Machine learning based traffic exposure correction and risk attribution method for autonomous takeover scenarios

Abstract

This invention discloses a machine learning-based method for traffic exposure correction and risk attribution in autonomous driving takeover scenarios, relating to the field of autonomous driving safety technology. The invention extracts road network data from a target area, divides it into spatiotemporal observation units, matches takeover events, calculates the traffic exposure of each unit, generates a fixed exposure bias term, injects it with a fixed coefficient into the logarithmic probability space of a probability model, constructs environmental feature data, and trains independent probability prediction models for different takeover types. An interpretable attribution method is used to calculate the contribution of environmental features and generate operational-level output results. This effectively eliminates attribution bias caused by exposure confounding, improves the accuracy of takeover risk attribution, and provides reliable support for autonomous driving operation design domain optimization and safe operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving safety technology, specifically to a machine learning-based method for traffic exposure correction and risk attribution in autonomous driving takeover scenarios. Background Technology

[0002] With the development of autonomous driving technology and the advancement of large-scale road testing, human takeover events of autonomous vehicles have become core data sources reflecting the safety operation boundaries of the system and identifying weaknesses in the operational design domain. Conducting spatial-dimensional risk attribution analysis on takeover events to locate high-risk triggering scenarios in the road environment is a key technical aspect in the deployment optimization, operational parameter calibration, and safe operation and maintenance management of autonomous driving systems, and is widely used in the operational iteration of autonomous driving on urban roads.

[0003] Existing technologies typically conduct hotspot identification and environmental correlation analysis based on the spatial distribution of takeover events, directly using the frequency or count of takeover events within a road segment as the core indicator for risk assessment. This approach fails to correct for differences in autonomous driving traffic exposure across different road segments, resulting in exposure confounding bias in spatial attribution results. It cannot accurately distinguish whether the high frequency of takeover events in a road segment is due to a large volume of autonomous vehicle traffic or a high level of risk in the road environment itself, making it difficult to support precise operational design domain delineation and road scenario risk management. Summary of the Invention

[0004] The purpose of this invention is to provide a traffic exposure correction and risk attribution method for autonomous driving takeover scenarios based on machine learning, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a traffic exposure correction and risk attribution method for autonomous driving takeover scenarios based on machine learning, comprising the following steps:

[0006] Step 100: Extract the road network of the target area and divide it into several spatial units. Divide each spatial unit into multiple spatiotemporal observation units according to a preset time window. Match each takeover event in the autonomous driving takeover event data to the corresponding spatiotemporal observation unit.

[0007] Step 200: Calculate the traffic exposure of each of the spatiotemporal observation units, perform a logarithmic transformation on the traffic exposure to obtain a fixed exposure bias term; inject the fixed exposure bias term as a fixed offset term into the log-probability space of the probability model, the coefficient of the fixed exposure bias term is always 1.0 and remains unchanged during model training;

[0008] Step 300: Construct environmental feature data for each spatiotemporal observation unit, and construct corresponding event response labels for at least two different types of takeovers; combine the fixed exposure bias term to train independent probability prediction models for each type of takeover, and obtain the residual contribution function of environmental features for each type of takeover;

[0009] Step 400: Based on the various probability prediction models that have been trained, calculate the contribution of each environmental feature to the corresponding type of takeover event using an interpretable attribution method, and generate operation and maintenance level output results for the deployment of autonomous driving systems.

[0010] Preferably, step 100 includes:

[0011] Step 110: Extract road network data of the target area, extract road centerlines and road attribute information, and divide each road into several road segment units according to the preset segment length;

[0012] Step 120: Divide each road segment unit into time segments according to a preset time window to obtain multiple road segment-time period observation units as the spatiotemporal observation units;

[0013] Step 130: Match each takeover event to the corresponding road segment-time period observation unit and remove takeover events whose matching distance exceeds the preset maximum distance.

[0014] Preferably, step 200 includes:

[0015] Step 210: Determine the calculation dimension of traffic exposure, wherein the traffic exposure is any one of the following: the number of autonomous vehicle trips within the corresponding spatiotemporal observation unit, the total mileage of autonomous vehicles, and the estimated value of the overall traffic flow of the road segment;

[0016] Step 220: Perform a smoothed logarithmic transformation on the traffic exposure data to obtain a fixed exposure bias term, calculated using the following formula:

[0017] ;

[0018] In the formula, This is the fixed exposure bias term for the i-th spatiotemporal observation unit. Let be the traffic exposure of the i-th spatiotemporal observation unit;

[0019] Step 230: The fixed exposure bias term is injected into the log-odds space of the probability model as a fixed offset term. The coefficient of the fixed exposure bias term is constant at 1.0 throughout the entire training process and does not participate in the model weight update.

