An interpretable autonomous driving decision-making method based on causal knowledge

Abstract

The present invention belongs to the field of autonomous driving testing technology, and more specifically, is an explainable autonomous driving decision-making method based on causal knowledge. The method comprises the following steps: Step 1: Discovering causal relationships for autonomous decision-making knowledge; Step 2: Constructing an explainable autonomous driving decision-making model; Step 3: Evaluating the explainability of the autonomous driving decision-making system; and Step 4: Evaluating the performance of the autonomous driving decision-making system. The present invention includes four steps: discovering causal relationships for autonomous driving decision-making knowledge, constructing an explainable autonomous driving decision-making model, evaluating the explainability of the autonomous driving decision-making system, and evaluating the performance of the autonomous driving decision-making system. Based on the causal discovery model of decision observation data, the method finds key variables in the causal relationship that influence the vehicle's driving behavior from the data, defines explainability standards and decision-making performance evaluation indicators, evaluates the explainability and decision-making performance of the autonomous driving decision-making system, and continuously optimizes and iterates the autonomous driving decision-making system.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving testing technology, and specifically provides an explainable autonomous driving decision-making method based on causal knowledge. Background Art

[0002] The continuous advancement of deep learning theory, the decreasing costs of sensors and computing hardware resources, and the emergence of policies and regulations supporting the development of autonomous driving technology have accelerated the commercialization of autonomous driving technology. Rule-based autonomous driving decision-making methods, due to their over-reliance on manually written rules, are unable to adapt to complex driving environments. With the continuous development of massive amounts of driving data and large models, end-to-end autonomous driving technology is expected to become the mainstream technology for the implementation of intelligent vehicles. Intelligent vehicles must consider the influence of the surrounding environment to make safe and reliable decisions. Traditional data-driven end-to-end autonomous driving technology uses environmental information as input and outputs a sequence of driving actions. However, much environmental information is irrelevant to the vehicle's decision-making, and this large amount of irrelevant data affects model performance. Furthermore, because end-to-end large models essentially learn correlations between data rather than causal relationships, the interpretability of autonomous driving decision-making systems remains a significant barrier to the development of high-level autonomous driving systems.

[0003] To accelerate the development of autonomous driving technology, developing explainable decision-making methods for autonomous driving and comprehensively and effectively evaluating the system's interpretability and decision-making performance are key technologies currently attracting research attention. Causal knowledge-based explainable decision-making methods for autonomous driving aim to provide a clear and transparent decision-making process. By leveraging massive amounts of driving data, they learn the causal knowledge that influences autonomous vehicle behavior decisions. This causal knowledge is then embedded into a spatiotemporal graph neural network model, intuitively reflecting the impact of each node or feature on the decision, thereby improving the model's interpretability. By constructing a causal discovery model for autonomous driving decision-making knowledge, the causal relationship between the causes and consequences of the vehicle's behavior is revealed. By building an explainable autonomous driving decision-making model and understanding the model's decision-making process, safe and trustworthy autonomous driving systems can be achieved. Explainability evaluation criteria and system evaluation metrics are developed to assess the interpretability and performance of autonomous driving decision-making systems.

[0004] There are relatively few patents for explainable autonomous driving decision-making methods based on causal knowledge in the field of intelligent vehicle decision-making. Chinese patent CN202211445291.6 discloses an explainable autonomous driving decision-making system and method thereof. By extracting key feature vectors from the traffic environment, inputting them into a deep Q-network model, and outputting corresponding decision instructions. Chinese patent CN202210201601.3 discloses an autonomous driving decision-making method and system based on a knowledge graph. The knowledge graph of the knowledge graph achieves the explainability of autonomous driving decision information and increases the credibility of the autonomous driving decision-making system. Chinese patent CN202410487199.9 discloses a risk estimation-based reinforcement learning autonomous driving safety explainable decision-making method. It makes understandable optimal decisions while ensuring safety constraints. The above three patents can achieve explainable autonomous driving decisions. Data-driven decision-making systems can only learn correlations between variables, not causal relationships. Therefore, embedding decision causal knowledge into deep learning models to achieve explainable autonomous driving decisions is still very critical. Summary of the Invention

[0005] The present invention provides an explainable autonomous driving decision-making method based on causal knowledge, which can derive a safe and explainable autonomous driving mode, solving the problem that traditional data-driven autonomous driving decision-making systems cannot utilize prior knowledge and the resulting weak user acceptability.

[0006] The technical solution of the present invention is described as follows in conjunction with the accompanying drawings:

[0007] A method for explainable autonomous driving decision-making based on causal knowledge includes the following steps:

[0008] Step 1: Discover the causal relationship for automated decision-making knowledge;

[0009] Step 2: Build an explainable autonomous driving decision model;

[0010] Step 3: Evaluate the interpretability of the autonomous driving decision-making system;

[0011] Step 4: Evaluate the performance of the autonomous driving decision system.

[0012] Furthermore, the specific method of step one is as follows:

[0013] S11. Collect multi-source heterogeneous driving data;

[0014] S12. Build a model for causal discovery based on decision-making observation data;

[0015] S13. Integrate autonomous driving decision-making knowledge.

[0016] Furthermore, the specific method of S11 is as follows:

[0017] S111, collecting virtual simulation data;

[0018] By building a driving simulator and simulation platform, various driving scenarios are designed, including urban roads, highways, and rural roads, covering different weather conditions, traffic flows, and road conditions; dynamic elements are added, including vehicles, pedestrians, and bicycles; and real traffic environments are simulated, including roads, buildings, traffic signs, and signal lights. A complete virtual driving environment is constructed, and the characteristic information of all scene elements constitutes a multi-source heterogeneous driving data set in the virtual simulation environment. v , as shown in formula (1):

[0019] (1)

[0020] Where, h e 、 h n 、 h b 、 h p 、 h l 、 h s are the feature sets of the scene elements of the vehicle, adjacent vehicles, adjacent bicycles, pedestrians, traffic lights and signboards respectively; among them, h e ={ x e , y e , vx e, vy e, s e , t e , b e}, h n ={ s n , x n , y n , vx n, vy n}, h b ={ x b , y b , vx b, vy b}, h p ={ x p , y p , vx p, vy p}, hl ={ x l , y l , s l}, h s ={ x s , y s , s s}; x e 、y e 、vx e、vy e、s e 、t e and b e They are the lateral position, longitudinal position, lateral speed, longitudinal speed, steering wheel angle, accelerator pedal opening and brake pedal opening of the vehicle; s n 、x n 、y n 、vx n、vy n are the type, lateral position, longitudinal position, lateral speed, and longitudinal speed of the adjacent vehicles respectively; x b 、y b , vx b, vy b are the Lateral position, longitudinal position, lateral speed, longitudinal speed; x p 、y p 、vx p、vy p They are the pedestrian's lateral position, longitudinal position, lateral speed, and longitudinal speed respectively; x l 、y l 、s l They are the horizontal position, vertical position and status of the traffic light respectively; x s 、y s 、s s They are the horizontal position, vertical position and status of the traffic sign respectively;

[0021] S112, collecting real vehicle data;

[0022] The raw data obtained by the inertial navigation positioning device, laser radar, millimeter wave radar, and camera are processed to obtain target-level data, and a multi-source heterogeneous driving data set Ω is obtained in a real vehicle environment. r ;

[0023] S113, classifying the scene;

[0024] The simulation data Ω v and multi-source heterogeneous driving data set, namely real vehicle driving data Ω r Slice the scene according to the road topology and determine the starting and ending points of each scene slice. The road topology is divided into roads of different curvatures, intersections with and without traffic lights, T-intersections, and roundabouts. Ensure that the road, traffic signs, signal lights and other elements within the slice are complete and consistent. Further classify the scene segments according to the number of vehicles in the surrounding environment, and the same number of vehicles in the surrounding environment are classified into the same type of scene.

[0025] S114, integrating multi-source heterogeneous driving data;

[0026] The generated scene slices are stored as independent simulation scenes, and a slice index is established to save the type, location, and environment information of each slice. The slice library is expanded as needed to add new slice types and scenes to obtain a classification set Ω for different road sections and different numbers of surrounding vehicles. v,n and Ω r,n ;

[0027] The specific method of S12 is as follows:

[0028] S121, preprocessing the data;

[0029] The obtained Ω v,n and Ω r,n Perform data cleaning and data standardization; first, fill the missing values with the Lagrange interpolation method, and use data standardization to convert data of different scales into a unified scale. The calculation formula is (2):

[0030] (2)

[0031] Where, X max and X min are the maximum and minimum values of the data respectively;

[0032] S122, build a causal relationship model;

[0033] A kernel regression model is used as a nonlinear model, with surrounding vehicles, traffic lights, and traffic signs as variables. Based on the ego vehicle's driving behavior on different road sections, with different numbers of surrounding vehicles, and with or without traffic lights and traffic signs, the nonlinear causal relationship between the variables that influence ego vehicle driving behavior is analyzed.

[0034] S123, fitting and evaluating the model;

[0035] The Granger causality test is used to fit the model parameters and the causal skeleton structure is determined through significance testing. A causal graph is generated to represent the causal relationship between the variables that affect the driving behavior of the vehicle, and the degree of influence of different variables on the driving behavior of the vehicle is quantified. The final output is a directed acyclic graph representing the causal relationship.

