation metrics;

[0010] Step 4: Set filters to filter out small disturbances and establish a multi-objective optimization model;

[0011] Step 5: Use a multi-layer perceptron architecture to set up a vehicle safety performance metric prediction model based on the complexity of the metric;

[0012] Step 6: Use a multi-objective search algorithm to search the first vehicle test operation scenario built in step 2, collect vehicle configuration combinations and labels, pre-train the vehicle safety performance metric prediction model built in step 5, and set model evaluation indicators;

[0013] Step 7: Use a multi-objective search algorithm to search the second vehicle test operation scenario built in step 2, collect vehicle configuration combinations and labels, and use a transfer learning method to adjust the weights of the pre-trained model in step 6 to provide appropriate safety performance metrics in the second vehicle test operation scenario;

[0014] Step 8: Freeze the model that meets the model evaluation index during the training process of step 7 and generate an unsafe configuration.

[0015] Furthermore, in step 1, an appropriate high-fidelity simulator is selected as a virtual test environment, the autonomous driving system algorithm to be tested is integrated into the communication middleware, and the communication middleware is embedded in the simulator, or development is directly carried out based on the simulator to build an autonomous driving joint simulation platform.

[0016] Furthermore, the virtual vehicle in step 2 is composed of a vehicle dynamics model, which is provided by a simulator and has an initial value set; for the selected variable vehicle configuration, an initial value and a variation range domain are set for each vehicle configuration parameter, wherein the parameter variation range domain is determined based on expert experience or specific parameter values ​​of the vehicle production line.

[0017] Furthermore, the metrics used to evaluate vehicle safety performance in step 3 include safetyDegree, TET, TIT, and aveDece;

[0018] The safetyDegree indicator is measured by the stopping distance before the collision with the obstacle when the tested vehicle has not collided. In the case of a collision, the negative speed at the time of the collision is used to measure it. The expression of safetyDegree is:

[0019] ;

[0020] Where res represents the simulation result in the current scenario, d represents the Euclidean distance between the autonomous driving vehicle and the obstacle, and cs represents the instantaneous speed when the autonomous driving vehicle collides with the obstacle;

[0021] The TET and TIT indicators are used to assess the collision risk of vehicles; TET measures the duration of a vehicle's exposure to a potential collision within a specific time frame, while TIT considers the severity and duration of a potential collision; the collision time TTC is used to calculate the TET and TIT indicators, and the calculation formula for TTC is:

[0022] ;

[0023] Among them, the subscript i represents the i-th timestamp in the current time interval, L s is the relative distance, V L represents the relative speed;

[0024] The TET calculation formula is:

[0025] ;

[0026] ;

[0027] is a Boolean function. At time t, if TTC i (t) between 0 and the specified threshold TTC * Between The value is 1, otherwise it is 0. Indicates the duration of a frame;

[0028] The calculation formula of TIT is:

[0029] ;

[0030] aveDece is used to measure the impact of changes in deceleration during braking on the safety and comfort of an autonomous vehicle. The calculation formula is:

[0031] ;

[0032] ;

[0033] Among them, during the duration T, when the ADS issues a braking command, The value of is 1, otherwise, its value is 0. represents the deceleration at time t, Used to indicate whether the autonomous driving car has performed a braking operation at time t.

[0034] Furthermore, the definition of the filter in step 4 is as follows:

[0035] ;

[0036] in The original value configured for all selected vehicles, the search value is ,When the absolute difference between the original value and the search value exceeds the threshold, the change is considered different, otherwise the original value is kept for this search.

[0037] Furthermore, the multi-objective optimization model established in step 4 is:

[0038] ;

[0039] in, Represents the configuration value after the search, , , To evaluate the safety performance of vehicles, is the percentage of the maximum change difference in the changed parameter, is the number of parameters changed;

[0040] Five optimization goals are set: the first goal is to minimize the safety metric of the autonomous vehicle, the second goal is to maximize TET, the third goal is to maximize TIT, the fourth goal is to minimize the percentage of the maximum change difference of the selected parameters, and the fifth goal is to minimize the number of changed parameters. The multi-objective optimization model is converted into the optimization goal:

[0041] .