[0020] Preferably, step 300 includes:

[0021] Step 310: Construct an exogenous environmental feature vector for each spatiotemporal observation unit as the environmental feature data. The exogenous environmental feature vector includes at least one of the following groups: road geometry feature group, intersection and topology feature group, interaction context feature group, traffic state feature group, and environmental condition feature group. All features are exogenous variables that do not depend on the takeover event.

[0022] Step 320: For each type of target takeover, construct a corresponding binary response variable as the event response label based on whether the spatiotemporal observation unit has experienced a corresponding type of takeover event.

[0023] Step 330: Group all samples according to road segment identification, and divide them into training set and test set. All time period samples of the same road segment are completely included in the training set or test set; use class weight balancing strategy to deal with the problem of sparse positive samples.

[0024] Step 340: In conjunction with the fixed exposure bias term, train independent binary classification probability models for each type of takeover on the training set as the probability prediction model, and obtain the residual contribution function of environmental features corresponding to each type of takeover.

[0025] Preferably, in step 400, calculating the contribution of each environmental characteristic to the corresponding type of takeover event using an interpretable attribution method includes:

[0026] An interpretable attribution method based on Shapley values ​​is used to calculate the contribution of each environmental feature to each sample; the average absolute contribution value is used as the global importance index of the corresponding environmental feature; the contribution values ​​are aggregated by feature group to obtain the contribution ratio of each group of features to the corresponding type of takeover event; the interaction contribution value between selected feature pairs is calculated to identify composite triggering conditions.

[0027] Preferably, the operation and maintenance level output results include a map of weak points in the operation design domain, a road segment-level risk threshold table, and a list of high-risk road segments;

[0028] Run the map of weak points in the design domain where the probability of takeover exceeds a preset threshold for road sections;

[0029] The road segment-level risk threshold table includes the decomposition results of risk contribution factors for each road segment and corresponding operational parameter adjustment suggestions;

[0030] The list of high-risk road sections is arranged in descending order of the predicted probability of being taken over.

[0031] Preferably, when the binary classification probability model adopts a gradient boosting tree model, the fixed exposure bias term is injected into the model as a baseline marginal parameter;

[0032] When the binary classification probability model adopts a generalized additive model, a logistic regression model, or a neural network model, the fixed exposure bias term is added as an offset term to the linear predictor with a fixed coefficient of 1.0.

[0033] Preferably, step 100 further includes:

[0034] We construct intersection entrance lane-time period units and grid-time period units in parallel to form a multi-scale observation system; the attribution results of observation units at different scales are used for cross-validation.

[0035] Preferably, step 400 further includes a relative risk quantification step:

[0036] The zero-environmental baseline takeover probability is constructed by using the global intercept of the probability prediction model and the fixed exposure bias term; the predicted takeover probability of each spatiotemporal observation unit is compared with the corresponding zero-environmental baseline takeover probability to obtain the pure environmental relative risk multiple after stripping the system baseline; the relative risk difference of the same road segment under different takeover types is calculated to identify the risk specificity of the corresponding road segment.

[0037] Preferably, a standardized list of high-risk road sections is generated based on the relative risk multiple of the pure environment. The standardized list of high-risk road sections is accompanied by risk-specific labels for the corresponding road sections, which are used for horizontal benchmarking and scenario-based parameter calibration across vehicle fleets and system versions in the operation design domain.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] By extracting road networks to construct spatiotemporal observation units and matching takeover events, calculating and correcting traffic exposure, and injecting fixed bias term coefficients into the probability model, multiple environmental features are constructed and independent probability models are trained for different takeover types. This approach, relying on interpretable attribution methods to calculate feature contribution and generate operation-level output results, effectively eliminates attribution bias caused by mixed traffic exposure, avoids signal distortion caused by mixed attributions of different types of takeover events, accurately identifies environmental impact factors corresponding to various takeovers, achieves standardized quantification and differentiated judgment of road environmental risks, improves the accuracy and reliability of autonomous driving takeover risk attribution, and provides reliable technical support for autonomous driving operation design domain optimization and scenario-based parameter calibration. Attached Figure Description

[0040] Figure 1 The main flowchart of a traffic exposure correction and risk attribution method for autonomous driving takeover scenarios based on machine learning provided in an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] The overall architecture of this invention consists of four stages: construction of road segment-time period observation units, calculation and correction of traffic exposure, typified probabilistic modeling, interpretable attribution, and closed-loop output. It can be applied to the identification of weak points in the Operational Design Domain (ODD) of L3-L4 level autonomous vehicles in real-world urban road operation scenarios, as well as road segment risk assessment and operational parameter optimization.

