Intelligent driving responsibility judgment model construction method, device and equipment based on data verification processing

Abstract

The invention relates to the technical field of intelligent driving, in particular to an intelligent driving responsibility judgment model construction method, device and equipment based on data verification processing, and is used for solving the problems that in the prior art, systematic quality verification and multi-dimensional preprocessing cannot be carried out on original vehicle operation data before an intelligent driving responsibility judgment model is constructed, and the model accuracy is low. The method comprises the following steps: acquiring multi-source heterogeneous original vehicle operation data in an intelligent driving scene; according to a preset verification item and a preset verification mode, carrying out quality verification on the original vehicle operation data, and determining target data meeting a quality requirement; pre-processing the target data, and determining the pre-processed target data; and carrying out fusion matching and analysis modeling on the preprocessed target data, and determining an intelligent driving responsibility judgment model. According to the method, the accuracy of intelligent driving responsibility judgment and the accident backtracking efficiency can be improved, and a brand new technical normal form is provided for intelligent driving responsibility judgment.

Description

Technical Field

[0001] The present invention relates to the field of intelligent driving technology, and in particular to a method, device and equipment for constructing an intelligent driving responsibility judgment model based on data verification processing. Background Art

[0002] With the rapid development of technologies such as artificial intelligence, connected vehicles, big data, and automated control, intelligent driving technology, a key development direction in the automotive industry, is gradually evolving from assisted driving to highly automated and unmanned driving. In intelligent driving systems, vehicles can autonomously perform environmental perception, path planning, and decision-making in complex traffic environments, thereby replacing some or all driving tasks. This not only improves driving efficiency and safety but also provides key support for the development of smart transportation and smart cities.

[0003] As intelligent driving technology becomes increasingly popular, the number of traffic incidents and safety accidents that accompany it is gradually increasing. In order to clarify the boundary of responsibility between the intelligent driving system and the driver in an accident, a variety of intelligent driving responsibility determination solutions have emerged in existing technologies. These solutions are mainly based on on-board sensor data, vehicle speed, steering, braking and other operating information, combined with rule models or machine learning methods to analyze and determine the cause of the accident. Some solutions attempt to introduce a responsibility labeling system, using data mining, causal reasoning and other means to improve the accuracy and intelligence of responsibility determination. However, most of these methods rely on high-quality data input in a fixed format and lack comprehensive quality control of the data source, resulting in unstable model training results and weak generalization ability, making it difficult to adapt to the responsibility determination needs in multiple scenarios.

[0004] A key issue with existing technologies in the construction of intelligent driving accountability models is the lack of systematic data quality verification and preprocessing of raw vehicle operation data. Directly using unverified historical or real-time data for modeling and analysis can easily lead to data redundancy, distortion, or loss, thereby impacting the accuracy of the final model and the credibility of accountability identification. Existing solutions struggle to ensure the representativeness and reliability of the data used for model training, especially when data dimensions are inconsistent, data fields are incomplete, or outliers are present. Therefore, there is an urgent need for a method for constructing intelligent driving accountability models based on a data verification and processing mechanism to ensure data quality at the source. This method, through a hierarchical, multi-dimensional data verification and processing strategy, can effectively improve the stability, rationality, and interpretability of the accountability determination model. Furthermore, existing technologies lack analysis of real-time performance and multi-dimensional logical consistency during the data verification process, leading to the risk of misjudgment in complex traffic scenarios. For example, when the vehicle's position changes abnormally at zero speed, existing methods are unable to effectively identify and eliminate abnormal data.

[0005] Patent number CN 115510044A discloses a method for verifying shared traffic data. This method incorporates a shared data acquisition unit to perform data detection and verification on the received data. By incorporating a data cleaning and conversion system, it verifies and extracts incomplete, erroneous, and duplicate data, improving data accuracy. By incorporating a source interface data verification unit, it resolves anomalies in the accessed data. It also preprocesses the stored shared data. By incorporating a road network attribute data verification unit, it surveys the entered data. Combined with manual data inspection and processing, this method improves the accuracy of shared data. However, this method's data verification unit does not analyze the real-time nature of the data. The purpose of real-time analysis is to ensure data quality. Real-time data is compared with historical data to ensure logical consistency between different data points. For example, if the vehicle speed is 0, the position change should also be 0. If the logic is inconsistent, verification efficiency will be reduced.