[0042] Furthermore, three models are built to predict and analyze the safetyDegree indicator: 1) collision signal prediction model; 2) collision regression model; 3) non-collision regression model.

[0043] Furthermore, the parameters of the multi-objective search algorithm configured in step 6 include setting the population size, evolutionary generations, crossover operator and its crossover rate.

[0044] Furthermore, in step 7, a multi-objective search algorithm is used to search the second vehicle test operation scenario built in step 2 to collect vehicle configuration information. If the collected vehicle configuration information meets the quantity requirements, the pre-trained model obtained in step 6 is loaded and all layers except the last two layers are frozen. The model weights are adjusted using a transfer learning method to provide appropriate vehicle safety performance metrics in the second vehicle test operation scenario.

[0045] Furthermore, in step 8, if the model after transfer learning can meet the evaluation index set in step 6, the model that meets the conditions is frozen and an unsafe configuration is generated.

[0046] Compared with the prior art, the present invention has the following significant effects: 1. The search target is improved in the multi-target search algorithm, which better helps automotive engineers locate key configuration parameters. 2. The transfer learning algorithm is combined with specific scenarios to build a customized safety performance metric prediction model, which reduces the number of simulation runs, greatly saves time and money, and helps automotive engineers optimize vehicle design. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0048] Figure 2 This is a safety performance metric prediction model architecture diagram of the present invention;

[0049] Figure 3 A search flow chart for a vehicle configuration of the present invention;

[0050] Figure 4 A schematic diagram of the combined process of transfer learning and multi-objective search algorithm of the present invention. DETAILED DESCRIPTION

[0051] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0052] Figure 1 The overall flow chart of the lightweight autonomous driving simulation test method combining transfer learning and vehicle configuration search of the present invention includes the following steps:

[0053] Step (1): Select a high-fidelity simulator and the autonomous driving system to be tested and perform joint simulation.

[0054] Step (2): Select a commonly used vehicle dynamics model, set the initial value and variation range for each vehicle configuration parameter according to the selected variable vehicle configuration, and build a vehicle test operation scenario.

[0055] Step (3): Set up vehicle safety performance evaluation metrics.

[0056] Step (4): Set up filters to filter out disturbances that are too small and establish a multi-objective optimization model.

[0057] Step (5): Use a multi-layer perceptron architecture to set up a vehicle safety performance metric prediction model based on the complexity of the metric.

[0058] Step (6): Use a multi-objective search algorithm to search for scene 1 built in step (2), collect vehicle configuration combinations and labels, pre-train the model built in step (5) and set model evaluation indicators.

[0059] Step (7): Use a multi-objective search algorithm to search for scenario 2 constructed in step (2), collect vehicle configuration combinations and labels, and use transfer learning methods to adjust the weights of the pre-trained model obtained in step (6) to provide appropriate safety performance metrics in scenario 2.

[0060] Step (8): Freeze the model that meets the model evaluation index during the training process of step (7) and generate an unsafe configuration.

[0061] Furthermore, in step (1), an appropriate high-fidelity simulator is selected as a virtual test environment, the autonomous driving system algorithm to be tested is integrated into the communication middleware and embedded into the simulator, or developed based on the simulator to build an autonomous driving joint simulation platform for testing.

[0062] Furthermore, the virtual vehicle in step (2) is composed of a vehicle dynamics model, which can be provided by a simulator and set with initial values. The vehicle configuration includes numerous physical parameters, and for the selected variable vehicle configuration, an initial value and a variation range are set for each vehicle configuration parameter, wherein the parameter range can be determined based on expert experience or specific parameter values ​​of a vehicle production line.

[0063] Furthermore, in step (2), a vehicle test operation scenario is constructed, wherein the elements of the scenario may include: static elements (such as road structure, weather, time, etc.) and dynamic participants (such as pedestrians and other vehicles, etc.).