[0043] The input to this method is a set of target takeover events. Each takeover event in the set contains at least the time of occurrence, the location of the event (latitude and longitude coordinates), and the event type label. The set of target takeover events can be derived from the output of a takeover classifier based on kinematic rules, the output of a takeover classifier based on machine learning, manually labeled takeover event records, and mode switching events in vehicle takeover logs.

[0044] Please see Figure 1 This invention provides a machine learning-based method for traffic exposure correction and risk attribution in autonomous driving takeover scenarios, the method being applied to, including:

[0045] Step 100: Extract the road network of the target area and divide it into several spatial units. Divide each spatial unit into multiple spatiotemporal observation units according to a preset time window. Match each takeover event in the autonomous driving takeover event data to the corresponding spatiotemporal observation unit.

[0046] Specifically, step 100 includes:

[0047] Step 110: Extract road network data of the target area, extract road centerlines and road attribute information, and divide each road into several road segment units according to the preset segment length;

[0048] Step 120: Divide each road segment unit into time segments according to a preset time window to obtain multiple road segment-time period observation units as the spatiotemporal observation units;

[0049] Step 130: Match each takeover event to the corresponding road segment-time period observation unit and remove takeover events whose matching distance exceeds the preset maximum distance.

[0050] The road network data extracted in step 110 can be obtained from OpenStreetMap (OSM) or other open source / commercial road network data sources. The extracted road attribute information includes road topology attributes, road level, number of lanes, and direction of travel. The preset segment length of the road is in the range of 50m to 200m, with a preferred segment length of 100m.

[0051] In this embodiment, the residual segments generated after road segmentation are merged. If the length of the segment after segmentation is less than the minimum residual segment threshold, the residual segment is merged into the adjacent previous segment unit. The minimum residual segment threshold ranges from 10m to 50m, preferably 20m, to avoid statistical errors caused by extremely short segment units and to ensure the rationality of spatial unit division.

[0052] Understandably, the preset time window range for dividing the road segment units in step 120 is 15 min to 6 h, with a preferred time window of 1 hour. At the same time, the time window can be further aggregated into macro-level time periods such as peak hours, off-peak hours, and nighttime hours according to the analysis needs, thereby refining the analysis granularity of the time dimension.

[0053] Furthermore, the spatial matching process in step 130 is completed based on equidistant projection, preferably using the UTM universal transverse Mercator projection; sampling is performed at fixed intervals along the centerline of each road segment and a cKDTree spatial index is established; the nearest neighbor matching rule is used to associate the takeover event with the road segment-time period observation unit; the maximum threshold value of the matching distance is in the range of 20m to 50m, preferably 30m; takeover events exceeding this threshold are judged as off-line segment data and are removed to ensure the accuracy of spatial matching.

[0054] Furthermore, step 100 also includes:

[0055] We construct intersection entrance lane-time period units and grid-time period units in parallel to form a multi-scale observation system; the attribution results of observation units at different scales are used for cross-validation.

[0056] In this step, in addition to the basic road segment-time period observation unit, a multi-scale spatial observation unit is constructed to form a complete multi-scale observation system, and analysis is carried out from different spatial granularities.

[0057] In this embodiment, the intersection approach lane-time period unit takes the intersection stop line as the starting point and defines a range of 50m to 200m upstream as the spatial boundary, which is specifically designed for detailed analysis of intersections, a high-incidence scenario for takeover incidents; the grid-time period unit uses a fixed-size grid to divide the space, with the preferred grid size being 500m×500m, which is suitable for overall risk assessment of large areas.

[0058] Understandably, the simultaneous modeling and attribution analysis of three different scales of observation systems—road segment units, intersection approach lane units, and grid units—and the cross-validation of multiple sets of attribution results can effectively reduce the analytical bias caused by a single spatial granularity and improve the reliability of the overall attribution conclusions.

