CN-121989976-A - Large-model-driven automatic driving failure mode cluster mining and classifying method

CN121989976ACN 121989976 ACN121989976 ACN 121989976ACN-121989976-A

Abstract

The invention belongs to the technical field of intelligent driving, and particularly relates to a large-model-driven automatic driving failure mode cluster mining and classifying method, which comprises the steps of firstly acquiring automatic driving test data, detecting failure events according to preset conditions such as AEB triggering, manual takeover and the like, and generating failure fragments through time window slicing; converting the fragments into unified format data according to a structured template containing dimensions such as Ego vehicle tracks, inputting a large model to generate semantic information such as failure abstract, preliminary attribution and risk causal chain, splicing, encoding into a fixed length vector, clustering into failure clusters by combining cosine similarity and HDBSCAN algorithm, and finally refining cluster-level failure modes, typical triggering conditions and the like by means of the large model to output a standardized failure cluster dictionary and a differential analysis report. The invention can solve the problem that the existing automatic driving failure analysis mode lacks of excavating and classifying the automatic driving complete failure cluster.

Inventors

SHI JIA
ZHOU JINYING
ZHANG QIANG
LI CHAOBIN
HUANG JUNFU
LIU YAN
XIONG YINGZHI
ZHANG YUNFEI
CHENG RUI

Assignees

中汽院智能网联科技有限公司
中国汽车工程研究院股份有限公司
中汽院(江苏)汽车工程研究院有限公司

Dates

Publication Date: 20260508
Application Date: 20260209

Claims (9)

1. A large model driven automatic driving failure mode cluster mining and classifying method is characterized by comprising the following steps: S1, acquiring test data in an automatic driving mode, detecting failure events according to preset trigger conditions, and calling a time window to slice each failure event to generate a plurality of failure fragments; S2, describing each failure fragment according to a uniform structural description template to generate structural data in a uniform format; s3, inputting the structured data into a large model, and respectively outputting failure abstract, failure preliminary attribution and risk causal link through the large model to generate semantic information corresponding to the failure fragment; s4, splicing semantic information of each failure segment into a text, converting the text into a fixed length vector, calculating semantic distance between any two failure segments by taking the fixed length vector as a characteristic, and dividing all the failure segments into a plurality of failure clusters by adopting a clustering algorithm; S5, describing the failure mode of each failure cluster based on a large model, refining typical trigger conditions, selecting representative scenes and prompting systematic weaknesses to obtain analysis summary results of the failure clusters; And S6, outputting a failure cluster dictionary and a failure analysis report according to the analysis summary result.
2. The method for mining and classifying large model driven automatic driving failure mode clusters according to claim 1, wherein S1 comprises: S1-1, acquiring continuous test data in an automatic driving mode, and automatically detecting failure events in the test data according to preset triggering conditions, wherein the preset triggering conditions comprise, but are not limited to, an automatic emergency braking AEB triggering mark of 1, a manual take-over mark of 1, a minimum vehicle distance falling below a set threshold value in preset time, TTC falling in preset time, longitudinal or transverse acceleration exceeding a preset comfort threshold value, and a system entering a degradation mode or a protection mode; S1-2, extracting a trigger event from the acquired failure event, and selecting preset time lengths before and after the moment of the trigger event as a failure fragment time window, wherein the expression is as follows: Wherein, the In order to trigger the moment of triggering of the event, For a preset duration of the trigger event before the trigger time in the failure event, The method comprises the steps of triggering a triggering event in a failure event for a preset time length; S1-3, obtaining a plurality of failure fragments from the failure event according to the failure fragment time window divided by S1-2.
3. The method for mining and classifying large model driven automatic driving failure mode clusters according to claim 1, wherein S2 comprises: S2-1, constructing a structural description template based on dimensions of Ego vehicle track characteristics, target object track characteristics, system control and decision data, scene semantic information and key event node information; S2-2, organizing the failure fragments into structured data in a unified format according to a structured description template.
4. The method for mining and classifying large model driven automatic driving failure mode clusters according to claim 1, wherein S3 comprises: s3-1, preprocessing the structured data and inputting the preprocessed structured data into a large model; S3-2, extracting scene core features, system module behavior deviation and failure results from the preprocessed structured data by the large model, and generating a failure abstract in a short text form according to logic of the scene, the deviation and the results; S3-3, extracting failure evidences of a sensing module, a prediction module, a decision module and a control module of the automatic driving system from the preprocessed structured data by the large model, and judging the association degree of module failure and results by combining with the logic of the automatic driving system to generate failure preliminary attribution; s3-4, the large model is connected with scene inducements, system module behaviors, physical quantity changes and failure results in series according to the time sequence of the preprocessed structured data, and corresponding indexes in the structured data are bound in each link to generate a risk causal link; And S3-5, integrating the failure abstract, the failure primary reason and the risk causal chain to generate semantic information of a corresponding failure fragment.
5. The method for mining and classifying large model driven automatic driving failure mode clusters according to claim 4, wherein S4 comprises: S4-1, acquiring semantic information of an ith failure fragment, splicing failure abstract, failure preliminary reason and risk causal chain in the semantic information, and passing through separator in the middle Distinguishing, generating complete semantic text The expression is: Wherein, the The content of the failure abstract in the semantic information of the ith failure fragment; preliminary attribution content for invalidation in semantic information of the ith invalidation segment; The risk causal link content in the semantic information of the ith failure fragment; s4-2, converting the complete semantic text into a fixed length vector through a large model, wherein the expression is as follows: Wherein, the For the output fixed-length vector to be a constant, Is a coding function in a large model, and a fixed length vector is used for the coding function Performing normalization processing to obtain normalized vector ; S4-3, calculating semantic distances between standardized vectors of any two failure fragments based on cosine similarity, and generating semantic vector similarity; S4-4, dividing all failure fragments into a plurality of failure clusters by adopting HDBSCAN clustering algorithm based on semantic vector similarity, wherein the expression is: Wherein each failure cluster Representing a semantically similar failure mode, Is the number of clusters.
6. The method for mining and classifying large model driven automatic driving failure mode clusters according to claim 5, wherein S5 comprises: S5-1, generating cluster-level failure mode description through semantic high-frequency word extraction and large model aggregation based on semantic information in each failure cluster; S5-2, extracting scene trigger conditions, target trigger conditions and system state trigger conditions through a statistical analysis method and a threshold fitting method based on the structured data in each failure cluster, and integrating to generate typical trigger conditions; S5-3, calculating the distance between the semantic vector in each failure cluster and the center of the cluster, and screening a plurality of representative failure fragments to be used as a representative scene of the cluster; S5-4, combining cluster-level failure mode description of each failure cluster, typical triggering conditions and automatic driving system module logic, and generating a systematic weak point prompt through large model and domain knowledge matching; And S5-5, integrating cluster-level failure mode description, typical trigger conditions, representative scenes and systematic weak point prompts, and storing analysis summary results for generating failure clusters according to a unified format.
7. The method for mining and classifying large model driven automatic driving failure mode clusters according to claim 6, wherein S6 comprises: s6-1, filling analysis and summary results of failure clusters output in the step S5 according to fields in a preset failure cluster dictionary; s6-2, designing a failure analysis report based on the user differential requirements, integrating analysis summary results of the failure clusters according to the designed failure analysis report, and outputting a failure analysis report corresponding to the failure clusters.
8. An electronic device comprising a processor and a memory, wherein the memory stores programs or instructions, and the processor performs a large model driven autopilot failure mode cluster mining and classification method according to any one of claims 1-7 by invoking the programs or instructions stored in the memory.
9. A computer-readable storage medium storing a program or instructions for causing a computer to perform a large model driven automatic driving failure mode cluster mining and classification method according to any one of claims 1-7.

