Auxiliary driving accident tracing method fusing feature screening and dynamic Bayesian network

Abstract

The invention relates to an auxiliary driving accident tracing method fusing feature screening and a dynamic Bayesian network, and the method comprises the following steps: S1, constructing a mixed training data set containing a positive sample of accident data and a negative sample of normal driving data, and quantitatively screening out a key data field having correlation with the accident from massive signals; s2, constructing a generative directed edge pointing to objective performance from subjective causes, and constructing a double-layer dynamic Bayesian network skeleton; s3, constructing a special data set only containing the accident data, discretizing key data fields, inputting the discretized key data fields into a DBN empty skeleton, and outputting a prior DBN model filled with a complete conditional probability table; and S4, outputting a causal chain of'objective performance caused by subjective cause 'with the maximum probability. According to the invention, refined and automatic distinguishing of accident liabilities is realized, and the method has great technical value and social significance for improving traffic safety, perfecting an auxiliary driving system and clarifying man-machine responsibilities in the future.

Description

Technical Field

[0001] This invention relates to the fields of vehicle safety technology and artificial intelligence, and more specifically, to an assisted driving accident tracing method that integrates feature screening and dynamic Bayesian networks. Background Technology

[0002] With the widespread adoption of Advanced Driver Assistance Systems (ADAS), vehicles can record massive amounts of dynamic data before and after accidents, providing unprecedented opportunities for refined accident analysis. However, existing accident tracing methods are still largely constrained by a high degree of reliance on expert experience, which is not only subjective and inefficient, but also extremely difficult to determine liability, especially in complex "human-machine co-driving" scenarios. Existing rule bases or fault tree methods are also inadequate for handling the complex interactions in human-machine co-driving scenarios. Furthermore, the causal mechanisms are drastically different for different accident types, so the same type of accident often exhibits relatively correlated data changes. Existing technologies struggle to match different accident types with their characteristic data, thereby failing to use time-series data for accident tracing. While current mainstream deep learning or black-box models perform well in pattern recognition, they struggle to distinguish between correlation and causation. In current "human-machine co-driving" accidents, the traditional "black-box" problem of liability determination is becoming increasingly inadequate due to the deep involvement of ADAS, making traditional liability determination methods that rely on traffic police experience insufficient. Traffic police cannot determine the exact moment an accident occurred through visual observation whether it was due to driver negligence (unwillingness to take over), system failure too quickly (inability to take over), or "misled" takeover caused by system "silent failure". Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide an assisted driving accident tracing method that integrates feature screening and dynamic Bayesian network. It can deeply integrate data-driven, vehicle dynamics principles and expert knowledge for preset accident types to achieve automated, refined and probabilistic determination of accident causes.

[0004] The technical solution adopted by this invention to solve its technical problem is: to construct an assisted driving accident source tracing method that integrates feature screening and dynamic Bayesian networks, including the following steps: S1. Key feature screening based on mixed datasets: For the preset homogeneous accident types, a mixed training dataset containing positive samples of accident data and negative samples of normal driving data is constructed. The dataset is trained using machine learning algorithms. By ranking the feature importance output by the model, key data fields that are related to this type of accident are quantitatively screened from massive signals. S2. Construct a two-layer DBN topology structure of "objective observable level - subjective cause definition level". Use the key data fields selected in step S1 as objective observable nodes and establish their physical causal connections and temporal dependencies between slices based on vehicle dynamics. At the same time, introduce latent variable cause nodes and construct generative directed edges from subjective causes to objective performance to construct a two-layer dynamic Bayesian network skeleton. S3. DBN parameter learning based on accident-specific dataset: Construct a special dataset containing only accident data of this type, discretize the key data fields and input them into the DBN empty skeleton, use maximum likelihood estimation to learn the objective observable node parameters, and use the expectation-maximization EM algorithm to iteratively train the latent variable parameters in the case of missing causal labels, and output a prior DBN model that fills the complete conditional probability table. S4. Probabilistic tracing of new accidents: When a new accident occurs and is classified into this type, the corresponding key evidence data is extracted and discretized. This data is then used as evidence input into the trained DBN model. The inference algorithm is used to calculate the posterior probability distribution of each node. Based on the probability of occurrence of the latent variable nodes, the causal chain with the highest probability of "subjective cause leading to objective performance" is output, thereby achieving automated accident determination.