[0064] Furthermore, the metrics used to evaluate vehicle safety performance in step (3) include safetyDegree, TET, TIT, and aveDece.

[0065] SafetyDegree is measured by the stopping distance before the collision with the obstacle when the tested vehicle has not collided. In the event of a collision, the negative value of the speed at the time of collision is used to measure it. The calculation formula of safetyDegree is as follows:

[0066] (1);

[0067] In formula (1), res represents the simulation result in the current scenario, d represents the Euclidean distance between the autonomous driving vehicle and the obstacle, and cs represents the instantaneous speed when the autonomous driving vehicle collides with the obstacle.

[0068] Time Exposed Time-to-collision (TET) and Time Integrated Time-to-collision (TIT) are used to assess the collision risk of a vehicle. TET measures the duration that a vehicle is exposed to a potential collision within a specific time frame, while TIT considers the severity and duration of a potential collision. By monitoring factors such as the distance, speed, and direction between the vehicle and the potential collision hazard, TIT provides a more comprehensive assessment of collision risk. In order to calculate TET and TIT, the concept of time-to-collision (TTC) measurement is adopted. The calculation formula for TTC is given in Equation (2):

[0069] (2);

[0070] Among them, the subscript i represents the i-th timestamp in the current time interval, L s is the relative distance, V L Represents the relative speed, and its specific calculation formula is given in formula (3):

[0071] (3);

[0072] TET refers to the duration of time during which the vehicle under test approaches a potential collision risk and its TTC value is lower than the threshold TTC* within a time period H, which means that when the TET value is lower, the situation is safer within H. The specific calculation formula is given in formula (4):

[0073] (4);

[0074] ;

[0075] is a Boolean function. At time t, if TTC i (t) between 0 and the specified threshold TTC * Between The value is 1 if yes, otherwise it is 0.

[0076] The TIT indicator uses the TTC score to calculate the safety level. TIT takes into account the safety level below a certain threshold (i.e., TTC * ), the threshold can be used to determine the collision risk level during time H. The specific calculation formula is given in formula (5):

[0077] (5);

[0078] Indicates the duration of a frame, which is 0.05s;

[0079] aveDece is used to measure the impact of the deceleration change of the autonomous vehicle during braking on its safety and comfort. Its specific calculation formula is given in formula (6):

[0080] (6);

[0081] ;

[0082] Among them, during the duration T, when the ADS issues a braking command, The value of is 1. Otherwise, its value is 0. represents the deceleration at time t, Used to indicate whether the autonomous driving car has performed a braking operation at time t.

[0083] Furthermore, in step (4), since some small disturbances may not reduce the safety of the autonomous driving vehicle, a filter is set to avoid invalid searches. The filter is defined as follows:

[0084] ;

[0085] The original value of the vehicle configuration is , the search value is , only when the absolute difference between the original value and the search value exceeds the threshold, the change is considered different, otherwise the original value is kept for this search.

[0086] Furthermore, the multi-objective optimization model established in step (4) is:

[0087] ;

[0088] in, Represents the configuration value after the search, , , To evaluate the safety performance of vehicles, is the percentage of the maximum change difference in the changed parameter, is the number of parameters changed.

[0089] According to the multi-objective optimization model, the following five optimization objectives can be obtained: the first objective is to minimize the safety metric of the autonomous driving vehicle, the second objective is to maximize TET, the third objective is to maximize TIT, the fourth objective is to minimize the percentage of change of the selected parameters, and the fifth objective is to minimize the number of changed parameters. The above multi-objective optimization model is transformed into the optimization objectives:

[0090] .