Description

Large-model-driven automatic driving failure mode cluster mining and classifying method Technical Field The invention belongs to the technical field of intelligent driving, and particularly relates to a large-model-driven automatic driving failure mode cluster mining and classifying method. Background With the rapid evolution of the automatic driving technology to high-grade (L3 and above), a train enterprise and a testing mechanism can generate massive test data daily in the research and development and verification stage, and the test data comprise simulation test data, closed field test data and actual road test data. Of these data, only a few fragments belong to the "dead fragments" that have a critical value for the system safety analysis and iterative optimization, i.e. specific time fragments that have a significant decrease in performance, a significant increase in risk, triggering manual takeover or intervention of safety mechanisms during the operation of the intelligent driving system. Currently, the safety performance of an automatic driving system has become a core bottleneck of technology landing, the safety verification requirement of a supervision organization on high-level automatic driving is becoming strict, and the industry is in urgent need of efficient and accurate failure mode analysis technology so as to mine system common weaknesses from massive test data and support system-level safety assessment and pre-production inspection. The automatic driving failure analysis modes commonly adopted in the industry at present mainly comprise three types, but all have obvious technical defects, and cannot meet the large-scale, high-precision and systematic failure analysis requirements: The method is very low in multiplexing efficiency, cannot support massive test data generated daily, and cannot compare and analyze failure data of a vehicle type and a cross item due to the fact that different engineers lack of uniform formats for description of the failure reasons. Screening and analyzing based on simple rules or threshold values, namely screening suspected risk fragments through a preset numerical threshold value, and then manually further judging whether the suspected risk fragments are invalid fragments or not. Although the method realizes the preliminary screening of the suspected fragments, the existence of the abnormal working condition can be identified, the core cause of the abnormality cannot be explained, the threshold value setting depends on experience, and the screening omission or the screening error is easy to occur. Based on rough classification of scene conditions, classifying the failure fragments according to basic attributes of the test scene. The method can only realize 'surface layer scene classification', cannot deeply mine the intrinsic mechanism of failure, and cannot automatically induce 'failure mode clusters' from a large amount of failure data, so that engineers are difficult to identify systematic weaknesses of the system in a specific scene. Whereas in the closest prior art, it mainly comprises: 1. the threshold-based abnormality detection automatically screens out 'abnormal conditions' by setting a minimum distance, a TTC threshold, an acceleration threshold and the like, and has the disadvantage that only abnormality can be identified but the cause of the abnormality cannot be explained. 2. Track clustering based on numerical features, which performs cluster analysis on failure tracks, such as using Dynamic Time Warping (DTW) to measure differences between tracks, dividing tracks into clusters. The method has the defects that the method only can reflect the similarity of the track shape, and cannot reveal mechanisms at the aspects of perception, prediction, decision and the like. 3. The accident case aggregation based on text performs natural language analysis on a public accident report or an internal accident text, extracts accident types and scene elements, and has the defects of no binding with specific system operation data and lacking fine-grained failure analysis of 'behavior process level'. Therefore, in the automatic driving failure analysis mode in the prior art, the lack of a set of slave test data cannot be formed existsFailure fragmentExplanation of the mechanismFailure cluster "complete method chain. Disclosure of Invention The invention aims to solve the technical problem of lack of automatic driving complete failure cluster mining and classifying in the existing automatic driving failure analysis mode by providing a large model driven automatic driving failure mode cluster mining and classifying method. The invention provides a basic scheme that a large model driven automatic driving failure mode cluster mining and classifying method comprises the following steps: S1, acquiring test data in an automatic driving mode, detecting failure events according to preset trigger conditions, and calling a time window to slice each failure event to generate