[0005] According to the above scheme, in step S1, data-driven methods are used to mine key observable features from massive amounts of data to distinguish between "preset accident types" and "normal driving".

[0006] According to the above scheme, the method of using data-driven methods to mine key observable features from massive amounts of data to distinguish between "preset accident types" and "normal driving" includes the following steps: S101. Construct a hybrid training dataset by extracting positive and negative samples from historical accident data. The positive samples are accident segment data labeled with "preset accident type," and the negative samples are normal driving segment data. All data are time-series aligned, and the original data fields from vehicle bus CAN, ADAS sensor, and environmental perception data are extracted as the original feature set. X raw Construct the target vector Y label As labels for supervised learning, accident samples are labeled as 1, and normal samples are labeled as 0; S102. Feature Importance Ranking: Supervised learning algorithms in machine learning are used to perform binary classification training on the mixed dataset. After training, the feature importance index output by the model is extracted, and the original features are sorted from high to low importance to obtain a feature importance ranking list.

[0007] According to the above scheme, in step S2, the structure of the dynamic Bayesian network DBN at the objective physical level is constructed to complete the construction of the two-layer network structure, specifically as follows: S201. Determine the objectively observable nodes by selecting the top-N data fields ranked in the sorting list in step S1 as the objectively observable nodes of DBN. S202. Construct the DBN objective layer structure. Within the same time slice T, establish physical causal connections between nodes and establish temporal dependencies of the same node between adjacent time slices T and T+1 to form the temporal evolution structure between DBN slices. S203. Define hidden variable nodes, which represent the causes of the accident, including driver behavior factors, vehicle system status or environmental factors. S204. Construct the subjective layer structure of DBN, and construct directed edges from the subjective layer to the objective layer. The internal structure of DBN slices is expanded into a two-layer structure of "objective layer-subjective layer", forming a complete empty skeleton of DBN. According to the above scheme, in step S3, pure accident data is used to perform parameter training and conditional probability table (CPT) learning and filling on the empty DBN skeleton generated in step S2.

[0008] The method for parameter training and CPT learning to fill the generated DBN empty skeleton according to the above scheme includes the following steps: S301. Construct an accident dataset, extract historical accident data belonging to the "preset accident type", excluding normal driving data, and retain the data fields selected in step S2; S302. Perform data discretization processing, and discretize the continuous data to transform it into state values ​​suitable for Bayesian network processing. S303. Perform parameter learning by inputting the discretized accident dataset into the DBN empty skeleton. According to the above scheme, in step S303, the physical causal relationship between objectively observable nodes is learned by using maximum likelihood estimation (MLE) to learn the conditional probability table. For structures containing latent variable nodes, the expectation-maximization (EM) algorithm is used for iterative training to automatically learn the probabilistic dependency relationship between latent variables and observable variables, thereby obtaining a prior DBN model for a preset accident type that fills the complete conditional probability table.