[0091] Furthermore, in step (5), since the attribute of the safety performance metric safetyDegree is discontinuous, three models need to be built to predict and analyze the metric. Therefore, a total of five different models need to be built to predict the safety performance metric, namely: 1) collision signal prediction model; 2) collision regression model; 3) non-collision regression model; 4) TIT regression model; 5) TET regression model. Multilayer perceptrons are used for construction. Due to the differences and complexity of each metric, two different model structures are used to build the above five models. The collision signal prediction model uses a common binary classification model structure to predict whether the configuration will cause a collision. If the predicted value is less than 0.5, the configuration is classified as an unsafe configuration, otherwise it is classified as a safe configuration. The collision / non-collision regression model and the TIT / TET regression model are both built using a general regression model because they face similar prediction challenges and metric calculations when predicting indicators. In order to cope with the inherent challenges of predicting collision and non-collision scenarios, the complexity of the model is enhanced by increasing the number of MLP layers.

[0092] Furthermore, the parameters of the multi-objective search algorithm that need to be configured in step (6) include setting the population size, evolutionary generations, crossover operator and its crossover rate, etc. In the scenario 1 built in step (2), the multi-objective search algorithm is used to solve the multi-objective model, and the process is as follows:

[0093] (1) Initialize the population according to the upper and lower bounds;

[0094] (2) Convert the multi-objective model into a fitness function;

[0095] (3) Iterate to obtain the Pareto optimal solution.

[0096] According to the search results, vehicle configuration combinations and labels are collected, and the above five models are pre-trained. Since the units of collision and non-collision situations are different in the definition of safetyDegree, safetyDegree is used to extend the collision attribute CollisionSign. Set safetyDegree less than 0 (collision situation) to 0, and set non-zero values ​​to 1 (non-collision situation):

[0097] ;

[0098] Finally, by repeatedly running the original dataset of scenario 1 and analyzing the inherent random nature of the data, fixed thresholds were set for the five models as evaluation indicators to guide the models to achieve specific levels of accuracy and error metrics in subsequent training processes.

[0099] Furthermore, in step (7), the multi-objective search algorithm with the same configuration is used to solve the multi-objective model in scenario 2 constructed in step (2). Vehicle configuration information is collected in scenario 2. If the collected vehicle configuration information meets the quantity requirement, the pre-trained model obtained in step (6) is loaded and all layers except the last two layers are frozen. The weights are adjusted using the transfer learning method to provide appropriate safety performance metrics in scenario 2.

[0100] Furthermore, in step (8), if the model after transfer learning can meet the fixed threshold set in step (6), that is, the model evaluation index, the model that meets the conditions is frozen and an unsafe configuration is generated.

[0101] The following will introduce the specific implementation process of each step:

[0102] Step 1: Select a high-fidelity simulator and the autonomous driving system to be tested for joint simulation.

[0103] Step 1.1: CARLA is chosen as the simulator for the experiment because it provides conditions very close to real-world driving conditions, making it easier for researchers to conduct experiments and test scenarios in a safe and controlled environment. World on Rail (WOR) is an end-to-end autonomous driving system built using CARLA. In terms of performance, WOR outperforms other end-to-end ADS, so WOR is chosen as the autonomous driving system to be tested. CARLA and WOR have strong compatibility in terms of communication interfaces and file formats, so they can be effectively integrated for joint simulation.

[0104] Step 1.2: After determining the joint simulation software, deploy CARLA on a Linux system and WOR on another Linux system. Since CARLA itself provides a Python API for communication with WOR, you only need to ensure that the two machines communicate in the same LAN to complete the joint simulation of the two software platforms.

[0105] Step 2: Select a commonly used vehicle dynamics model, set the initial value and variation range for each vehicle configuration parameter according to the selected variable vehicle configuration, and build a vehicle test operation scenario.