[0036] The specific method of S13 is as follows:

[0037] S131. Define road segment driving knowledge;

[0038] Integrate the knowledge that influences autonomous vehicle behavior decisions on a road segment; consider the interactions between the ego vehicle and surrounding vehicles and traffic signs, use a causal diagram to represent the causal relationships between the variables that influence the ego vehicle's driving behavior, and categorize the causal diagram based on the presence or absence of traffic signs and the number of vehicles in the surrounding environment;

[0039] S132. Define intersection decision knowledge;

[0040] Define the decision-making knowledge that affects the ego vehicle's driving behavior in intersection scenarios; consider the impact of traffic lights, pedestrians, and environmental vehicles on the ego vehicle; and classify and save causal graphs that can represent the causal relationship between variables that affect the ego vehicle's driving behavior based on the presence or absence of traffic lights, the number of pedestrians, and the number of environmental vehicles.

[0041] Furthermore, the specific method of step 2 is as follows:

[0042] S21. Constructing driving scenarios for autonomous vehicles;

[0043] S22. Build an autonomous driving decision model;

[0044] S23. Evaluation model offline test results.

[0045] Furthermore, the specific method of S21 is as follows:

[0046] S211, extracting scene elements;

[0047] Extract scene elements from the autonomous vehicle driving scene. Dynamic scene elements include the autonomous vehicle, adjacent vehicles, pedestrians, bicycles, and traffic lights. Static scene elements include traffic signs.

[0048] S212, extracting causal relationships;

[0049] The similarity between the integrated autonomous driving decision causal graph and the spatiotemporal graph structure is calculated, and the causal graph with high similarity is selected. The autonomous driving decision causal knowledge is initialized to the spatiotemporal graph. The similarity calculation is shown in formula (11):

[0050] (11)

[0051] Where, G c and G d They are cause-effect diagram and driving scenario diagram respectively; mcs ( G c , G d ) is the number of nodes in the largest common subgraph of the two graphs; max (| G c |,| G d |) is the ratio of the product of the number of nodes in the two graphs;

[0052] S213, representing an autonomous driving driving scenario based on a spatiotemporal graph structure;

[0053] The driving scene of the autonomous driving vehicle is represented by a spatiotemporal graph. For a certain moment, the scene elements are represented by the nodes of the graph, and the relationship between the scene elements is represented by the directed edges between the nodes. The dynamic scene modeling is as follows: on the basis of the graph structure, the time dimension is added to form a sequential graph model. The graph structure is represented as shown in formula (12):

[0054] (12)

[0055] Where, V ={ v e , v n , v b , v p , v l , v s};node v e , v n , v b , v p , v l ,v s They are the own car, neighboring car, bicycle, pedestrian, traffic light and sign; E ={ ee n , ee b , ee p , ee l , ee s}, ee n , ee b , ee p , ee l , ee s are directed edges between the vehicle and neighboring vehicles, bicycles, pedestrians, traffic lights, and traffic signs; the connection relationship of the directed edges is determined by the distance between the nodes. v e and v i distance d ( v e , v i ) is calculated using formulas (13) and (14):

[0056] (13)

[0057] (14)

[0058] when d ( v e , v i )<100m, v e and v i There are directed edges between them; the weights of the directed edges are initialized by the extracted causal graph to ensure that the sum of the weights of the directed edges of all nodes is 1;

[0059] The specific method of S22 is as follows:

[0060] S221. Build a dataset for autonomous driving decision-making;

[0061] The collected autonomous vehicle driving data is divided into data sets; a fixed-size sliding window is used to create training samples; the historical 3-second autonomous vehicle driving data is used to predict the driving behavior action sequence for the next 2 seconds; 70%, 20%, and 10% of the sample size are divided into training set, test set, and validation set respectively; the sample size N tot The calculation formula is shown in (15):

[0062] (15)

[0063] Where, N is the total length of the time series; the window size W The number of time steps included for each sample; step size S The number of time steps for each window movement;

[0064] S222, embedding causal knowledge in autonomous driving decision making;

[0065] The causal relationship of the causal graph with high similarity is extracted and embedded into the spatiotemporal graph model, including the relationship between nodes and edges that affect the driving behavior of the vehicle.

[0066] S223, learning the spatiotemporal correlation between “ego vehicle and scene elements”;

[0067] The spatiotemporal interaction between the ego vehicle and scene elements is learned through a spatiotemporal graph neural network. First, a temporal attention mechanism is used to calculate the feature representation of the ego vehicle node. Then, a graph convolutional neural network is designed to mine the graph structure at each moment to obtain the feature representation of the ego vehicle node that aggregates surrounding nodes. A temporal attention mechanism is used to learn the importance of the ego vehicle features at each moment. Finally, a channel attention mechanism is used to fuse the spatiotemporal correlations between the ego vehicle and scene elements.

[0068] S224, mapping the driving behavior of the vehicle;

[0069] The learned encoding vector is mapped into a three-dimensional vector through dimensional transformation, deformation and linear transformation, representing the steering wheel angle, accelerator pedal opening and brake pedal opening of the vehicle respectively;

[0070] The specific method of S23 is as follows:

[0071] S231. Define evaluation indicators;

[0072] The mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are used as indicators to evaluate the difference between the model prediction value and the true value. The calculation formulas (16), (17), and (18) of MAE, RMSE, and MAPE are as follows:

[0073] (16)

[0074] (17)

[0075] (18)

[0076] Where, and They are the actual driving action sequence and the predicted driving behavior action sequence of the autonomous driving vehicle respectively; N Indicates the number of samples;

[0077] S232, performing offline testing on the model;

[0078] By analyzing MAE, RMSE, and MAPE, the degree of deviation between the predicted values and the actual values of the steering wheel angle, accelerator pedal opening, and brake pedal opening of the autonomous driving vehicle is measured; the dataset for autonomous driving decisions is divided into several parts, and the model is trained and verified on different data subsets; the dataset is evenly divided into k mutually exclusive subsets; k-1 subsets are selected each time for training, and the remaining subset is used for verification; the model is trained and verified k times, and each time the verification set is different; finally, the average of the k verification results is taken as the model evaluation indicator.

[0079] Furthermore, the specific method of step three is as follows:

[0080] S31. Define interpretability criteria;

[0081] S32. Provide prior explanation of the autonomous driving decision model;

[0082] S33, provide post-hoc explanation of the autonomous driving decision model;

[0083] S34. Evaluate the autonomous driving decision system.

[0084] Furthermore, the specific method of S31 is as follows:

[0085] S311, analytical model transparency;

[0086] Analyze the model's own decision-making mechanism and obtain explanations by studying the model's internal structure and parameters; use the model's prior explanation results as the model transparency analysis results; evaluate and score the model's input layer, graph attention layer, temporal convolution layer, spatiotemporal fusion layer, fully connected layer, normalization layer, activation layer, residual connection layer, and output layer. The specific formula and operation of graph convolution explain how node features are aggregated through neighboring nodes; similar input data has stable output results at each layer; output model transparency analysis results U ;

[0087] S312, analyze the sensitivity of the model;

[0088] By introducing the perturbation analysis feature importance as an indicator for evaluating the model's ex post interpretability; using the counterfactual explanation method to change the model's input features before prediction and analyze the model's prediction changes; using the counterfactual explanation results as the model sensitivity analysis results, and outputting the model sensitivity analysis results V ;

[0089] S313, conduct objective evaluation;

[0090] Objectively evaluate the autonomous driving system; score the scene elements that affect the self-driving behavior according to the preset situation. Each scene element has a unique corresponding score. The degree of influence on the self-driving behavior is classified into strong, general and weak according to the score. Finally, objective evaluation results of the scene elements that affect the self-driving behavior in different scenarios are obtained. W ;

[0091] S314. Define the comprehensive evaluation model of interpretability;

[0092] Define an interpretability comprehensive evaluation model, comprehensively evaluate the transparency, sensitivity and objective evaluation results, and output the comprehensive evaluation results; analyze the key features obtained U 、 V and W Sort the importance, sorting rules, and calculation formula (33) are:

[0093] (33)

[0094] Where, S The intersection set of key features obtained by S311, S312 and S313, the remaining features are based on the principle of two-by-two intersection, and the priority size is defined as > > > > U > W > V , obtain the feature importance ranking results;

[0095] The specific method of S32 is as follows:

[0096] S321, analyze the model structure;

[0097] The defined graph structure is analyzed. The edges of the graph structure are key factors in message transmission. The connecting edges between nodes are determined by the interactive relationships between scene elements with different semantics and features. To distinguish the importance of different edges in message transmission, the fixed weights of different edges are first initialized according to the causal relationship model. During the model training process, the weights are parameterized, and the edge weights are updated through backpropagation and gradient descent algorithms. The learned edge weights are then visualized.

[0098] S322, analyzing and visualizing node importance;

[0099] Select and extract features that have a significant impact on model predictions and visualize the results. Based on the adjacency matrix between the ego vehicle node and other scene element nodes, the ego vehicle node is used as the central node, and the features of surrounding nodes are aggregated. The importance of the nodes is visualized through a heat map.