[0009] According to the above scheme, in step S4, when a newly occurring accident is classified into a "preset accident type", the DBN model trained in steps S1 to S3 is used for probabilistic source tracing and attribution. According to the above scheme, the method for probabilistic source tracing and attribution is as follows: S401. Key feature extraction and discretization: Extract time series data within a specific time period before and after the accident, extract data fields corresponding to the observable nodes of the model, and discretize them using the same standard as in step S3. S402, Evidence Input and Probabilistic Inference: Input the discretized time series data into the trained DBN model, call the inference algorithm, and calculate the posterior probability distribution of each node. S403, Causal chain determination, outputs the probability of occurrence of all hidden variable nodes, that is, the probability of potential upper-level causes of the accident, outputs the dependency strength between nodes, and constructs the most likely causal chain from "subjective cause" to "objective manifestation". This invention also provides a divide-and-conquer probabilistic accident tracing system for assisted driving that integrates feature filtering and dynamic Bayesian networks, comprising: The key field filtering module is based on key feature filtering of mixed datasets. It is used to construct a mixed training dataset containing positive samples of accident data and negative samples of normal driving data for homogeneous accident types preset by experts. The dataset is trained using machine learning algorithms, and the key fields that are relevant to this type of accident are quantitatively filtered from massive signals by ranking the feature importance output by the model. A two-layer dynamic Bayesian network skeleton is constructed to build a two-layer DBN topology structure of "objective observable level - subjective cause definition level". The selected key data fields are used as objective observable nodes, and their physical causal connections and temporal dependencies between slices are established based on vehicle dynamics. At the same time, latent variable cause nodes are introduced, and generative directed edges are constructed from subjective causes to objective manifestations, thus constructing a complete two-layer dynamic Bayesian network skeleton that covers physical laws and potential cause logic. The output prior DBN model module learns DBN parameters based on the accident-specific dataset, constructs a special dataset containing only this type of accident data, discretizes the key data fields and inputs them into the DBN empty skeleton, uses maximum likelihood estimation to learn the objectively observable node parameters, and uses the expectation-maximization EM algorithm to iteratively train the latent variable parameters in the case of missing causal labels, and finally outputs the prior DBN model with a complete conditional probability table. The automated accident determination module uses probabilistic tracing of new accidents. When a new accident occurs and is classified into this type, it extracts and discretizes the corresponding key evidence data, inputs it as evidence into the trained DBN model, and uses inference algorithms to calculate the posterior probability distribution of each node. Based on the probability of occurrence of latent variable nodes, it outputs the causal chain with the highest probability of "subjective cause leading to objective performance", thus realizing the automated determination of accidents.

[0010] The method for tracing the source of assisted driving accidents by integrating feature screening and dynamic Bayesian networks, as described in this invention, has the following beneficial effects: 1. This invention opens the "black box" for the first time through probabilistic reasoning of DBN. It can quantitatively and objectively distinguish these different modes of "human failure", providing unprecedented technical evidence for solving this core contradiction. 2. This invention provides a scientific quantitative basis for legal adjudication and insurance claims. The output of the posterior probability of the upper-level cause is not to replace judges in making legal judgments, but to serve as a scientific and objective "expert technical evidence". This report can help judicial institutions and insurance companies make fairer and more efficient decisions in the face of massive and complex vehicle data. It is not only powerful evidence to protect car companies from unreasonable defenses, but also a key tool to protect drivers from unfair accusations of system defects. 3. One of the most important practical implications of the safety iteration and industry supervision of the driving assistance system of this invention lies in its ability to feed back into the industry. By performing batch attribution on a large amount of accident data, systemic safety hazards can be revealed. If the model calculates a significantly higher probability of "vehicle system failure" in multiple similar accidents, this will become a strong data signal to trigger automakers to carry out OTA (over-the-air) upgrades or initiate large-scale recalls. It also provides government regulatory agencies with a scientific basis for assessing the maturity of ADAS technology, introducing new safety standards, or upgrading road infrastructure. 4. This invention clarifies the automated analysis path from "data" to "responsibility," constructs a complete end-to-end automated process from raw high-dimensional data to quantitative attribution, objectively selects features through machine learning, and performs probabilistic reasoning through a DBN model that integrates expert knowledge, ultimately achieving refined and automated differentiation of accident responsibility; it pushes accident investigation from subjective qualitative analysis to a new stage of objective combined with subjective quantitative analysis, which has significant technical value and social significance for improving traffic safety, perfecting driver assistance systems, and clarifying future human-machine rights and responsibilities. Attached Figure Description

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the assisted driving accident tracing method that integrates feature screening and dynamic Bayesian networks according to the present invention; Figure 2 This is a schematic diagram of the internal structure of the Bayesian network slice for side collisions in this invention. Detailed Implementation

[0012] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0013] like Figure 1As shown, the assisted driving accident tracing method of the present invention, which integrates feature screening and dynamic Bayesian networks, includes the following steps: S1. Key feature screening based on hybrid datasets: For homogeneous accident types preset by experts, a hybrid training dataset containing positive samples of accident data and negative samples of normal driving data is constructed. Machine learning algorithms are used for training, and key fields with strong correlation to this type of accident are quantitatively screened from massive signals by ranking the feature importance output by the model.