[0106] Step 2.1: Select a common vehicle dynamics model and set the initial value and range of each vehicle configuration parameter. According to the vehicle dynamics model provided by CARLA, select physical parameters and determine the range of parameters. A total of 12 physical parameters are selected, including max_rpm (maximum engine speed), dampRateFullT (damping ratio when the throttle is maximum), dampRateZeroT_CE (damping ratio when the throttle is zero when the clutch is engaged), dampRate_zeroT_CD (damping ratio when the throttle is zero when the clutch is disengaged), gearSwitchTime (switching time between gears), clutchStrength (clutch strength of the vehicle), mass (vehicle mass), dragCoeff (drag coefficient of the vehicle chassis), tireFric (scalar value of wheel friction), dampRate (damping rate of the wheel), radius (wheel radius), maxBrakeTorque (maximum braking torque).

[0107] Table 1 shows the selected vehicle configuration parameters, parameter units, parameter initial values ​​and parameter variation ranges.

[0108] .

[0109] Expose the corresponding interface for the selected variable vehicle configuration for subsequent calls.

[0110] Step 2.2: Build a crash test running scenario. First, you need to select the basic map environment of the simulator. In the selected map, you can customize the elements in the map. The elements that can be set include: static elements (such as road structure, weather, time, etc.), dynamic participants (pedestrians and NPC vehicles, etc., where the behavior and position of dynamic participants such as pedestrians and NPC vehicles can be customized). Finally, based on the NHTSA 37 pre-crash scenario as a reference, set up two accident-prone running test scenarios as scenario 1 and scenario 2 for testing.

[0111] Step 3: Set up vehicle safety performance evaluation metrics.

[0112] Step 3.1: safetyDegree is used to measure the distance the vehicle under test stops before colliding with an obstacle. In the event of a collision, safetyDegree is measured as the negative value of the speed at the time of collision. The calculation formula for safetyDegree is as follows:

[0113] ;

[0114] Step 3.2: Other safety performance metrics include TET and TIT, which are used to measure the hazard exposure time of the scenario, based on TTC and fed back by the simulator:

[0115] ;

[0116] ;

[0117] Step 3.3: aveDece measures the impact of changes in the vehicle configuration of the tested vehicle on its safety and comfort:

[0118]

[0119] ;

[0120] represents the deceleration at time t, Used to indicate whether the autonomous driving car has performed a braking operation at time t.

[0121] Step 4: Set filters to filter out small disturbances and establish a multi-objective optimization model.

[0122] Step 4.1: Limit the parameter changes that are too small by setting the change threshold, that is, determine the physical parameters that are ultimately involved in the change. The filter is defined as follows:

[0123] ;

[0124] Step 4.2: Feature change threshold is calculated by the domain D i =[ , ] is determined by the percentage change, that is:

[0125] ;

[0126] The change accuracy of the threshold It depends on the range of the value of the physical parameter, that is:

[0127] ;

[0128] Step 4.3: Establish a multi-objective optimization model:

[0129] .

[0130] Step 4.4: According to the multi-objective optimization model, the following five optimization objectives can be obtained:

[0131] .

[0132] Step 5: Use a multi-layer perceptron architecture to set up a vehicle safety performance metric prediction model based on the complexity of the metric.

[0133] Step 5.1: Since the attribute of the safety performance metric safetyDegree is discontinuous, three models need to be built to predict and analyze the metric. Therefore, a total of five different models need to be built to predict the safety performance metric: 1) collision signal prediction model; 2) collision regression model; 3) non-collision regression model; 4) TIT regression model; 5) TET regression model.

[0134] Step 5.2: The collision signal prediction model uses a common binary classification model structure to predict whether the configuration will lead to a collision.

[0135] Step 5.3: The collision / non-collision regression model and the TIT / TET regression model are both built using a common regression model.

[0136] Step 6: Use a multi-objective search algorithm to search for scene 1 built in step 2, collect vehicle configuration combinations and labels, pre-train the model and set model evaluation indicators.

[0137] Step 6.1: Select the multi-objective search algorithm NSGA-II and set the algorithm parameters including: population size, evolutionary generations, crossover operator and its crossover rate, etc.

[0138] Step 6.2: Use the PYMOO framework to implement the multi-objective search algorithm and initialize the algorithm parameters mentioned in step 6.1.