[0100] S323, obtaining prior explanation results;

[0101] The nodes with the greatest impact on the vehicle's driving behavior in the same scenario are counted and sorted according to the calculation results of the node importance. The frequency distribution of characteristic variables with different degrees of influence on the vehicle's behavior decision is displayed through histograms and box plots; the ranking results of the influence of different node features on the model prediction results are determined. U , generate a model explanation report, describing in detail the distribution of each feature on the prediction results;

[0102] The specific method of S33 is as follows:

[0103] S331, input key variables;

[0104] Input key variables; specifically: trained autonomous driving decision model f , key variables defined X i ;

[0105] S332, constructing counterfactual scenarios;

[0106] By constructing counterfactual scenarios, using do operators to generate counterfactuals, and intervening in multiple features X 1, X 2, …, X k , generate new input features, and the other features remain unchanged. The calculation formula (34) is:

[0107] (34)

[0108] Where, v i is the value after intervention, and the generated counterfactual state Is a Similar inputs but with some features interfered with; then using the intervened input state to change the vehicle’s driving environment and calculate the counterfactual decision result , the calculation formula (35) is:

[0109] (35)

[0110] In order to find the best counterfactual state, an optimization algorithm is used to minimize the comprehensive loss function. The loss function calculation formula (36) is:

[0111] (36)

[0112] Where the first term is the decision difference and the second term is the similarity of the input. λ is a hyperparameter that balances these two terms; use the gradient descent algorithm to solve and find ; is the input feature; is the state after intervention;

[0113] Use weighted Euclidean distance evaluation X and The similarity of , is calculated by formula (37):

[0114] (37)

[0115] S333, key factors for evaluation;

[0116] Evaluate key variables in ego vehicle driving behavior; first, evaluate counterfactual decision outcomes The impact of, including safety, efficiency and compliance, defines an impact function The difference value is quantified, and the difference value quantification calculation formula (38) is:

[0117] (38)

[0118] Where, s 、 e and l are the safety, efficiency and compliance evaluation score functions respectively; w s 、 w e and w l are the weight coefficients of safety, efficiency and compliance respectively; find the counterfactual decision based on the difference value safety, efficiency and compliance of key variables;

[0119] S334, obtain post-explanation results;

[0120] Analyze key variables, use histograms and box plots to count key variables that affect safety, efficiency, and compliance, and record counterfactual instance scenarios and key variables; determine the ranking results of key variables that affect the safety, efficiency, and compliance of the vehicle under the same scenario V ;

[0121] The specific method of S34 is as follows:

[0122] S341. Define a questionnaire; obtain evaluation scores based on the questionnaire for the objective evaluation of the autonomous driving decision-making system; the questionnaire is designed as follows: for lane-changing scenarios, determine the scene elements and characteristics that affect the ego vehicle's lane changing; for emergency braking scenarios, determine the scene elements and characteristics that affect the ego vehicle's braking; and for intersection scenarios with traffic lights, determine the scene elements and characteristics that affect the ego vehicle's driving behavior.

[0123] S342. Statistically analyze the evaluation results of the questionnaire in S41, obtain a ranking result W of scene elements and key feature variables based on the evaluation scores, and intuitively display the degree of influence of different scene elements on vehicle driving behavior through a graphical visualization method.

[0124] Furthermore, the specific method of step 4 is as follows:

[0125] S41, test the autonomous driving decision system;

[0126] S42. Setting evaluation indicators for the autonomous driving decision system;

[0127] S43. Analyze the performance of the autonomous driving decision system.

[0128] Furthermore, the specific method of S41 is as follows:

[0129] S411, perform model-in-the-loop testing;

[0130] By inputting the self-driving car, surrounding vehicles, pedestrians and traffic light signals into the trained model through the interface in the simulation software, the model calculates and outputs steering wheel angle, accelerator pedal opening and brake pedal opening signal control instructions, and sends the control signals to the autonomous driving vehicle through the virtual interface to realize the status update of the autonomous driving vehicle; set the starting point and end point in the simulation software, randomly generate traffic participants with different densities and different flow rates, and pass through sections of road with different curvatures, intersections with traffic lights, intersections without traffic lights, and roundabout sections; the autonomous driving vehicle drives from the starting point to the end point, and the simulation ends when a collision occurs. If there is no collision and the vehicle drives in accordance with traffic rules, it is recorded as a successful test, and the number of collisions is recorded;

[0131] S412, conduct real vehicle online testing;

[0132] The algorithm is tested online on the road sections where data was collected. The autonomous driving decision-making system is tested for long-distance driving, lane keeping, lane changing, and overtaking in highway environments, as well as its ability to navigate congested sections, multi-lane intersections, and intersections with or without traffic lights in urban road environments. Real-vehicle tests are conducted under varying traffic densities and flow rates, with set starting and end points. If the autonomous driving decision-making system is unable to cope with the current driving scenario, the driver is allowed to take over control of the vehicle at any time. When the driver takes over, it indicates that the autonomous driving decision-making system has failed in the current scenario, and the number of times the driver takes over is recorded.

[0133] The specific method of S42 is as follows:

[0134] S421, conduct driving safety evaluation;

[0135] In the model-in-the-loop test, the collision rate is used as an indicator to evaluate the driving safety of the autonomous vehicle. In the real vehicle online test, the takeover rate is used as an indicator to evaluate the driving safety. The calculation formulas (39) and (40) of the collision rate and takeover rate are as follows:

[0136] (39)

[0137] (40)

[0138] Where, T tot is the total number of trials; T c is the number of collisions that occurred in the virtual simulation test; T t The number of times the driver took over during the actual vehicle test;

[0139] S422, conduct driving efficiency evaluation;

[0140] The performance of the decision-making system is evaluated by the driving efficiency of the autonomous vehicle. The driving efficiency calculation formula (41) is as follows:

[0141] (41)

[0142] Where, T e The number of tests required to meet the minimum travel time; T suc is the number of successes;

[0143] S423, conduct driving compliance evaluation;

[0144] The performance of the decision-making system is evaluated by the driving compliance of the autonomous vehicle. The driving compliance calculation formula (42) is as follows:

[0145] (42)

[0146] Where, T l The number of tests required to meet traffic regulations;

[0147] The specific method of S43 is as follows:

[0148] S431. Obtaining model-in-the-loop test results;

[0149] Evaluate the decision-making performance of the autonomous vehicle by analyzing the results of driving safety, driving efficiency, and driving compliance evaluations, calculate the overall collision rate, evaluate the system's overall collision situation in all test scenarios, and draw a histogram or pie chart to show the distribution of the overall collision rate; calculate the collision rate for each specific scenario separately, evaluate the system's collision situation in different scenarios, draw a bar chart or line chart to show the difference in collision rate in each scenario, and record the collision accident's vehicle, adjacent vehicles, pedestrians, and environmental information;

[0150] S432, obtaining the actual vehicle online test results;

[0151] By drawing histograms or pie charts, the driving safety, driving efficiency, and driving compliance evaluation results of autonomous vehicles in various scenarios are statistically analyzed. During the driving process of the autonomous vehicle, the information of the autonomous vehicle, adjacent vehicles, pedestrians, and the environment when the driver takes over is recorded.

[0152] S433. Optimizing result feedback;

[0153] Analyze and record the model's in-the-loop test results and real-vehicle online test results; study the model's decision-making mechanism in failed use cases, locate the cause of decision failure, and improve the algorithm's performance; analyze the cause of each collision, draw a grouped bar chart, and display the number of collisions caused by different reasons.

[0154] The beneficial effects of the present invention are:

[0155] 1. The causal knowledge-based explainable autonomous driving decision-making method described in this invention includes four steps: causal discovery of autonomous driving decision-making knowledge, construction of an explainable autonomous driving decision-making model, evaluation of the explainability of the autonomous driving decision-making system, and evaluation of the performance of the autonomous driving decision-making system. Based on the causal discovery model of decision observation data, the method identifies key variables in the causal relationship that influences the vehicle's driving behavior. Explainability standards and decision-making performance evaluation indicators are defined, and the explainability and decision-making performance of the autonomous driving decision-making system are evaluated, continuously optimizing and iterating the autonomous driving decision-making system.

[0156] 2. The causal knowledge-based explainable autonomous driving decision-making method described in this invention constructs a causal discovery model for autonomous driving decision-making knowledge, providing key variables that influence the driving behavior of the vehicle for the decision-making model based on spatiotemporal graph neural network;

[0157] 3. The causal knowledge-based explainable autonomous driving decision-making method described in this invention constructs an explainable autonomous driving decision-making model, implements the embedded representation of decision-making causal knowledge and the spatiotemporal correlation learning of "self-vehicle-scene elements", and outputs safe and accurate driving behavior action sequences based on a spatiotemporal graph neural network;

[0158] 4. The causal knowledge-based explainable autonomous driving decision-making method described in this invention, by defining interpretability criteria, enables both ex ante and ex post explanations of the autonomous driving decision-making model, as well as objective evaluation of the autonomous driving decision-making system. It also identifies key features that influence autonomous vehicle behavior decisions in different scenarios and conducts statistical and visual analysis.

[0159] 5. The causal knowledge-based explainable autonomous driving decision-making method described in the present invention analyzes the decision-making performance of the autonomous driving decision-making system through safety, comfort, and compliance evaluation indicators through model-in-the-loop testing and real-vehicle online testing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0160] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0161] Figure 1 It is a schematic diagram of the process of the present invention;

[0162] Figure 2 Schematic diagram of step 1 of the present invention;

[0163] Figure 3 This is a schematic diagram of step 2 of the present invention;

[0164] Figure 4 Schematic diagram of step three of the present invention;

[0165] Figure 5 Schematic diagram of step 4 of the present invention;

[0166] Figure 6 This is an example diagram of the causal discovery model algorithm based on decision observation data;

[0167] Figure 7 This is an example picture of a driving scene;

[0168] Figure 8 This is an example diagram of an explainable autonomous driving decision algorithm. DETAILED DESCRIPTION

[0169] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It will be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all structures.