[0014] Using a data-driven approach, key observable features that distinguish this "preset accident type" from "normal driving" are extracted from massive amounts of data. The specific steps include: S101. Construct a hybrid training dataset. Extract two types of data from historical accident data: accident segments labeled with the "preset accident type" are used as positive samples, and a large number of normal driving segments are used as negative samples. Perform time-series alignment on all data and extract raw data fields from vehicle bus (CAN), ADAS sensor, and environmental perception data as the original feature set. X raw Construct the target vector Y label As labels for supervised learning, accident samples are labeled as 1, and normal samples are labeled as 0.

[0015] S102. Feature Importance Ranking. A supervised learning algorithm from machine learning is used to train the above mixed dataset for binary classification. After training, the feature importance index output by the model is extracted, and the original features are sorted from high to low importance to obtain a feature importance ranking list. This list provides quantitative data support for the subsequent selection of "observable nodes" in the Bayesian network.

[0016] S2. Construct a two-layer DBN topology structure of "objective observable level - subjective cause definition level". Use the key data fields selected in step one as objective observable nodes and establish their physical causal connections and temporal dependencies between slices based on vehicle dynamics. At the same time, introduce expert knowledge to define unobservable latent variable cause nodes and construct generative directed edges from subjective causes to objective manifestations, thereby further constructing a complete two-layer dynamic Bayesian network skeleton that covers physical laws and potential cause logic.

[0017] Based on the screening results and combined with vehicle physics principles, a dynamic Bayesian network (DBN) structure is constructed to describe the objective physical layer of an accident. This structure supports the incorporation of expert knowledge, allowing the objective physical layer's manifestations to be traced back to the subjective causes, thus completing the construction of the two-layer network structure. Specifically, the following steps are included: S201. Determine Objective Observable Nodes. Select the top-N data fields from the ranking list in step S1 as objectively observable nodes of the DBN. These nodes represent the physical quantities that are most critical, significant, and relevant to this type of accident. Considering the potential conflict between data-driven approaches and expert knowledge, the feature ranking list generated in step S1 serves as an input reference for experts. When defining the nodes of the DBN skeleton, technical personnel ultimately have the right to select, add, or delete from this ranking list to ensure the causal integrity and interpretability of the model.

[0018] S202. Construct the objective layer structure of the DBN. Based on vehicle dynamics principles, traffic engineering logic, and industry consensus, establish physical causal connections between nodes within the same time slice T. This constitutes the network structure within the objective layer slice of the quantitative DBN. Based on the Markov assumption, establish the temporal dependencies of the same node between adjacent time slices (T to T+1), thus forming the temporal evolution structure between slices of the DBN.

[0019] S203. Define hidden variable nodes. Based on the underlying mechanism of the "preset similar accident," domain experts define a set of unobservable upper-level cause nodes as hidden variable nodes in the DBN structure. These nodes represent potential upper-level causes of the accident, including but not limited to driver behavior factors such as "improper lane changing" and "improper speed control," vehicle system states such as "brake failure," or environmental factors such as "complex traffic environment."

[0020] S204. Construct the subjective layer structure of DBN. Construct directed edges from the subjective layer to the objective layer. This connection represents the direct causal relationship from "cause" to "manifestation". Through this step, the intra-slice structure of DBN is expanded into a two-layer structure of "objective layer-subjective layer", forming a complete empty skeleton of DBN.

[0021] S3. DBN parameter learning based on accident-specific dataset: Construct a special dataset containing only accident data of this type. After discretizing the key data fields, input them into the empty DBN skeleton. Use maximum likelihood estimation to learn the objectively observable node parameters. Use the expectation-maximization (EM) algorithm to iteratively train the latent variable parameters in the absence of causal labels. Finally, output a prior DBN model that fills the complete conditional probability table.

[0022] Using pure accident data, the DBN empty skeleton generated in step S2 is subjected to parameter training and conditional probability table (CPT) learning and filling to enable it to have quantitative reasoning ability. The specific steps include: S301. Construct an accident dataset. Extract only historical accident data belonging to this "preset accident type", excluding normal driving data, and retain only the data fields selected in step two.

[0023] S302. Data Discretization. Continuous data is discretized and transformed into state values ​​suitable for Bayesian network processing.

[0024] S303, Parameter Learning. The discretized accident dataset is input into the DBN empty skeleton. For the physical causal relationships between objectively observable nodes, maximum likelihood estimation (MLE) is used to learn the conditional probability table. For structures containing latent variable nodes, the expectation-maximization (EM) algorithm is used for iterative training to automatically learn the probabilistic dependencies between latent and observable variables. Finally, a prior DBN model for this preset accident type is obtained, filled with a complete conditional probability table.