[0059] Step 200: Calculate the traffic exposure of each of the spatiotemporal observation units, perform a logarithmic transformation on the traffic exposure to obtain a fixed exposure bias term; inject the fixed exposure bias term as a fixed offset term into the log-probability space of the probability model, the coefficient of the fixed exposure bias term is always 1.0 and remains unchanged during model training.

[0060] Specifically, step 200 includes:

[0061] Step 210: Determine the calculation dimension of traffic exposure, wherein the traffic exposure is any one of the following: the number of autonomous vehicle trips within the corresponding spatiotemporal observation unit, the total mileage of autonomous vehicles, and the estimated value of the overall traffic flow of the road segment;

[0062] Step 220: Perform a smoothed logarithmic transformation on the traffic exposure data to obtain a fixed exposure bias term, calculated using the following formula:

[0063] ;

[0064] In the formula, This is the fixed exposure bias term for the i-th spatiotemporal observation unit. Let be the traffic exposure of the i-th spatiotemporal observation unit;

[0065] Step 230: The fixed exposure bias term is injected into the log-odds space of the probability model as a fixed offset term. The coefficient of the fixed exposure bias term is constant at 1.0 throughout the entire training process and does not participate in the model weight update.

[0066] Step 210 provides three optional traffic exposure levels. The definition method can be selected as needed based on the actual available data sources: the preferred method is... That is, the number of trips of different autonomous vehicles within the corresponding spatiotemporal observation unit, counted by deduplication using vehicle ID + trip ID, is used to characterize the autonomous driving traffic load borne by that unit; alternative method A is... , which corresponds to the total mileage of all autonomous vehicles within the spatiotemporal observation unit, is used to characterize the total amount of autonomous driving traffic work; Alternative method B is the overall traffic flow estimate of the road segment, which is suitable for scenarios without fine-grained autonomous driving trajectory data. This traffic flow data can be obtained through loops, floating cars, or traffic assignment models.

[0067] In this embodiment, step 220 adopts The logarithmic transformation is performed in the formal form, and a smoothing term is added. The purpose of adding the value "1" is to avoid the smoothing effect when... To address the issue of meaningless logarithmic operations at zero exposure levels, we ensure that all samples can be calculated correctly.

[0068] Understandably, this invention constructs a complete computational relationship in the logit space of the probability model, and the overall model formula is as follows:

[0069] ;

[0070] In the formula, Let logit be the value of the i-th spatiotemporal observation unit. Let be the probability of a target-type takeover event occurring in the i-th spatiotemporal observation unit. For the model's global intercept, This is the residual contribution function learned by the model from environmental features.

[0071] Furthermore, the coefficient of the fixed exposure bias term remains constant at 1.0 throughout the process and does not participate in model iteration updates. It is not treated as a regular feature for the model to learn weights autonomously. This design forces the model to reassess the additional takeover probability information provided by environmental features, given the known autonomous driving traffic volume within the spatiotemporal observation unit. This thoroughly distinguishes between two different influencing factors: "large scale of autonomous driving traffic" and "high complexity of road environment." If the environmental features have no actual correlation with the takeover event, the residual contribution function of the environmental features approaches 0, and the model prediction result is determined solely by traffic exposure, which aligns with real traffic operation patterns and eliminates attribution bias caused by exposure confounding at its source.

[0072] Step 300: Construct environmental feature data for each spatiotemporal observation unit, and construct corresponding event response labels for at least two different types of takeovers; combine the fixed exposure bias term to train independent probability prediction models for each type of takeover, and obtain the residual contribution function of environmental features for each type of takeover.

[0073] Specifically, step 300 includes:

[0074] Step 310: Construct an exogenous environmental feature vector for each spatiotemporal observation unit as the environmental feature data. The exogenous environmental feature vector includes at least one of the following groups: road geometry feature group, intersection and topology feature group, interaction context feature group, traffic state feature group, and environmental condition feature group. All features are exogenous variables that do not depend on the takeover event.

[0075] Step 320: For each type of target takeover, construct a corresponding binary response variable as the event response label based on whether the spatiotemporal observation unit has experienced a corresponding type of takeover event.

[0076] Step 330: Group all samples according to road segment identification, and divide them into training set and test set. All time period samples of the same road segment are completely included in the training set or test set; use class weight balancing strategy to deal with the problem of sparse positive samples.