[0170] See Figure 1 The present invention provides an explainable autonomous driving decision-making method based on causal knowledge, comprising the following steps:

[0171] See Figure 2 and Figure 6 ,Step 1: Discover the causal relationship of automatic decision-making knowledge, as follows:

[0172] S11. Collect multi-source heterogeneous driving data, as follows:

[0173] S111, collecting virtual simulation data;

[0174] By building a driving simulator and simulation platform, various driving scenarios are designed, including urban roads, highways, and rural roads, covering different weather conditions, traffic flows, and road conditions; dynamic elements are added, including vehicles, pedestrians, and bicycles; and real traffic environments are simulated, including roads, buildings, traffic signs, and signal lights. A complete virtual driving environment is constructed, and the characteristic information of all scene elements constitutes a multi-source heterogeneous driving data set in the virtual simulation environment. v , as shown in formula (1):

[0175] (1)

[0176] Where, h e 、 h n 、 h b 、 h p 、 h l 、 h s are the feature sets of the scene elements of the vehicle, adjacent vehicles, adjacent bicycles, pedestrians, traffic lights and signboards respectively; among them, h e ={ x e , y e , vx e, vy e, se , t e , b e}, h n ={ s n , x n , y n , vx n, vy n}, h b ={ x b , y b , vx b, vy b}, h p ={ x p , y p , vx p, vy p}, h l ={ x l , y l , s l}, h s ={ x s , y s , s s}; x e 、y e 、vx e、vy e、s e 、t e and b e They are the lateral position, longitudinal position, lateral speed, longitudinal speed, steering wheel angle, accelerator pedal opening and brake pedal opening of the vehicle; s n 、x n 、y n 、vx n、vy n are the type, lateral position, longitudinal position, lateral speed, and longitudinal speed of the adjacent vehicles respectively; x b 、y b , vx b, vy b are theLateral position, longitudinal position, lateral speed, longitudinal speed; x p 、y p 、vx p、vy p They are the pedestrian's lateral position, longitudinal position, lateral speed, and longitudinal speed respectively; x l 、y l 、s l They are the horizontal position, vertical position and status of the traffic light respectively; x s 、y s 、s s They are the horizontal position, vertical position and status of the traffic sign respectively;

[0177] S112, collecting real vehicle data;

[0178] The raw data obtained by the inertial navigation positioning device, laser radar, millimeter wave radar, and camera are processed to obtain target-level data. The data type is the same as step 111), and a multi-source heterogeneous driving data set Ω is obtained in a real vehicle environment. r ;

[0179] Among them, the processing can be carried out through sensor fusion algorithms based on deep learning or other algorithms.

[0180] S113, classifying the scene;

[0181] The simulation data collected by S111 and S112 are v and multi-source heterogeneous driving data set, namely real vehicle driving data Ω r Slice the scene according to the road topology and determine the starting and ending points of each scene slice. The road topology is divided into roads of different curvatures, intersections with and without traffic lights, T-intersections, and roundabouts. Ensure that the road, traffic signs, signal lights and other elements within the slice are complete and consistent. Further classify the scene segments according to the number of vehicles in the surrounding environment, and the same number of vehicles in the surrounding environment are classified into the same type of scene.

[0182] S114, integrating multi-source heterogeneous driving data;

[0183] The scene slices generated in S113 are stored as independent simulation scenes, and a slice index is established to save the type, location, and environment information of each slice for easy query and management. The slice library is expanded as needed to add new slice types and scenes to improve the richness and coverage of the data, and a classification set Ω of surrounding vehicles of different road sections and different numbers is obtained. v,n and Ω r,n ;

[0184] Expanding the slice library refers to slicing by scenario type and scenario. For example, a library has three categories of books: literature, engineering, and linguistics. When a new history book is stored in the library, a new location (i.e., ID) is assigned to it.

[0185] S12. Establish a causal discovery model based on decision observation data, as follows:

[0186] S121, preprocessing the data;

[0187] The obtained Ω v,n and Ω r,n Perform data cleaning and data standardization; first, fill the missing values with the Lagrange interpolation method, and use data standardization to convert data of different scales into a unified scale. The calculation formula is (2):

[0188] (2)

[0189] Where, X max and X min are the maximum and minimum values of the data respectively;

[0190] S122, build a causal relationship model;

[0191] A kernel regression model is used as a nonlinear model, with surrounding vehicles, traffic lights, and traffic signs as variables. Based on the ego vehicle's driving behavior on different road sections, with different numbers of surrounding vehicles, and with or without traffic lights and traffic signs, the nonlinear causal relationship between the variables that influence ego vehicle driving behavior is analyzed.

[0192] An exemplary embodiment of this step is as follows Figure 6 shown.

[0193] The Hilbert-Schmidt Independence Criterion (HSIC) is used to determine the input feature variables. X ={ h e , h n , h b , h p , h l , h s , h r}, and the target feature variable Y={ s e , t e , b e}, h e 、 h n 、 h b 、 h p 、 h l 、 h s 、 h r 、 s e 、 t e and b e From the first step, step by step, step by step, definition. First, for each feature X p and Y q , define the kernel matrix K p and L q Formulas (3) and (4) are:

[0194] (3)

[0195] (4)

[0196] Where, k p Features X p The kernel function of l q is the target variable Y q The kernel function of n is the dimension of the input feature variable;

[0197] The Laplace kernel function is used to perform nonlinear mapping on the input feature variables and the target feature variables. The calculation formulas (5) and (6) of the Laplace kernel function are as follows:

[0198] (5)

[0199] (6)

[0200] Where ||.|| is the Euclidean distance between two input vectors; x andy They are nuclear center; and The bandwidth parameter that controls the smoothness of the kernel function;

[0201] In order to eliminate the bias, the kernel matrix is centered and the calculation formulas are (7) and (8):

[0202] (7)

[0203] (8)

[0204] Where, I is the identity matrix; A is a vector of all 1s, .

[0205] For each feature and each target variable, the HSIC between the two is calculated using formula (9):

[0206] (9)

[0207] In the formula, the larger the HSIC value, the better the feature X p With the target variable Y q The stronger the dependence between them; tr(.) is the trace of the matrix, that is, the sum of the diagonal elements of the matrix.

[0208] Using the Hilbert-Schmidt independence criterion and the elastic network regression model, a nonlinear feature selection model is constructed. The optimization objective function formula (10) is:

[0209] (10)

[0210] Where, β q For the q The regression coefficient of the target variable; is the regularization parameter; is the trade-off parameter between L1 and L2 regularization.

[0211] Through this optimization objective function, select the target variable Y The most dependent nonlinear feature. The coordinate descent method is used to solve the objective function. First, the regression coefficients are initialized. β is a zero matrix, For each target variable Y q , calculate the loss term and regularization term in the objective function, when the regression coefficient matrix β When the update amplitude is small enough, the iteration is stopped. The final regression coefficient matrixβ Represents each feature X p With each target variable Y q The dependency between β pq Representation characteristics X p right Y q Have a strong impact.

[0212] S123, fitting and evaluating the model;

[0213] The Granger causality test is used to fit the model parameters and determine the causal skeleton structure through significance test. A causal graph is generated to represent the causal relationship between the variables that affect the driving behavior of the vehicle, and the degree of influence of different variables on the driving behavior of the vehicle is quantified. Finally, a directed acyclic graph (DAG) representing the causal relationship is output. G c ;

[0214] S13. Integrate autonomous driving decision-making knowledge, as follows:

[0215] S131. Define road segment driving knowledge;

[0216] Integrate the knowledge that influences autonomous vehicle behavior decisions on a road segment; consider the interactions between the ego vehicle and surrounding vehicles and traffic signs, use a causal diagram to represent the causal relationships between the variables that influence the ego vehicle's driving behavior, and categorize the causal diagram based on the presence or absence of traffic signs and the number of vehicles in the surrounding environment;

[0217] S132. Define intersection decision knowledge;

[0218] Define the decision-making knowledge that affects the ego vehicle's driving behavior in intersection scenarios; consider the impact of traffic lights, pedestrians, and environmental vehicles on the ego vehicle; and classify and save causal graphs that can represent the causal relationship between variables that affect the ego vehicle's driving behavior based on the presence or absence of traffic lights, the number of pedestrians, and the number of environmental vehicles.

[0219] Step 2: Refer to Figure 3 and Figure 7 , build an explainable autonomous driving decision model. The specific method is as follows:

[0220] S21. Construct a driving scenario for an autonomous vehicle, as follows:

[0221] S211, extracting scene elements;

[0222] Extract scene elements from the autonomous vehicle driving scene. Dynamic scene elements include the autonomous vehicle, adjacent vehicles, pedestrians, bicycles, and traffic lights. Static scene elements include traffic signs. Scene element features are the same as those defined in S111.

[0223] S212, extracting causal relationships;

[0224] The similarity between the autonomous driving decision causal graph and the spatiotemporal graph structure integrated by S13 is calculated. The causal graph with high similarity (customizable, greater than 0.8 is considered high similarity) is selected, and the autonomous driving decision causal knowledge is initialized to the spatiotemporal graph. The similarity calculation is shown in Equation (11):

[0225] (11)

[0226] Where, G c and G d They are cause-effect diagram and driving scenario diagram respectively; mcs ( G c , G d ) is the number of nodes in the largest common subgraph of the two graphs; max (| G c |,| G d |) is the ratio of the product of the number of nodes in the two graphs;

[0227] S213, representing an autonomous driving driving scenario based on a spatiotemporal graph structure;

[0228] The driving scene of the autonomous vehicle is represented by a spatiotemporal graph. At a certain moment, the scene elements are represented by nodes of the graph, and the relationship between the scene elements is represented by directed edges between nodes. The dynamic scene modeling is as follows: on the basis of the graph structure, the time dimension is added to form a sequential graph model, such as Figure 7 As shown, the graph structure is expressed as shown in formula (12):

[0229] (12)

[0230] Where, V ={ v e , v n , v b , v p , v l , v s};node v e , v n , v b , v p , v l , v s They are the own car, neighboring car, bicycle, pedestrian, traffic light and sign; E ={ ee n , ee b , ee p , ee l , ee s}, ee n , ee b , ee p , ee l , ee s are directed edges between the vehicle and neighboring vehicles, bicycles, pedestrians, traffic lights, and traffic signs; the connection relationship of the directed edges is determined by the distance between the nodes. v e and v i distance d ( v e , v i ) is calculated using formulas (13) and (14):