[0025] S4. Probabilistic tracing of new accidents: When a new accident occurs and is classified into this type, the corresponding key evidence data is extracted and discretized. This data is then used as evidence input into the trained DBN model. The inference algorithm is used to calculate the posterior probability distribution of each node. Finally, based on the probability of occurrence of the latent variable nodes, a causal chain with the highest probability of "subjective cause leading to objective performance" is output, thereby achieving automated accident determination.

[0026] In practical application, when a newly occurring accident is categorized into a "preset accident type," the DBN model is used for probabilistic attribution and causation. This includes the following steps: S401. Key Feature Extraction and Discretization. Extract time-series data within a specific time period before and after the accident, extract data fields corresponding to the observable nodes of the model, and discretize them using the same criteria as in step S3.

[0027] S402, Evidence Input and Probabilistic Inference. The discretized time-series data is input into the trained DBN model, and the inference algorithm is called to calculate the posterior probability distribution of each node.

[0028] S403. Causal Chain Determination. Output the probability of occurrence of all latent variable nodes, i.e., the probability of potential upper-level causes of the accident. Output the dependency strength between nodes, constructing the most probable causal chain from "subjective cause" to "objective manifestation", thereby achieving automated and refined determination of the cause of the accident.

[0029] Example Taking the pre-defined accident type of "side collision" as an example, this paper further elaborates on the specific implementation process of this method in real-world scenarios, from model training to new accident inference: S1. Key Feature Filtering for Side Collisions Based on Hybrid Datasets. A hybrid dataset with balanced positive and negative samples is constructed, and a machine learning model is used to filter out objectively observable nodes that are strongly correlated with side collisions. Details are as follows: S101. Constructing the mixed dataset (Mixed_Set_SideImpact), including: (1) Accident Sample Set (Set_SideImpact_Y1): Traverse 200 accident cases marked as "side collision" in the historical database. Using the accident occurrence time T=0 as the baseline, extract time window data from the first 10 seconds to the last 5 seconds (a total of 15 seconds). Set the time slice frequency to 10Hz, that is, generate 150 time slices for each accident, and extract 75 full data fields containing vehicle kinematics and environmental perception, and mark the target vector Traffic_Accident=1.

[0030] (2) Normal Sample Set (Set_SideImpact_Y0): Considering that the core kinematic features of side collisions are mostly related to "lane change", safe and accident-free normal lane change segments are selectively extracted from the "normal driving data database". 200 segments with the same duration of 150 seconds and frequency of 10Hz are randomly selected, and the same 75 data fields are extracted and the target vector Traffic_Accident=0 is marked.

[0031] (3) Merge: Merge the two above to form a mixed training set Mixed_Set_SideImpact containing 400 samples (1:1 balanced).

[0032] S102, Model Training and Feature Ranking: Using all 75 data fields as feature inputs (X) and Traffic_Accident as the target (Y), a random forest model (RF-SideImpact) specifically for side collisions was trained. After training, the model's "global feature importance ranking" was extracted. The results show that the weights of features such as lateral acceleration, SAS steering angular velocity, and lateral distance are significantly higher than those of longitudinal features.

[0033] S103, Key Feature Output: Based on importance ranking and expert screening, the top 35 data fields were selected as the key features most directly related to side collisions, forming the Key_List_SideImpact list (see Appendix Table 1 for details). These 35 fields will serve as the objectively observable nodes for subsequent Bayesian network operations.

[0034] Table 1 Node Numbers and Connection Relationships

[0035] S2. Construct an "objective-subjective" two-layer DBN topology. Based on the selected key evidence and combined with expert knowledge, construct a two-layer causal network skeleton specifically for side collisions. Details are as follows: S201. Construct the objective layer by defining the 35 key pieces of evidence in Key_List_SideImpact as the underlying objective observable nodes, i.e., Y nodes.

[0036] (1) Intra-slice structure: Based on vehicle dynamics, establish physical connections. For example, establish directed edges from "vehicle longitudinal acceleration" to "vehicle velocity". The node numbering, meaning, and connection relationships are shown in Appendix 1. Figure 2 As shown, the Y nodes and their interconnected blue arrows constitute the structure of the objective layer.