[0077] Step 340: In conjunction with the fixed exposure bias term, train independent binary classification probability models for each type of takeover on the training set as the probability prediction model, and obtain the residual contribution function of environmental features corresponding to each type of takeover.

[0078] In step 310, all five feature groups are exogenous variables. The feature value calculation process does not depend on the takeover event itself, which can effectively avoid post-event information leakage and ensure the authenticity and generalization ability of model training. The specific indicators included in each feature group are as follows: The road geometry feature group includes at least one of the following: maximum horizontal curvature, curvature, average slope, and slope change rate, where curvature is the ratio of the road segment path length to the straight distance between the start and end points; the intersection and topology feature group includes at least one of the following: distance to the nearest intersection, number of approach lanes at the intersection where the road segment is located, distance to the nearest upstream traffic light, number of lanes, and road grade code; the interaction context feature group includes at least one of the following: distance to the nearest bus stop, distance to the nearest pedestrian crossing, and number of various POIs (points of interest) within a specified radius of the road segment; the traffic status feature group includes at least one of the following: time period attribute, congestion index, and speed limit value of the road segment; the environmental condition feature group includes at least one of the following: hourly precipitation, visibility level, and lighting conditions.

[0079] In this embodiment, step 320 constructs separate binary response variables for different pipe connection types. The variable value rule is as follows: if the k-th type of takeover event occurs in the i-th spatiotemporal observation unit, it is marked as 1; otherwise, it is marked as 0. Common takeover types include braking takeover and acceleration takeover. The environmental triggering logic of the two types of takeover is significantly different. Constructing separate labels can avoid the mutual cancellation of signals caused by mixed attribution and improve the reliability of attribution results.

[0080] Understandably, step 330 uses a group sampling method to divide the training set and the test set. Specifically, it uses GroupShuffleSplit or Group K-Fold cross-validation to divide all time period samples corresponding to the same road segment into the training set or the test set. This avoids the sample leakage problem caused by the correlation of data from multiple time periods in the same road segment and improves the generalization ability of the model. Autonomous driving takeover events are sparse events, and the proportion of positive samples in the road segment-time period observation unit is usually less than 0.2%. Therefore, a class weight balancing strategy is used to optimize the model training effect. The preferred setting of positive class weights is the ratio of the number of negative samples to the number of positive samples.

[0081] Furthermore, in step 340, an independent binary classification probability model is trained for each type of takeover event. The models for different takeover types are independent of each other and do not interfere with each other. Finally, the residual contribution function of environmental features corresponding to each type of takeover is output, so as to realize the differentiated environmental factor identification for different takeover types.

[0082] Optionally, the feature group combination method can be flexibly selected according to the analysis needs. Only a single type of feature group or multiple types of feature groups can be used to adapt to attribution scenarios with different accuracy requirements.

[0083] Furthermore, when the binary classification probability model adopts a gradient boosting tree model, the fixed exposure bias term is injected into the model as a baseline marginal parameter.

[0084] When the binary classification probability model adopts a generalized additive model, a logistic regression model, or a neural network model, the fixed exposure bias term is added as an offset term to the linear predictor with a fixed coefficient of 1.0.

[0085] The probability model of this invention can be selected from a variety of mainstream machine learning models, and the main logic of the framework does not change with the model selection.

[0086] In this embodiment, the gradient boosting tree model includes commonly used types such as XGBoost, LightGBM, and CatBoost. This type of model injects a fixed exposure bias term into the model through the built-in base_margin parameter, keeping the bias term coefficient at 1.0 throughout the process.

[0087] Understandably, when using generalized additive models, logistic regression, neural networks, etc., a fixed exposure bias term is added as an offset to the linear predictor of the model, with the coefficient strictly limited to 1.0, to ensure the consistency of the exposure correction mechanism under different models.

[0088] Furthermore, the model selection can be flexibly switched according to the data scale, computational efficiency, and accuracy requirements, expanding the applicable scenarios of this method.

[0089] Step 400: Based on the various probability prediction models that have been trained, calculate the contribution of each environmental feature to the corresponding type of takeover event using an interpretable attribution method, and generate operation and maintenance level output results for the deployment of autonomous driving systems.

[0090] In step 400, the contribution of each environmental feature to the corresponding type of takeover event is calculated using an interpretable attribution method, including:

[0091] An interpretable attribution method based on Shapley values ​​is used to calculate the contribution of each environmental feature to each sample; the average absolute contribution value is used as the global importance index of the corresponding environmental feature; the contribution values ​​are aggregated by feature group to obtain the contribution ratio of each group of features to the corresponding type of takeover event; the interaction contribution value between selected feature pairs is calculated to identify composite triggering conditions.