[0139] Step 6.3: Search for scenario 1. After 30*100 iterations, convergence is roughly achieved and the algorithm obtains the optimal solution, which reduces the safety performance of the autonomous driving vehicle and meets the parameter values ​​selected in the vehicle configuration for multi-objective optimization. Collect vehicle configuration combinations and labels, and pre-process the safetyDegree label.

[0140] Step 6.4: Use the TORCH framework to complete model pre-training. During the iterative training process, the inherent random nature of the data is analyzed, and fixed thresholds are set for the five models as model evaluation indicators to guide the models to achieve a specific level of accuracy and error measurement in the subsequent training process.

[0141] Step 7: Use a multi-objective search algorithm to search for Scenario 2 built in Step 2, collect vehicle configuration combinations and labels, and use transfer learning methods to adjust the weights of the pre-trained model obtained in Step 6 to provide appropriate safety performance metrics in Scenario 2.

[0142] Step 7.1: Use the multi-objective search algorithm with the same configuration as step 6 to search scene 2 and collect vehicle configuration combinations and labels that meet the conditions.

[0143] Step 7.2: Load the pre-trained model obtained in step 6 and freeze all layers except the last two layers. Use the TORCH framework to implement the transfer learning method to adjust the weights to provide appropriate safety performance metrics in scenario 2.

[0144] Step 8: Freeze the model that meets the model evaluation index during the training process of step 7 and generate an unsafe configuration.

[0145] Step 8.1: If the model after transfer learning can meet the fixed threshold set in step 6, that is, the model evaluation index, the model that meets the conditions is frozen and an unsafe configuration is generated.

[0146] The lightweight autonomous driving simulation test method combining transfer learning and vehicle configuration search of the present invention further studies the impact of vehicle configuration changes on ADS safety by using transfer learning and multi-objective optimization problems. Transfer learning is integrated into the search process of the NGSA-Ⅱ multi-objective search algorithm, solving the problem of consuming a large amount of computing resources when using a high-fidelity simulation simulator for virtual testing. It helps vehicle engineers locate key parameters and their range of variation in the configuration more quickly and accurately, and improves the robustness of the autonomous driving algorithm.

[0147] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. An autonomous driving simulation test method combining transfer learning and vehicle configuration search, characterized in that: The testing method comprises the following steps: Step 1: Select a high-fidelity simulator and the autonomous driving system to be tested for joint simulation; Step 2: Select a vehicle dynamics model, set the initial value and variation range for each vehicle configuration parameter for the selected variable vehicle configuration, and build at least two vehicle test operation scenarios; Step 3: Set up vehicle safety performance evaluation metrics; Step 4: Set filters to filter out small disturbances and establish a multi-objective optimization model; Step 5: Use a multi-layer perceptron architecture to set up a vehicle safety performance metric prediction model based on the complexity of the metric; Step 6: Use a multi-objective search algorithm to search the first vehicle test operation scenario built in step 2, collect vehicle configuration combinations and labels, pre-train the vehicle safety performance metric prediction model built in step 5 and set model evaluation indicators; the parameters of the multi-objective search algorithm configured in step 6 include setting the population size, evolutionary generations, crossover operator and its crossover rate; Step 7: Use a multi-objective search algorithm to search the second vehicle test operation scenario built in step 2, collect vehicle configuration combinations and labels, and use a transfer learning method to adjust the weights of the pre-trained model in step 6 to provide appropriate safety performance metrics in the second vehicle test operation scenario; Step 8: Freeze the model that meets the model evaluation index during the training process of step 7 and generate an unsafe configuration.

2. The automatic driving simulation test method combining transfer learning and vehicle configuration search according to claim 1, characterized in that: In step 1, an appropriate high-fidelity simulator is selected as a virtual test environment, the autonomous driving system algorithm to be tested is integrated into the communication middleware, and the communication middleware is embedded in the simulator, or development is directly carried out based on the simulator to build an autonomous driving joint simulation platform.