[0231] (13)

[0232] (14)

[0233] when d ( v e , v i )<100m, v e and v i There are directed edges between them; the weights of the directed edges are initialized by the extracted causal graph to ensure that the sum of the weights of the directed edges of all nodes is 1;

[0234] S22. Build an autonomous driving decision model, as follows:

[0235] S221. Build a dataset for autonomous driving decision-making;

[0236] The collected autonomous driving vehicle driving data is divided into data sets; a fixed-size sliding window is used to create training samples, wherein the window can be slid at a frequency of 60 Hz, that is, every 60 consecutive discrete points; the historical 3 seconds of autonomous driving vehicle driving data is used to predict the driving behavior action sequence of the next 2 seconds; 70%, 20% and 10% of the sample number are divided into training set, test set and validation set respectively; the number of samples is 100%. N tot The calculation formula is shown in (15):

[0237] (15)

[0238] Where, N is the total length of the time series; the window size W The number of time steps included for each sample; step size S The number of time steps for each window movement;

[0239] S222, embedding causal knowledge in autonomous driving decision making;

[0240] Calculate S212 to obtain a causal graph with high similarity, extract the causal relationship of the causal graph, and embed the causal relationship into the spatiotemporal graph model; including the relationship between nodes and edges that affect the driving behavior of the vehicle;

[0241] S223, learning the spatiotemporal correlation between “ego vehicle and scene elements”;

[0242] The spatiotemporal interaction between the ego vehicle and scene elements is learned through a spatiotemporal graph neural network. First, a temporal attention mechanism is used to calculate the feature representation of the ego vehicle node. Then, a graph convolutional neural network is designed to mine the graph structure at each moment to obtain the feature representation of the ego vehicle node that aggregates surrounding nodes. A temporal attention mechanism is used to learn the importance of the ego vehicle features at each moment. Finally, a channel attention mechanism is used to fuse the spatiotemporal correlations between the ego vehicle and scene elements.

[0243] S224, mapping the driving behavior of the vehicle;

[0244] The encoding vector learned in S223 is mapped into a three-dimensional vector through dimensional transformation, deformation and linear transformation, which respectively represents the steering wheel angle, accelerator pedal opening and brake pedal opening of the vehicle;

[0245] S23. Evaluation model offline test results, as follows:

[0246] S231. Define evaluation indicators;

[0247] Model evaluation metrics are defined, using mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) as indicators to evaluate the difference between the model's predicted value and the true value. The calculation formulas (16), (17), and (18) for MAE, RMSE, and MAPE are as follows:

[0248] (16)

[0249] (17)

[0250] (18)

[0251] Where, and They are the actual driving action sequence and the predicted driving behavior action sequence of the autonomous driving vehicle respectively; N Indicates the number of samples;

[0252] S232, performing offline testing on the model;

[0253] The authors analyzed MAE, RMSE, and MAPE to measure the deviation between the predicted and true values of the steering wheel angle, accelerator pedal opening, and brake pedal opening of autonomous vehicles. The autonomous driving decision-making dataset was divided into several parts, and the model was trained and validated on different data subsets. The dataset was evenly divided into k mutually exclusive subsets. Each time, k-1 subsets were selected for training, and the remaining subset was used for validation. The model was trained and validated k times, each with a different validation set. Finally, the average of the k validation results was taken as the model evaluation metric. The value of k was 5 or 10.

[0254] Exemplary embodiments of S22 and S23 are as follows Figure 8 shown.

[0255] S22 builds an autonomous driving decision model. According to the self-driving car driving data in S221, after dividing the data set, the scene fragments at each moment are represented by a graph structure. The representation of the graph is represented by S213. The continuous scene composed of each sample is represented by a list of graphs. When constructing the graph structure and feature data, all tensor parameters are copied to the GPU device. When training the model, first load the data set constructed by the self-driving car. According to the causal relationship graph extracted by S212, the causal relationship is used to guide the decision of the model. Then, the self-driving car time correlation learning is performed. The input is based on the self-driving car driving data. The dimension of the data is [ B × T ×F ], B Indicates the size of Batchsize, T represents the length of the input sequence, F Represents the dimension of the sequence, that is, the number of features of the ego vehicle. First, the matrix W is used to map the ego vehicle feature dimension, and the ego vehicle feature dimension is mapped to D dimension, D is a hyperparameter. Then, the temporal attention mechanism is used to learn the impact of the vehicle’s historical data on future data. The temporal attention weight calculation formula (19) is expressed as:

[0256] (19)

[0257] Where, W Q , W K are the weight matrices to be trained respectively; α t,i For the moment t and i relationship;

[0258] The feature representation of the ego-vehicle node is calculated through the temporal attention mechanism, and the calculation formula (20) is:

[0259] (20)

[0260] Then, the spatial correlation learning of “car-scene elements” is performed. First, all the feature dimensions of heterogeneous nodes are mapped to the same dimension. Then, the features are aggregated according to the different connection relationships between nodes. Assuming that each type of edge For a certain connection relationship, define a message constructor MsgE to generate v i To the vehicle node v e The message constructor formula (21) can be expressed as:

[0261] (twenty one)

[0262] Where, E i,v The feature of the edge.

[0263] For the self-driving node v e , all types are E i Aggregate the messages of the edges to get the aggregated message. The aggregation formula (22) is:

[0264] (twenty two)

[0265] Where Agg Ei is an aggregation function, such as mean aggregation or max pooling.

[0266] The node features are updated using the aggregated messages of all types of edges, and the calculation formula (23) is:

[0267] (twenty three)

[0268] Where, For the l Layer Node v e The hidden state of ; Update is an update function.

[0269] Finally, propagation is performed through the graph convolutional network, and the calculation formula (24) is:

[0270] (twenty four)

[0271] Where, and is a learnable weight matrix representing the activation function.

[0272] Finally, the spatiotemporal correlation of the vehicle and scene elements is fused through the channel attention mechanism to obtain the fused representation result. and Perform global average pooling to obtain two vectors Z 1 and Z 2, the calculation formulas (25) and (26) are:

[0273] (25)

[0274] (26)

[0275] Where, Z 1 and Z The shape of 2 is 1× f .

[0276] Then, a fully connected layer and a nonlinear activation function are used to learn the channel weights, and the calculation formulas (27) and (28) are:

[0277] (27)

[0278] (28)

[0279] Formula (27) and Formula (28) use ReLU activation function and Sigmoid activation function respectively. W r andW s is the weight matrix of the fully connected layer.

[0280] Use the learned channel weights S 1 and S 2 pairs of input tensors and Perform weighted fusion to obtain the final representation result. The calculation formula (29) is:

[0281] (29)

[0282] Where, is element-wise multiplication, .

[0283] Then, two LSTM networks are used to map the driving behavior features and output driving behavior action information, including steering wheel angle, accelerator pedal opening, and brake pedal opening. The calculation formulas (30) and (31) are:

[0284] (30)

[0285] (31)

[0286] Where, is the model prediction result, , ; is the length of the output sequence.

[0287] Loss function during training L The calculation formula (32) is:

[0288] (32)

[0289] Where, and are the true value and predicted value output by the model respectively.

[0290] The MAE, RMSE, and MAPE in S23 are used to evaluate the training effect of the model. If the training loss of the model changes by less than 0.0001 within 5 consecutive epochs, or when the set maximum number of epochs is reached, the training is stopped. Usually, after each epoch, the performance of the model on the validation set is checked. If the performance is better than the previous best performance, the current model weights are saved. A set of hyperparameters for the model is determined, the model is trained with the training set, the optimal function that minimizes the loss function is found, the performance of the optimal function is measured on the validation set, and the optimal specified hyperparameter combination is determined. The model parameters are saved, loaded, and verified on the test set and validation set. The model with the smallest error on the validation set is selected, and the training set and validation set are merged as the overall training model to find the optimal function; the generalization performance of the optimal function is measured on the test set.

[0291] Step 3: Refer to Figure 4 , evaluate the interpretability of autonomous driving decision systems. The specific methods are as follows:

[0292] S31. Define the interpretability criteria as follows:

[0293] S311, analytical model transparency;

[0294] Analyze the model's own decision-making mechanism and obtain explanations by studying the model's internal structure and parameters; use the model's prior explanation results as the model transparency analysis results; evaluate and score the model's input layer, graph attention layer, temporal convolution layer, spatiotemporal fusion layer, fully connected layer, normalization layer, activation layer, residual connection layer, and output layer. The specific formula and operation of graph convolution explain how node features are aggregated through neighboring nodes; similar input data has stable output results at each layer; output model transparency analysis results U ;

[0295] S312, analyze the sensitivity of the model;

[0296] By introducing the perturbation analysis feature importance as an indicator for evaluating the model's ex post interpretability; using the counterfactual explanation method to change the model's input features before prediction and analyze the model's prediction changes; using the counterfactual explanation results as the model sensitivity analysis results, and outputting the model sensitivity analysis results V ;

[0297] S313, conduct objective evaluation;

[0298] Objectively evaluate the autonomous driving system; score the scene elements that affect the self-driving behavior according to the preset situation. Each scene element has a unique corresponding score. The degree of influence on the self-driving behavior is classified into strong, general and weak according to the score. Finally, objective evaluation results of the scene elements that affect the self-driving behavior in different scenarios are obtained. W ;

[0299] S314. Define the comprehensive evaluation model of interpretability;

[0300] Define an interpretability comprehensive evaluation model, comprehensively evaluate the transparency, sensitivity and objective evaluation results, and output the comprehensive evaluation results; analyze the key features obtained U 、 V and W Sort the importance, sorting rules, and calculation formula (33) are:

[0301] (33)

[0302] Where, S The intersection set of key features obtained by S311, S312 and S313, the remaining features are based on the principle of two-by-two intersection, and the priority size is defined as > > > > U > W > V , obtain the feature importance ranking results;

[0303] S32. Provide a preliminary explanation of the autonomous driving decision model, as follows:

[0304] S321, analyze the model structure;

[0305] The graph structure defined in step S211 is analyzed. The edges of the graph structure are key factors in message transmission. The connecting edges between nodes are determined by the interactive relationships between scene elements with different semantics and features. To distinguish the importance of different edges in message transmission, the fixed weights of different edges are first initialized according to the causal relationship model constructed in step S122. During the model training process, the weights are parameterized, and the edge weights are updated through backpropagation and gradient descent algorithms. The learned edge weights are then visualized.