[0037] (2) Inter-slice structure: Establish temporal evolution connections. For example, establish T The "lateral distance of the vehicle on the left" at time 1 is an autoregressive connection to the node at time T.

[0038] S202, Subjective Layer and Dual-Layer Extension: (1) Definition of hidden variables: Based on the side collision mechanism, domain experts define the hidden variable nodes of the upper-level causes, such as "improper lane change operation", "improper speed control", "ADAS system failure", etc.

[0039] (2) Causal Derivation (X→Y): Constructing generative connections from the subjective level to the objective level. For example, defining a connection where the latent variable "improper speed control" directly points to the observable node "vehicle speed," and the latent variable "braking system failure" directly points to the observable node "brake pedal opening." See Appendix 2 for details on the definitions of latent variable nodes and their connections. Figure 2 The yellow arrows connecting nodes X and Y represent the causal relationship between the subjective and objective layers.

[0040] Table 2: Latent Variable Node Numbers, Definitions, and Connection Relationships

[0041] S203, Output Skeleton: The final result is a two-layer DBN model skeleton, DBN_Skeleton_SideImpact, specifically designed for side collision detection. Figure 1 As shown.

[0042] S3. DBN Parameter Learning Based on Accident-Specific Data Sets. Using pure accident data, the parameters of the network skeleton are trained to enable it to grasp the physical laws and causal logic under side-impact accidents. Specific details are as follows: S301. Accident-specific set and discretization preparation: Using only the accident sample set (Set_SideImpact_Y1) from step S1, consisting of 200 pure accident data points, normal driving data and label columns were removed. Only the 35 key evidence fields defined in the network structure were retained to construct the training set DataFrame_SideImpact.

[0043] S302. Discretization Processing: A hybrid strategy is used to process continuous data. For example, "vehicle speed" is discretized using feature engineering and thresholding, incorporating road speed limits; "relative distance" is discretized using quartiles. Specific discretization rules are shown in Appendix 3. It should be noted that this discretization method is only an example and is not fixed.

[0044] Table 3. Discretization method of side collision instance data fields

[0045] S303, Parameter Learning: The discretized DataFrame_SideImpact is input into DBN_Skeleton_SideImpact. For the objective layer Y node, its physical coupling probability in the accident is calculated; for the subjective layer X node (data missing), the Expectation-Maximization (EM) algorithm is used for iterative calculation until the log-likelihood function converges. The final output is a prior model Trained_Model_SideImpact.pkl specifically for side collision attribution, filled with a complete conditional probability table (CPT).

[0046] S4. Online Probabilistic Accident Origin Tracing and Causal Chain Determination. Using a trained model, a newly occurring side-impact collision accident is analyzed as an example. Details are as follows: S401, Evidence Extraction and Model Reasoning: (1) Data preparation: Extract the original data of the new accident, lock 35 key pieces of evidence according to Key_List_SideImpact, and discretize them according to the same rules in step three.

[0047] (2) Inference calculation: Load the model Trained_Model_SideImpact.pkl and input the discretized data as observation evidence. The model is activated by the algorithm to infer the posterior probability of all nodes.

[0048] S402. Source tracing result output: (1) Causal chain determination: The posterior probability of the latent variable "improper lane change operation" is as high as 92%.

[0049] (2) Supporting evidence chain: Key evidence shows that 2 seconds before the accident, the driver still entered the SAS steering angle velocity with "state = 'extremely high'" under the combined conditions of "turn signal status = 'off'" and "BSM blind spot monitoring = 'issue alarm'".

[0050] Final conclusion report: Causal chain analysis using the DBN model shows that this factor was primarily caused by improper lane-changing maneuvering by the driver (posterior probability 92%). The physical evidence chain indicates that the driver forcibly performed a high-speed turn without activating the turn signal and ignoring the BSM warning, directly leading to the side collision.