3. The automatic driving simulation test method combining transfer learning and vehicle configuration search according to claim 1, characterized in that: In step 2, the vehicle dynamics model is provided by the simulator and an initial value is set; for the selected variable vehicle configuration, an initial value and a variation range are set for each vehicle configuration parameter, wherein the parameter variation range is determined based on expert experience or specific parameter values ​​of the vehicle production line.

4. The automatic driving simulation test method combining transfer learning and vehicle configuration search according to claim 1, characterized in that: The metrics for evaluating vehicle safety performance in step 3 include safetyDegree, TET, TIT, and aveDece; The safetyDegree indicator is measured by the stopping distance before the collision with the obstacle when the tested vehicle has not collided. In the case of a collision, the negative speed at the time of the collision is used to measure it. The expression of safetyDegree is: ； Where res represents the simulation result in the current scenario, d represents the Euclidean distance between the autonomous driving vehicle and the obstacle, and cs represents the instantaneous speed when the autonomous driving vehicle collides with the obstacle; The TET and TIT indicators are used to assess the collision risk of vehicles; TET measures the duration of a vehicle's exposure to a potential collision within a specific time frame, while TIT considers the severity and duration of a potential collision; the collision time TTC is used to calculate the TET and TIT indicators, and the calculation formula for TTC is: ； Among them, the subscript i represents the i-th timestamp within time t, L s is the relative distance, V L represents the relative speed; The TET calculation formula is: ； ； is a Boolean function. At time t, if TTC i (t) between 0 and the specified threshold TTC * Between The value is 1, otherwise it is 0. Indicates the duration of a frame; The calculation formula of TIT is: ； aveDece is used to measure the impact of changes in deceleration during braking on the safety and comfort of an autonomous vehicle. The calculation formula is: ； ； Among them, during the duration T, when the ADS issues a braking command, The value of is 1, otherwise, its value is 0. represents the deceleration at time t, Used to indicate whether the autonomous driving car has performed a braking operation at time t.

5. The automatic driving simulation test method combining transfer learning and vehicle configuration search according to claim 1, characterized in that: The definition of the filter in step 4 is as follows: ； in The original value configured for all selected vehicles, the search value is ,When the absolute difference between the original value and the search value exceeds the threshold, the change is considered different, otherwise the original value is kept for this search.

6. The automatic driving simulation test method combining transfer learning and vehicle configuration search according to claim 1, characterized in that: The multi-objective optimization model established in step 4 is: ； in, Represents the configuration value after the search, , , To evaluate the safety performance of vehicles, is the percentage of the maximum change difference in the changed parameter, is the number of parameters changed; Five optimization goals are set: the first goal is to minimize the safety metric of the autonomous vehicle, the second goal is to maximize TET, the third goal is to maximize TIT, the fourth goal is to minimize the percentage of the maximum change difference of the selected parameters, and the fifth goal is to minimize the number of changed parameters. The multi-objective optimization model is converted into the optimization goal: 。 7. The automatic driving simulation test method combining transfer learning and vehicle configuration search according to claim 4, characterized in that: Three models are built to predict and analyze the safetyDegree indicator: 1) collision signal prediction model; 2) collision regression model; 3) non-collision regression model.

8. The automatic driving simulation test method combining transfer learning and vehicle configuration search according to claim 1, characterized in that: In step 7, a multi-objective search algorithm is used to search the second vehicle test operation scenario built in step 2 to collect vehicle configuration information. If the collected vehicle configuration information meets the quantity requirements, the pre-trained model obtained in step 6 is loaded and all layers except the last two layers are frozen. The model weights are adjusted using a transfer learning method to provide appropriate vehicle safety performance metrics in the second vehicle test operation scenario.

9. The automatic driving simulation test method combining transfer learning and vehicle configuration search according to claim 1, characterized in that: In step 8, if the model after transfer learning can meet the evaluation index set in step 6, the model that meets the conditions is frozen and an unsafe configuration is generated.

Images