[0306] S322, analyzing and visualizing node importance;

[0307] Select and extract features that have a significant impact on model predictions and visualize the results. Based on the adjacency matrix between the ego vehicle node and other scene element nodes, the ego vehicle node is used as the central node, and the features of surrounding nodes are aggregated. The importance of the nodes is visualized through a heat map.

[0308] S323, obtaining prior explanation results;

[0309] The nodes with the greatest impact on the vehicle's driving behavior in the same scenario are counted and sorted according to the calculation results of the node importance. The frequency distribution of characteristic variables with different degrees of influence on the vehicle's behavior decision is displayed through histograms and box plots; the ranking results of the influence of different node features on the model prediction results are determined. U , generate a model explanation report, describing in detail the distribution of each feature on the prediction results;

[0310] S33. Provide a post-hoc explanation of the autonomous driving decision model, as follows:

[0311] S331, input key variables;

[0312] The following steps use the autonomous driving decision model trained by S22 f , key variables defined in S111 X i ;

[0313] S332, construct counterfactual scenarios;

[0314] By constructing counterfactual scenarios, using do operators to generate counterfactuals, and intervening in multiple features X 1, X 2, …, X k , generate new input features, and the other features remain unchanged. The calculation formula (34) is:

[0315] (34)

[0316] Where, v i is the value after intervention, and the generated counterfactual state Is a Similar inputs but with some features interfered with; then using the interfered input state to change the vehicle’s driving environment and calculate the counterfactual decision result , the calculation formula (35) is:

[0317] (35)

[0318] In order to find the best counterfactual state, an optimization algorithm is used to minimize the comprehensive loss function. The loss function calculation formula (36) is:

[0319] (36)

[0320] Where the first term is the decision difference and the second term is the similarity of the input. λ is a hyperparameter that balances these two terms; use the gradient descent algorithm to solve and find ; is the input feature; is the state after intervention;

[0321] Use weighted Euclidean distance evaluation X and The similarity of , is calculated by formula (37):

[0322] (37)

[0323] S333, key factors for evaluation;

[0324] Evaluate key variables in ego vehicle driving behavior; first, evaluate counterfactual decision outcomes The impact of, including safety, efficiency and compliance, defines an impact function The difference value is quantified, and the difference value quantification calculation formula (38) is:

[0325] (38)

[0326] Where, s 、 e and l are the safety, efficiency and compliance evaluation score functions respectively; w s 、 w e and w l are the weight coefficients of safety, efficiency and compliance respectively; find the counterfactual decision based on the difference value safety, efficiency and compliance of key variables;

[0327] S334, obtain post-explanation results;

[0328] Analyze key variables, use histograms and box plots to count key variables that affect safety, efficiency, and compliance, and record counterfactual instance scenarios and key variables; determine the ranking results of key variables that affect the safety, efficiency, and compliance of the vehicle under the same scenario V .

[0329] S34. Evaluate the autonomous driving decision-making system as follows:

[0330] S341, define questionnaire;

[0331] The evaluation scores were obtained based on a questionnaire designed to objectively evaluate the autonomous driving decision-making system. The questionnaire was designed as follows: the lane-changing scenario identified the scene elements and characteristics that affected the ego vehicle's lane change; the emergency braking scenario identified the scene elements and characteristics that affected the ego vehicle's braking; and the intersection scenario with traffic lights identified the scene elements and characteristics that affected the ego vehicle's driving behavior.

[0332] S342. Statistically analyze the evaluation results of the questionnaire in S41, obtain a ranking result W of scene elements and key feature variables based on the evaluation scores, and intuitively display the degree of influence of different scene elements on vehicle driving behavior through a graphical visualization method.

[0333] Step 4: Refer to Figure 5 , to evaluate the performance of the autonomous driving decision system. The specific method is as follows:

[0334] S41. The autonomous driving decision system is tested as follows:

[0335] S411, perform model-in-the-loop testing;

[0336] By using the interface in the simulation software to input the self-driving car, environmental vehicles, pedestrians and traffic light signals into the trained model, the model calculates and outputs the steering wheel angle, accelerator pedal opening and brake pedal opening signal control instructions, and sends the control signals to the autonomous driving vehicle through the virtual interface to achieve the status update of the autonomous driving vehicle. Set the starting point and end point in the simulation software, randomly generate traffic participants of different densities and different flow rates, and pass through sections of road with different curvatures, intersections with traffic lights, intersections without traffic lights, roundabouts and other sections. The number of experimental tests is set to 20,000 times. The autonomous driving vehicle drives from the starting point to the end point, and the simulation ends when a collision occurs. If there is no collision and the vehicle drives in accordance with traffic rules, it is recorded as a successful test, and the number of collisions is recorded;

[0337] S412, conduct real vehicle online testing;

[0338] This step conducts real-vehicle online testing of the algorithm on the road section where S11 collects data, and tests the functions of the autonomous driving decision-making system in long-distance driving, lane keeping, lane changing, overtaking, etc. in a highway environment, as well as the traffic capacity of congested sections, multi-lane intersections, and intersections with or without traffic lights in an urban road environment. Real-vehicle tests are conducted under different traffic densities and flow rates, with a set starting point and end point, and the number of tests is also set to 20,000 times. In order to ensure driving safety, when the autonomous driving decision-making system is unable to cope with the current driving scenario, the driver can take over control of the vehicle at any time; when the driver takes over, it means that the autonomous driving decision-making system has failed to make a decision in the current scenario, and the number of times the driver takes over is recorded;

[0339] S42. Set evaluation indicators for the autonomous driving decision system, as follows:

[0340] S421, conduct driving safety evaluation;

[0341] In the model-in-the-loop test, the collision rate is used as an indicator to evaluate the driving safety of the autonomous vehicle. In the real vehicle online test, the takeover rate is used as an indicator to evaluate the driving safety. The calculation formulas (39) and (40) of the collision rate and takeover rate are as follows:

[0342] (39)

[0343] (40)

[0344] Where, T tot is the total number of trials; T c is the number of collisions that occurred in the virtual simulation test; T t The number of times the driver took over during the actual vehicle test;

[0345] S422, conduct driving efficiency evaluation;

[0346] The performance of the decision-making system is evaluated by the driving efficiency of the autonomous vehicle. The driving efficiency calculation formula (41) is as follows:

[0347] (41)

[0348] Where, T e The number of tests required to meet the minimum travel time; T suc is the number of successes;

[0349] S423, conduct driving compliance evaluation;

[0350] The performance of the decision-making system is evaluated by the driving compliance of the autonomous vehicle. The driving compliance calculation formula (42) is as follows:

[0351] (42)

[0352] Where, T l The number of tests required to meet traffic regulations;

[0353] S43. Analyze the performance of the autonomous driving decision system as follows:

[0354] S431. Obtaining model-in-the-loop test results;

[0355] Evaluate the decision-making performance of autonomous vehicles by analyzing driving safety, driving efficiency, and driving compliance evaluation results, calculate the overall collision rate, evaluate the system's overall collision situation in all test scenarios, and draw a histogram or pie chart to show the distribution of the overall collision rate; calculate the collision rate for each specific scenario separately, evaluate the system's collision situation in different scenarios, draw a bar chart or line chart to show the difference in collision rate in each scenario, and record environmental information such as the vehicle involved in the collision, adjacent vehicles, pedestrians, and traffic lights;

[0356] S432, obtaining the actual vehicle online test results;

[0357] By drawing histograms or pie charts, the driving safety, driving efficiency, and driving compliance evaluation results of autonomous vehicles in various scenarios are statistically analyzed. During the driving process, the autonomous vehicle records environmental information such as the vehicle itself, adjacent vehicles, pedestrians, and traffic lights when the driver takes over.

[0358] S433. Optimizing result feedback;

[0359] This phase analyzes and records the results of model-in-the-loop testing and on-vehicle online testing. Failure cases are used to further investigate the model's decision-making mechanisms, identify the causes of decision failures, and continuously improve algorithm performance. The causes of each collision are analyzed, such as decision errors, sensor failures, and environmental interference. Grouped bar charts are then created to display the number of collisions caused by different reasons.