[0051] This completes a full-chain automated tracing system, encompassing multi-source heterogeneous time-series data and probabilistic analysis of accident causes. Through an innovative "divide and conquer" framework, it first objectively filters features based on preset accident types using machine learning and behavior matching sampling. Then, it creatively constructs a two-layer dynamic Bayesian network structure integrating "objective physical evolution" and "subjective causal derivation." This ensures the physical consistency of the inference logic at the bottom layer using vehicle dynamics constraints, while at the top layer, it achieves structured modeling of unobservable latent variables through expert definition. Next, the EM algorithm learns the true causal probability using only accident data. Finally, through dedicated model inference for newly occurring accidents, it achieves reverse probabilistic tracing from the bottom-level physical phenomena to the top-level subjective causes. This solution provides rigorous, interpretable, and scientifically reliable mathematical models and technical support for solving the problem of blurred responsibility boundaries in "human-machine co-driving" scenarios.

[0052] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for tracing the source of assisted driving accidents that integrates feature filtering and dynamic Bayesian networks, characterized in that, Includes the following steps: S1. Key feature screening based on mixed datasets: For the preset homogeneous accident types, a mixed training dataset containing positive samples of accident data and negative samples of normal driving data is constructed. The dataset is trained using machine learning algorithms. By ranking the feature importance output by the model, key data fields that are related to this type of accident are quantitatively screened from massive signals. S2. Construct a two-layer DBN topology structure of "objective observable level - subjective cause definition level". The key data fields selected in step S1 are used as objective observable nodes. Based on vehicle dynamics, physical causal connections and temporal dependencies between slices are established. At the same time, latent variable cause nodes are introduced to construct generative directed edges from subjective causes to objective performance, and a two-layer dynamic Bayesian network skeleton is constructed. S3. DBN parameter learning based on accident-specific dataset: Construct a special dataset containing only accident data of this type, discretize the key data fields and input them into the DBN empty skeleton, use maximum likelihood estimation to learn the objective observable node parameters, and use the expectation-maximization EM algorithm to iteratively train the latent variable parameters in the case of missing causal labels, and output a prior DBN model that fills the complete conditional probability table. S4. Probabilistic source tracing of new accidents: When a new accident occurs and is classified into this type, the corresponding key evidence data is extracted and discretized, and used as evidence input into the trained DBN model. The inference algorithm is used to calculate the posterior probability distribution of each node. Based on the probability of occurrence of the latent variable nodes, the causal chain with the highest probability of "subjective cause leading to objective performance" is output to realize the automatic judgment of the accident.

2. The method for assisted driving accident tracing by fusing feature screening and dynamic Bayesian networks according to claim 1, characterized in that, In step S1, data-driven methods are used to extract key observable features from massive amounts of data to distinguish between "preset accident types" and "normal driving".

3. The method for assisted driving accident tracing by fusing feature screening and dynamic Bayesian networks according to claim 2, characterized in that, A data-driven approach to extract key observable features from massive datasets that distinguish between "preset accident types" and "normal driving" includes the following steps: S101. Construct a hybrid training dataset by extracting positive and negative samples from historical accident data. The positive samples are accident segment data labeled with "preset accident type," and the negative samples are normal driving segment data. All data are time-series aligned, and the original data fields from vehicle bus CAN, ADAS sensor, and environmental perception data are extracted as the original feature set. X raw Construct the target vector Y label As labels for supervised learning, accident samples are labeled as 1, and normal samples are labeled as 0; S102. Feature Importance Ranking: Supervised learning algorithms in machine learning are used to perform binary classification training on the mixed dataset. After training, the feature importance index output by the model is extracted, and the original features are sorted from high to low importance to obtain a feature importance ranking list.

4. The method for assisted driving accident tracing by fusing feature screening and dynamic Bayesian networks according to claim 1, characterized in that, In step S2, the structure of the Dynamic Bayesian Network (DBN) at the objective physical level is constructed, completing the construction of the two-layer network structure, specifically as follows: S201. Determine the objectively observable nodes. Select the top-N data fields ranked in the sorting list in step S1 as the objectively observable nodes of DBN. S202. Construct the DBN objective layer structure. Within the same time slice T, establish physical causal connections between nodes and establish temporal dependencies of the same node between adjacent time slices T and T+1 to form the temporal evolution structure between DBN slices. S203. Define hidden variable nodes, which represent the causes of the accident, including driver behavior factors, vehicle system status, or environmental factors. S204. Construct the subjective layer structure of DBN, and construct directed edges from the subjective layer to the objective layer. The internal structure of DBN slices is expanded into a two-layer structure of "objective layer-subjective layer", forming a complete empty skeleton of DBN.