[0092] In this step, the Shapley value (SHAP)-based model interpretability method is used to calculate the feature contribution. This method can quantify the influence of each environmental feature on the prediction result of a single sample. TreeExplainer can be used for tree-type models, and KernelExplainer can be used for machine learning models with arbitrary structures.

[0093] In this embodiment, the contribution value of a single feature of a single sample is denoted as... The formula for calculating the global importance index of feature j for the k-th type of takeover event is: By eliminating the mutual cancellation of positive and negative contributions through the average absolute contribution value, the overall influence intensity of the feature is objectively reflected.

[0094] Understandably, by aggregating and statistically analyzing the contribution values ​​of individual features according to five major feature groups—road geometry, intersection topology, interaction context, traffic state, and environmental conditions—we can obtain the overall contribution ratio of each feature group to the target takeover event, thus intuitively forming the distribution characteristics of environmental impact factors for different takeover types.

[0095] Furthermore, this step also supports feature interaction analysis, which calculates the SHAP interaction value between two sets of features based on the SHAP interaction attribution method, and identifies composite triggering conditions, i.e., scenarios in which the attribution effect of one environmental factor will be significantly enhanced or weakened as the value of another factor changes; the interaction analysis results can be output in the form of SHAP dependency graph and interaction matrix to explore the takeover triggering rules in complex scenarios.

[0096] Furthermore, the operation and maintenance level output results include a map of weak points in the operation design domain, a road segment-level risk threshold table, and a list of high-risk road segments;

[0097] Run the map of weak points in the design domain where the probability of takeover exceeds a preset threshold for road sections;

[0098] The road segment-level risk threshold table includes the decomposition results of risk contribution factors for each road segment and corresponding operational parameter adjustment suggestions;

[0099] The list of high-risk road sections is arranged in descending order of the predicted probability of being taken over.

[0100] Among them, the three types of outputs in this step are operation and maintenance level outputs for the deployment of autonomous driving systems, which can directly support the operation design and safety supervision of autonomous driving systems.

[0101] In this embodiment, the Operation Design Domain (ODD) vulnerability map is presented in a spatial visualization form, and road segments with takeover prediction probabilities exceeding a preset threshold are uniformly marked. It can be directly applied to the review of autonomous driving operation design domain and vehicle driving route planning.

[0102] Understandably, the road segment-level risk threshold table decomposes the contribution values ​​of all environmental features of a single road segment, identifies the core factors of the dominant takeover risk within that road segment, and generates targeted suggestions for adjusting autonomous driving operation parameters based on the characteristics of the risk factors, providing a direct basis for the calibration of road segment-level operation parameters.

[0103] Furthermore, the list of high-risk road sections is sorted from high to low according to the predicted probability of takeover, which makes it easier for the operation and safety management team to carry out road section-level ODD assessments, vehicle operation parameter calibrations and other tasks in order of priority, thereby improving the efficiency of operation and maintenance.

[0104] Furthermore, step 400 also includes a relative risk quantification step:

[0105] The zero-environmental baseline takeover probability is constructed by using the global intercept of the probability prediction model and the fixed exposure bias term; the predicted takeover probability of each spatiotemporal observation unit is compared with the corresponding zero-environmental baseline takeover probability to obtain the pure environmental relative risk multiple after stripping the system baseline; the relative risk difference of the same road segment under different takeover types is calculated to identify the risk specificity of the corresponding road segment.

[0106] This step relies on the inherent parameters of the model to standardize the processing of risk indicators, which solves the problem that traditional absolute probability indicators are affected by the version of the autonomous driving system and the overall performance of the fleet, and cannot be compared horizontally.

[0107] In this embodiment, the zero environmental baseline takeover probability refers to the takeover probability determined solely by the model's global intercept and the fixed exposure bias term when all environmental features are in the baseline state, representing the basic operational level of the autonomous driving system itself.

[0108] Understandably, the pure environmental relative risk multiple is obtained by calculating the ratio of the predicted takeover probability to the baseline takeover probability. This indicator completely eliminates the interference of the autonomous driving system's own baseline and only reflects the level of risk brought about by the road environment itself.