[0360] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An explainable autonomous driving decision-making method based on causal knowledge, characterized by: The following steps are involved: Step 1: Discover the causal relationship for automated decision-making knowledge. The specific method is as follows: 11) Collect multi-source heterogeneous driving data, specifically: 111) collecting virtual simulation data; By building a driving simulator and simulation platform, various driving scenarios are designed, including urban roads, highways, and rural roads, covering different weather conditions, traffic flows, and road conditions; dynamic elements are added, including vehicles, pedestrians, and bicycles; and real traffic environments are simulated, including roads, buildings, traffic signs, and signal lights. A complete virtual driving environment is constructed, and the characteristic information of all scene elements constitutes a multi-source heterogeneous driving data set in the virtual simulation environment. v , as shown in formula (1): Ω v ={h e , h n , h b , h p , h l ,h s } (1) Where h e 、h n 、h b 、h p 、h l 、h s are the feature sets of the scene elements of the vehicle, adjacent vehicles, adjacent bicycles, pedestrians, traffic lights and signboards respectively; among them, h l ={x l ,y l ,s l }, h s ={x s ,y s ,s s };x e 、y e 、 s e , t e and b e are the lateral position, longitudinal position, lateral speed, longitudinal speed, steering wheel angle, accelerator pedal opening and brake pedal opening of the vehicle respectively; s n 、x n 、y n 、 are the type, lateral position, longitudinal position, lateral speed, and longitudinal speed of the adjacent vehicle respectively; x b 、y b 、 are the lateral position, longitudinal position, lateral speed, and longitudinal speed of the bicycle respectively; x p 、y p 、 are the pedestrian's lateral position, longitudinal position, lateral speed, and longitudinal speed respectively; x l 、y l 、s l are the horizontal position, vertical position and state of the traffic light respectively; x s 、y s 、s s They are the horizontal position, vertical position and status of the traffic sign respectively; 112) Collecting real vehicle data; The raw data obtained by the inertial navigation positioning device, laser radar, millimeter wave radar, and camera are processed to obtain target-level data, and a multi-source heterogeneous driving data set Ω is obtained in a real vehicle environment. r ; 113) Classify the scene; The simulation data Ω v and multi-source heterogeneous driving data set, namely real vehicle driving data Ω r Slice the scene according to the road topology and determine the start and end points of each scene slice. The road topology is divided into roads of different curvatures, intersections with and without traffic lights, T-intersections, and roundabouts. Ensure that the road, traffic signs, and signal light elements within the slice are complete and consistent. Further classify the scene segments according to the number of vehicles in the surrounding environment, and the same number of vehicles in the surrounding environment are classified into the same type of scene. 114) Integrate multi-source heterogeneous driving data; The generated scene slices are stored as independent simulation scenes, and a slice index is established to save the type, location, and environment information of each slice. The slice library is expanded as needed to add new slice types and scenes to obtain a classification set Ω for different road sections and different numbers of surrounding vehicles. v,n and Ω r,n ; 12) Build a causal discovery model based on decision-making observation data, specifically: 121) preprocessing the data; The obtained Ω v,n and Ω r,n Perform data cleaning and data standardization. First, fill the missing values with the Lagrange interpolation method, and use data standardization to convert data of different scales into a unified scale. The calculation formula is (2): Where, X max and X min are the maximum and minimum values of the data respectively; 122) Constructing causal relationship models; A kernel regression model is used as a nonlinear model, with surrounding vehicles, traffic lights, and traffic signs as variables. Based on the ego vehicle's driving behavior on different road sections, with different numbers of surrounding vehicles, and with or without traffic lights and traffic signs, the nonlinear causal relationship between the variables that influence ego vehicle driving behavior is analyzed. 123) Fit and evaluate the model; The Granger causality test is used to fit the model parameters and determine the causal skeleton structure through significance test; a causal graph is generated to represent the causal relationship between the variables that affect the driving behavior of the vehicle, and to quantify the degree of influence of different variables on the driving behavior of the vehicle, and finally output the DAG graph G c ; 13) Integrate autonomous driving decision-making knowledge, specifically: 131) Define road segment driving knowledge; Integrate the knowledge that influences autonomous vehicle behavior decisions on a road segment; consider the interactions between the ego vehicle and surrounding vehicles and traffic signs, use a causal diagram to represent the causal relationships between the variables that influence the ego vehicle's driving behavior, and categorize the causal diagram based on the presence or absence of traffic signs and the number of vehicles in the surrounding environment; 132) Define intersection decision-making knowledge; Define the knowledge that influences the ego vehicle's driving behavior in intersection scenarios; consider the impact of traffic lights, pedestrians, and surrounding vehicles on the ego vehicle; and categorize and save causal graphs that represent the causal relationships between variables that influence the ego vehicle's driving behavior, based on the presence or absence of traffic lights, the number of pedestrians, and the number of surrounding vehicles. Step 2: Build an explainable autonomous driving decision model; Step 3: Evaluate the interpretability of the autonomous driving decision-making system; Step 4: Evaluate the performance of the autonomous driving decision system.

2. The explainable autonomous driving decision-making method based on causal knowledge according to claim 1, characterized in that: The specific method of step 2 is as follows: 21) Constructing driving scenarios for autonomous vehicles; 22) Build an autonomous driving decision model; 23) Evaluate the model offline test results.

3. The explainable autonomous driving decision-making method based on causal knowledge according to claim 2, characterized in that: The specific method of step 21) is as follows: 211) extracting scene elements; Extract scene elements from the autonomous vehicle driving scene. Dynamic scene elements include the autonomous vehicle, adjacent vehicles, pedestrians, bicycles, and traffic lights. Static scene elements include traffic signs. 212) Extracting causal relationships; The similarity between the integrated autonomous driving decision causal graph and the spatiotemporal graph structure is calculated, and the causal graph with high similarity is selected. The spatiotemporal graph is initialized with the autonomous driving decision causal knowledge. The similarity calculation is shown in formula (11): Where G c and G d They are causal graph and driving scene graph respectively; mcs(G c ,G d ) is the number of nodes of the largest common subgraph of the two graphs; max(|G c |,|G d |) is the ratio of the product of the number of nodes in the two graphs; 213) represents an autonomous driving driving scenario based on a spatiotemporal graph structure; The driving scene of the autonomous vehicle is represented by a spatiotemporal graph. For a certain moment, the scene elements are represented by the nodes of the graph, and the relationship between the scene elements is represented by the directed edges between the nodes. The dynamic scene modeling is as follows: on the basis of the graph structure, the time dimension is added to form a sequential graph model. The graph structure is represented as shown in formula (12): G={V,E} (12) Where, V={v e ,v n ,v b ,v p ,v l ,v s }; node v e ,v n ,v b ,v p ,v l ,v s They are the own car, neighboring car, bicycle, pedestrian, traffic light and sign; e are directed edges between the vehicle and neighboring vehicles, bicycles, pedestrians, traffic lights, and traffic signs. The connection relationship of the directed edges is determined by the distance between the nodes. e and v i The distance d(v e ,v i ) is calculated using formulas (13) and (14) as follows: Where, v e,x is the lateral coordinate of the vehicle; v i,x is the node v connected to the vehicle i The horizontal coordinate of v e,y is the longitudinal coordinate of the vehicle; v i,y is the node v connected to the vehicle i The vertical coordinate of When d(v e ,v i )<100m, v e and v i There are directed edges between them; the weights of the directed edges are initialized by the extracted causal graph to ensure that the sum of the weights of the directed edges of all nodes is 1; The specific method of step 22) is as follows: 221) Build a dataset for autonomous driving decision-making; The collected autonomous vehicle driving data is divided into a data set; a fixed-size sliding window is used to create training samples; the historical 3s autonomous vehicle driving data is used to predict the driving behavior action sequence of the next 2s; 70%, 20% and 10% of the sample number are divided into training set, test set and validation set respectively; the number of samples N tot The calculation formula is shown in (15): Where N is the total length of the time series; the window size W is the number of time steps contained in each sample; the step size S is the number of time steps each time the window moves; 222) Embedding causal knowledge in autonomous driving decision making; The causal relationship of the causal graph with high similarity is extracted and embedded into the spatiotemporal graph model. Including the relationships between nodes and edges that affect the driving behavior of the vehicle; 223) Learning the spatiotemporal correlation between the vehicle and the scene elements; The spatiotemporal interaction relationship between the ego vehicle and scene elements is learned through a spatiotemporal graph neural network. First, the feature representation of the ego vehicle node is calculated through a temporal attention mechanism. Then, a graph convolutional neural network is designed to mine the graph structure at each moment, obtain the feature representation of the ego vehicle node that aggregates surrounding nodes, and learn the importance of the ego vehicle features at each moment through a temporal attention mechanism. Finally, a channel attention mechanism is used to fuse the spatiotemporal correlations between the ego vehicle and scene elements. 224) Mapping the driving behavior of the vehicle; The learned encoding vector is mapped into a three-dimensional vector through dimensional transformation, deformation and linear transformation, representing the steering wheel angle, accelerator pedal opening and brake pedal opening of the vehicle respectively; The specific method of step 23) is as follows: 231) Define evaluation indicators; The mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are used as indicators to evaluate the difference between the model prediction value and the true value. The calculation formulas (16), (17), and (18) of MAE, RMSE, and MAPE are as follows: Where y i and are the actual driving action sequence and the predicted driving behavior action sequence of the autonomous driving vehicle respectively; N represents the number of samples; 232) Perform offline testing on the model; By analyzing MAE, RMSE, and MAPE, the degree of deviation between the predicted values and the actual values of the steering wheel angle, accelerator pedal opening, and brake pedal opening of the autonomous driving vehicle is measured; the dataset for autonomous driving decisions is divided into several parts, and the model is trained and verified on different data subsets; the dataset is evenly divided into k mutually exclusive subsets; k-1 subsets are selected each time for training, and the remaining subset is used for verification; the model is trained and verified k times, and each time the verification set is different; finally, the average of the k verification results is taken as the model evaluation indicator.

4. The explainable autonomous driving decision-making method based on causal knowledge according to claim 1, characterized in that: The specific method of step three is as follows: 31) Define interpretability criteria; 32) Provide ex ante explanation of autonomous driving decision models; 33) Conduct post-hoc explanation of autonomous driving decision models; 34) Evaluate the autonomous driving decision-making system.