5. The method for assisted driving accident tracing by fusing feature screening and dynamic Bayesian networks according to claim 1, characterized in that, In step S3, pure accident data is used to train parameters and fill the conditional probability table (CPT) in the empty DBN skeleton generated in step S2.

6. The method for assisted driving accident tracing by fusing feature screening and dynamic Bayesian networks according to claim 5, characterized in that, The method for parameter training and CPT learning to fill in the generated DBN empty skeleton includes the following steps: S301. Construct an accident dataset, extract historical accident data belonging to the "preset accident type", excluding normal driving data, and retain the data fields selected in step S2; S302. Perform data discretization processing, and discretize the continuous data to transform it into state values ​​suitable for Bayesian network processing. S303. Perform parameter learning by inputting the discretized accident dataset into the DBN empty skeleton.

7. The method for assisted driving accident tracing by fusing feature screening and dynamic Bayesian networks according to claim 6, characterized in that, In step S303, the physical causal relationship between objectively observable nodes is learned by using maximum likelihood estimation (MLE) to learn the conditional probability table. For structures containing latent variable nodes, the expectation-maximization algorithm is used for iterative training to automatically learn the probabilistic dependency between latent variables and observable variables, thereby obtaining a prior DBN model for a preset accident type that fills the complete conditional probability table.

8. The method for assisted driving accident tracing by fusing feature screening and dynamic Bayesian networks according to claim 1, characterized in that, In step S4, when a newly occurring accident is classified into a "preset accident type", the DBN model trained in steps S1 to S3 is used for probabilistic source tracing and attribution.

9. The method for assisted driving accident tracing by fusing feature screening and dynamic Bayesian networks according to claim 8, characterized in that, The methods for probabilistic attribution and causation are as follows: S401. Key feature extraction and discretization: Extract time series data within a specific time period before and after the accident, extract data fields corresponding to the observable nodes of the model, and discretize them using the same standard as in step S3. S402, Evidence Input and Probabilistic Inference: Input the discretized time series data into the trained DBN model, call the inference algorithm, and calculate the posterior probability distribution of each node. S403, Causal chain determination, outputs the probability of occurrence of all hidden variable nodes, that is, the probability of potential upper-level causes of the accident, outputs the dependency strength between nodes, and constructs the most likely causal chain from "subjective cause" to "objective manifestation".

10. A divide-and-conquer probabilistic accident tracing system for assisted driving that integrates feature filtering and dynamic Bayesian networks, characterized in that, include: The key field filtering module is based on key feature filtering of mixed datasets. It is used to construct a mixed training dataset containing positive samples of accident data and negative samples of normal driving data for homogeneous accident types preset by experts. The dataset is trained using machine learning algorithms, and the key fields that are relevant to this type of accident are quantitatively filtered from massive signals by ranking the feature importance output by the model. A two-layer dynamic Bayesian network skeleton is constructed to build a two-layer DBN topology structure of "objective observable level - subjective cause definition level". The selected key data fields are used as objective observable nodes, and their physical causal connections and temporal dependencies between slices are established based on vehicle dynamics. At the same time, latent variable cause nodes are introduced, and generative directed edges are constructed from subjective causes to objective manifestations, thus constructing a complete two-layer dynamic Bayesian network skeleton that covers physical laws and potential cause logic. The output prior DBN model module learns DBN parameters based on the accident-specific dataset, constructs a special dataset containing only this type of accident data, discretizes the key data fields and inputs them into the DBN empty skeleton, uses maximum likelihood estimation to learn the objectively observable node parameters, and uses the expectation-maximization EM algorithm to iteratively train the latent variable parameters in the case of missing causal labels, and finally outputs the prior DBN model with a complete conditional probability table. The automated accident determination module uses probabilistic tracing of new accidents. When a new accident occurs and is classified into this type, it extracts and discretizes the corresponding key evidence data, inputs it as evidence into the trained DBN model, and uses inference algorithms to calculate the posterior probability distribution of each node. Based on the probability of occurrence of latent variable nodes, it outputs the causal chain with the highest probability of "subjective cause leading to objective performance", thus realizing the automated determination of accidents.

Images