[0109] Furthermore, by performing difference and ratio calculations on the relative risk multiples of different takeover types under the same road segment, the relative risk difference degree is obtained. Based on this, "generally high-risk road segments with all types of takeovers" and "specifically high-risk road segments with a single type of takeover" are distinguished, so as to achieve accurate identification of risk types.

[0110] Furthermore, based on the calculated relative risk multiple of the pure environment, a corresponding standardized list of high-risk road sections is generated; the standardized list of high-risk road sections is accompanied by risk-specific labels for the corresponding road sections, which are used for horizontal benchmarking and scenario-based parameter calibration across vehicle fleets and system versions of the operation design domain.

[0111] By generating new standardized results based on the relative risk multiple of the pure environment, the analysis results can be reused across different scenarios.

[0112] In this embodiment, the standardized high-risk road segment list is sorted based on the relative environmental risk multiple. Each road segment is simultaneously marked with a risk-specific label to clarify whether the road segment is mainly characterized by braking takeover risk, acceleration takeover risk, or high risk of all types of takeover.

[0113] Understandably, this standardized list breaks free from the limitations of a single fleet or a single system version, enabling horizontal benchmarking of operational design domains between different operating fleets and different iterations of autonomous driving systems, and unifying risk assessment standards for different entities.

[0114] Furthermore, by combining risk-specific tags, maintenance personnel can perform differentiated scenario-based parameter calibration for road sections with different risk types, ensuring driving safety while also taking into account the operating efficiency of autonomous vehicles.

[0115] In this embodiment, the present invention, relying on a holistic design that combines fixed bias correction of traffic exposure, independent modeling by type, interpretable attribution, and closed-loop operation and maintenance output, solves the problems of exposure mixing, distorted attribution of heterogeneous events, and inapplicability of analysis results in existing autonomous driving takeover spatial attribution technologies. Those skilled in the art can make conventional parameter adjustments and adaptations within the technical framework disclosed in this invention, based on actual operational scenarios, all of which fall within the protection scope of this invention.

[0116] Furthermore, it should be noted that the combination of the various technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0117] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is obviously not limited to the above embodiments, with many similar variations. All modifications that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should fall within the protection scope of this invention.

[0118] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A machine learning-based method for traffic exposure correction and risk attribution in autonomous driving takeover scenarios, characterized in that, Includes the following steps: Step 100: Extract the road network of the target area and divide it into several spatial units. Then, divide each spatial unit into multiple spatiotemporal observation units according to a preset time window. Match each takeover event in the autonomous driving takeover event data to its corresponding spatiotemporal observation unit; Step 200: Calculate the traffic exposure of each of the spatiotemporal observation units, and perform a logarithmic transformation on the traffic exposure to obtain a fixed exposure bias term; The fixed exposure bias term is injected as a fixed bias term into the log-odds space of the probability model. The coefficient of the fixed exposure bias term is always 1.0 and remains unchanged during model training. Step 300: Construct environmental feature data for each spatiotemporal observation unit, and construct corresponding event response labels for at least two different types of takeovers; combine the fixed exposure bias term to train independent probability prediction models for each type of takeover, and obtain the residual contribution function of environmental features for each type of takeover; Step 400: Based on the various probability prediction models that have been trained, calculate the contribution of each environmental feature to the corresponding type of takeover event using an interpretable attribution method, and generate operation and maintenance level output results for the deployment of autonomous driving systems.

2. The method for traffic exposure correction and risk attribution in autonomous driving takeover scenarios based on machine learning as described in claim 1, characterized in that, Step 100 includes: Step 110: Extract road network data of the target area, extract road centerlines and road attribute information, and divide each road into several road segment units according to the preset segment length; Step 120: Divide each road segment unit into time segments according to a preset time window to obtain multiple road segment-time period observation units as the spatiotemporal observation units; Step 130: Match each takeover event to the corresponding road segment-time period observation unit and remove takeover events whose matching distance exceeds the preset maximum distance.

3. The traffic exposure correction and risk attribution method for autonomous driving takeover scenarios based on machine learning as described in claim 1, characterized in that, Step 200 includes: Step 210: Determine the calculation dimension of traffic exposure, wherein the traffic exposure is any one of the following: the number of autonomous vehicle trips within the corresponding spatiotemporal observation unit, the total mileage of autonomous vehicles, and the estimated value of the overall traffic flow of the road segment; Step 220: Perform a smoothed logarithmic transformation on the traffic exposure data to obtain a fixed exposure bias term, calculated using the following formula: ； In the formula, This is the fixed exposure bias term for the i-th spatiotemporal observation unit. Let be the traffic exposure of the i-th spatiotemporal observation unit; Step 230: The fixed exposure bias term is injected into the log-odds space of the probability model as a fixed offset term. The coefficient of the fixed exposure bias term is constant at 1.0 throughout the entire training process and does not participate in the model weight update.