5. The explainable autonomous driving decision-making method based on causal knowledge according to claim 4, characterized in that: The specific method of step 31) is as follows: 311)Analysis model transparency; Analyze the model's decision-making mechanism and obtain explanations by studying the model's internal structure and parameters. Use the model's prior explanation results as the model transparency analysis results. Evaluate and score the model's input layer, graph attention layer, temporal convolution layer, spatiotemporal fusion layer, fully connected layer, normalization layer, activation layer, residual connection layer, and output layer. The specific formula and operation of graph convolution explain how node features are aggregated through neighboring nodes. Similar input data should have stable output results at each layer. Output the model transparency analysis result U. 312) Analyze the sensitivity of the model; By introducing perturbation analysis, feature importance is used as an indicator to evaluate the post-hoc interpretability of the model; Using the counterfactual explanation method, the model input features are modified before prediction to analyze the changes in model prediction; the counterfactual explanation results are used as the model sensitivity analysis results, and the model sensitivity analysis results V are output; 313) conduct objective evaluation; Objectively evaluate the autonomous driving system; score the scene elements that affect the ego vehicle's driving behavior based on preset conditions. Each scene element has a unique corresponding score. The degree of influence on the ego vehicle's driving behavior is classified according to the score as strong, medium, or weak. Ultimately, objective evaluation results W of the scene elements that affect the ego vehicle's driving behavior in different scenarios are obtained; 314) Define a comprehensive evaluation model for interpretability; Define an interpretability comprehensive evaluation model, and output the comprehensive evaluation results of transparency, sensitivity and objective evaluation. Then, sort the key feature analysis results U, V and W according to their importance. The sorting rules and calculation formula (33) are as follows: S=U∩V∩W (33) Where S is the intersection set of the key features obtained in step 311), step 312), and step 313. The remaining features are prioritized according to the principle of pairwise intersection, and the priority is defined as U∩V∩W>U∩V>V∩W>U∩W>U>W>V to obtain the feature importance ranking result. The specific method of step 32) is as follows: 321) Analyze the model structure; The defined graph structure is analyzed. The edges of the graph structure are key factors in message transmission. The connecting edges between nodes are determined by the interactive relationships between scene elements with different semantics and features. To distinguish the importance of different edges in message transmission, the fixed weights of different edges are first initialized according to the causal relationship model. During the model training process, the weights are parameterized, and the edge weights are updated through backpropagation and gradient descent algorithms. The learned edge weights are then visualized. 322) Analyze and visualize node importance; Select and extract features that have a significant impact on model predictions and visualize the results. Based on the adjacency matrix between the ego vehicle node and other scene element nodes, the ego vehicle node is used as the central node, and the features of surrounding nodes are aggregated. The importance of the nodes is visualized through a heat map. 323) obtain prior explanation results; The nodes with the greatest impact on the vehicle's driving behavior in the same scenario are counted and ranked according to the calculated results of the node importance. The frequency distribution of the characteristic variables with different degrees of influence on the vehicle's behavioral decision-making is displayed through histograms and box plots. The ranking result U of the influence of different node features on the model prediction results is determined, and a model explanatory report is generated, which details the distribution of each feature on the prediction results. The specific method of step 33) is as follows: 331) Input key variables; Input key variables; specifically: trained autonomous driving decision model f, defined key variables X i ; 332) Constructing counterfactual scenarios; By constructing counterfactual scenarios, we use the do operator to generate counterfactual scenarios and intervene in multiple features X1, X2, ..., X k , generate new input features, and the other features remain unchanged. The calculation formula (34) is: X i ′←do(X i ←v i ) i=1,2,...,k (34) Where, v i is the value after intervention, and the generated counterfactual state X i ′ is a i Similar inputs but with some features interfered with; then, using the interfered input state, the vehicle’s driving environment is changed and the counterfactual decision result Y′ is calculated using formula (35): Y′=f(X′) (35) In order to find the best counterfactual state, an optimization algorithm is used to minimize the comprehensive loss function. The loss function calculation formula (36) is: Where the first term is the decision difference, the second term is the input similarity, and λ is a hyperparameter that balances these two terms. Use the gradient descent algorithm to find X′; X i is the input feature; X′ is the state after intervention; w i is the weight vector to be learned; The weighted Euclidean distance is used to evaluate the similarity between X and X′, and the calculation formula (37) is: 333) Evaluate key factors; Evaluate the key variables of the ego vehicle's driving behavior. First, evaluate the impact of the counterfactual decision outcome Y′, including safety, efficiency, and compliance. Define an impact function E(Y, Y′) to quantify the difference value. The difference value quantification calculation formula (38) is: E(Y,Y′)=w s (s(Y)-s(Y′))+w e (e(Y)-e(Y′))+w l (l(Y)-l(Y′)) (38) Where s, e and l are the safety, efficiency and compliance evaluation score functions respectively; w s 、w e and w l are the weight coefficients of safety, efficiency and compliance respectively; according to the difference value, find the key variables that reduce the safety, efficiency and compliance of the counterfactual decision Y′; 334) Obtain post hoc interpretation results; Analyze key variables and use histograms and boxplots to count the key variables that affect safety, efficiency, and compliance. Record counterfactual instance scenarios and key variables. Determine the ranking result V of key variables that affect the safety, efficiency, and compliance of the vehicle under the same scenario. 341) Define questionnaire; The evaluation scores were obtained based on a questionnaire designed to objectively evaluate the autonomous driving decision-making system. The questionnaire was designed as follows: the lane-changing scenario identified the scene elements and characteristics that affected the ego vehicle's lane change; the emergency braking scenario identified the scene elements and characteristics that affected the ego vehicle's braking; and the intersection scenario with traffic lights identified the scene elements and characteristics that affected the ego vehicle's driving behavior. 342) The evaluation results of the questionnaire in step 41) are statistically analyzed, and the ranking results W of the scene elements and key feature variables are obtained according to the evaluation scores. The degree of influence of different scene elements on vehicle driving behavior is intuitively displayed through graphical visualization methods.

6. The explainable autonomous driving decision-making method based on causal knowledge according to claim 1, characterized in that: The specific method of step 4 is as follows: 41) Testing of autonomous driving decision-making systems; 42) Setting evaluation indicators for autonomous driving decision-making systems; 43) Analyze the performance of the autonomous driving decision system.

7. The explainable autonomous driving decision-making method based on causal knowledge according to claim 6, characterized in that: The specific method of step 41) is as follows: 411) perform model-in-the-loop testing; By inputting the self-driving car, surrounding vehicles, pedestrians, and traffic light signals into the trained model through the interface in the simulation software, the model calculates and outputs control instructions for the steering wheel angle, accelerator pedal opening, and brake pedal opening signals, and sends the control signals to the autonomous driving vehicle through the virtual interface to achieve the state update of the autonomous driving vehicle; the starting and end points are set in the simulation software, and traffic participants with different densities and different flow rates are randomly generated. The sections pass through different curvatures, including intersections with traffic lights, intersections without traffic lights, and roundabouts. The autonomous vehicle drives from the starting point to the end point, and the collision simulation ends. If there is no collision and the vehicle follows traffic rules, it is recorded as a successful test, and the number of collisions is recorded; 412) Conducting real vehicle online testing; The algorithm is tested online on the road sections where data was collected. The autonomous driving decision-making system is tested for long-distance driving, lane keeping, lane changing, and overtaking in highway environments, as well as its ability to navigate congested sections, multi-lane intersections, and intersections with or without traffic lights in urban road environments. Real-vehicle tests are conducted under varying traffic densities and flow rates, with set starting and end points. If the autonomous driving decision-making system is unable to cope with the current driving scenario, the driver is allowed to take over control of the vehicle at any time. When the driver takes over, it indicates that the autonomous driving decision-making system has failed in the current scenario, and the number of times the driver takes over is recorded. The specific method of step 42) is as follows: 421) Conduct driving safety evaluation; In the model-in-the-loop test, the collision rate is used as an indicator to evaluate the driving safety of the autonomous vehicle. In the real vehicle online test, the takeover rate is used as an indicator to evaluate the driving safety. The calculation formulas (39) and (40) of the collision rate and takeover rate are as follows: Where, T tot is the total number of trials; T c is the number of collisions in the virtual simulation test; T t The number of times the driver took over during the actual vehicle test; 422) Conduct driving efficiency evaluation; The performance of the decision-making system is evaluated by the driving efficiency of the autonomous vehicle. The driving efficiency calculation formula (41) is as follows: Where, T e The number of tests to meet the minimum travel time; T suc is the number of successes; 423) Conduct driving compliance evaluation; The performance of the decision-making system is evaluated by the driving compliance of the autonomous vehicle. The driving compliance calculation formula (42) is as follows: Where, T l The number of tests required to meet traffic regulations; The specific method of step 43) is as follows: 431) obtaining model-in-the-loop test results; Evaluate the decision-making performance of the autonomous vehicle by analyzing the results of driving safety, driving efficiency, and driving compliance evaluations, calculate the overall collision rate, evaluate the system's overall collision situation in all test scenarios, and draw a histogram or pie chart to show the distribution of the overall collision rate; calculate the collision rate for each specific scenario separately, evaluate the system's collision situation in different scenarios, draw a bar chart or line chart to show the difference in collision rate in each scenario, and record the collision accident's vehicle, adjacent vehicles, pedestrians, and environmental information; 432) Obtaining real vehicle online test results; By drawing histograms or pie charts, the driving safety, driving efficiency, and driving compliance evaluation results of autonomous vehicles in various scenarios are statistically analyzed. During the driving process of the autonomous vehicle, the information of the autonomous vehicle, adjacent vehicles, pedestrians, and the environment when the driver takes over is recorded. 433) Optimize result feedback; Analyze and record the model's in-the-loop test results and real-vehicle online test results; study the model's decision-making mechanism in failed use cases, locate the cause of decision failure, and improve the algorithm's performance; analyze the cause of each collision, draw a grouped bar chart, and display the number of collisions caused by different reasons.

Images