4. The method for traffic exposure correction and risk attribution in autonomous driving takeover scenarios based on machine learning as described in claim 1, characterized in that, Step 300 includes: Step 310: Construct an exogenous environmental feature vector for each spatiotemporal observation unit as the environmental feature data. The exogenous environmental feature vector includes at least one of the following groups: road geometry feature group, intersection and topology feature group, interaction context feature group, traffic state feature group, and environmental condition feature group. All features are exogenous variables that do not depend on the takeover event. Step 320: For each type of target takeover, construct a corresponding binary response variable as the event response label based on whether the spatiotemporal observation unit has experienced a corresponding type of takeover event. Step 330: Group all samples according to road segment identification, and divide them into training set and test set. All time period samples of the same road segment are completely included in the training set or test set; use class weight balancing strategy to deal with the problem of sparse positive samples. Step 340: In conjunction with the fixed exposure bias term, train independent binary classification probability models for each type of takeover on the training set as the probability prediction model, and obtain the residual contribution function of environmental features corresponding to each type of takeover.

5. The method for traffic exposure correction and risk attribution in autonomous driving takeover scenarios based on machine learning as described in claim 1, characterized in that, In step 400, the contribution of each environmental feature to the corresponding type of takeover event is calculated using an interpretable attribution method, including: An interpretable attribution method based on Shapley values ​​is used to calculate the contribution of each environmental feature to each sample; the average absolute contribution value is used as the global importance index of the corresponding environmental feature; the contribution values ​​are aggregated by feature group to obtain the contribution ratio of each group of features to the corresponding type of takeover event; the interaction contribution value between selected feature pairs is calculated to identify composite triggering conditions.

6. The traffic exposure correction and risk attribution method for autonomous driving takeover scenarios based on machine learning as described in claim 5, characterized in that, The operation and maintenance level output results include a map of weak points in the operation design domain, a road segment-level risk threshold table, and a list of high-risk road segments; Run the map of weak points in the design domain where the probability of takeover exceeds a preset threshold for road sections; The road segment-level risk threshold table includes the decomposition results of risk contribution factors for each road segment and corresponding operational parameter adjustment suggestions; The list of high-risk road sections is arranged in descending order of the predicted probability of being taken over.

7. The method for traffic exposure correction and risk attribution in autonomous driving takeover scenarios based on machine learning as described in claim 4, characterized in that, When the binary classification probability model adopts the gradient boosting tree model, the fixed exposure bias term is injected into the model as a baseline marginal parameter. When the binary classification probability model adopts a generalized additive model, a logistic regression model, or a neural network model, the fixed exposure bias term is added as an offset term to the linear predictor with a fixed coefficient of 1.

0.

8. The method for traffic exposure correction and risk attribution in autonomous driving takeover scenarios based on machine learning as described in claim 2, characterized in that, Step 100 further includes: We construct intersection entrance lane-time period units and grid-time period units in parallel to form a multi-scale observation system; the attribution results of observation units at different scales are used for cross-validation.

9. The traffic exposure correction and risk attribution method for autonomous driving takeover scenarios based on machine learning as described in claim 5, characterized in that, Step 400 also includes a relative risk quantification step: The zero-environmental baseline takeover probability is constructed by using the global intercept of the probability prediction model and the fixed exposure bias term; the predicted takeover probability of each spatiotemporal observation unit is compared with the corresponding zero-environmental baseline takeover probability to obtain the pure environmental relative risk multiple after stripping the system baseline; the relative risk difference of the same road segment under different takeover types is calculated to identify the risk specificity of the corresponding road segment.

10. The method for traffic exposure correction and risk attribution in autonomous driving takeover scenarios based on machine learning as described in claim 9, characterized in that, Based on the relative risk multiple of the pure environment, a standardized list of high-risk road sections is generated. The standardized list of high-risk road sections is accompanied by risk-specific labels for the corresponding road sections, which are used for horizontal benchmarking and scenario-based parameter calibration across vehicle fleets and system versions in the operation design domain.

Images