[0006] How to conduct systematic quality verification and multi-dimensional preprocessing of original vehicle operation data before building an intelligent driving accountability model to improve the accuracy of the model is an urgent problem to be solved. Summary of the Invention

[0007] The present invention provides a method, device and equipment for constructing an intelligent driving accountability model based on data verification processing, which is used to solve the problem that the existing technology cannot perform systematic quality verification and multi-dimensional preprocessing of original vehicle operation data before constructing the intelligent driving accountability model, resulting in low model accuracy.

[0008] In a first aspect, the present invention provides a method for constructing an intelligent driving responsibility judgment model based on data verification processing, the method comprising:

[0009] Acquire multi-source heterogeneous raw vehicle operation data in intelligent driving scenarios;

[0010] Performing quality verification on the original vehicle operation data according to preset verification items and preset verification methods to determine target data that meets quality requirements;

[0011] Preprocessing the target data to determine the preprocessed target data;

[0012] The pre-processed target data is fused, matched, analyzed and modeled to determine the intelligent driving responsibility judgment model.

[0013] Preferably, performing quality verification on the original vehicle operation data according to preset verification items and preset verification methods, and determining target data that passes the verification includes:

[0014] Classifying the original vehicle operation data into historical data and real-time data;

[0015] Performing multiple rounds of adaptive sampling on the historical data through a preset offline dynamic sampling engine to determine first verification qualified data;

[0016] Verifying the real-time data according to a preset real-time data streaming verification channel to determine second verification qualified data;

[0017] The target data is determined according to the first verified data and the second verified data.

[0018] Preferably, the performing multiple rounds of adaptive sampling on the historical data by a preset offline dynamic sampling engine to determine the first verified qualified data includes:

[0019] Obtaining a multi-dimensional feature space including a vehicle operating condition matrix, an environmental feature matrix, and a driving behavior matrix;

[0020] Performing vectorized mapping processing on the historical data according to the multi-dimensional feature space to obtain a first feature data set integrating vehicle state features;

[0021] Using a preset clustering algorithm, perform driving scene cluster analysis on the first feature data set to obtain a scene feature map including scene labels and density information;

[0022] Dynamically updating the sampling weight matrix of historical data according to the distribution density of each scene in the scene feature map to obtain a target sampling weight matrix;

[0023] Performing multiple rounds of adaptive sampling on the historical data according to the target sampling weight matrix to obtain a multiple rounds of sampling data set;

[0024] Performing spatiotemporal constraint verification on the multiple rounds of sampling data sets, eliminating data that does not meet the spatiotemporal constraint requirements, and obtaining the first verified qualified data.

[0025] Preferably, verifying the real-time data according to a preset real-time data streaming verification channel to determine the second verified qualified data includes:

[0026] Buffering and batching the real-time data according to the data access channel of the real-time data to obtain batches of real-time data to be verified;

[0027] Performing millisecond-level data integrity verification on the batch of real-time data to be verified to obtain a first verification data set, wherein the millisecond-level data integrity verification includes using a Bloom filter to perform rapid anomaly marking;

[0028] Performing spatiotemporal logic verification processing on the first verification data set to obtain a second verification data set, wherein the spatiotemporal logic verification processing is based on a pre-built spatiotemporal constraint rule library based on predicate logic;

[0029] Using a preset cross-modal attention alignment network, perform multi-sensor consistency verification on the second verification data set to obtain a third verification data set;

[0030] A time and space constraint check is performed on the third verification data set, and data that does not meet the time and space constraint requirements is eliminated to obtain the second verification qualified data.

[0031] Preferably, the preprocessing of the target data to determine the preprocessed target data includes:

[0032] Perform outlier detection and elimination processing on the target data to determine valid data;

[0033] Performing data completion processing on the missing fields in the valid data according to an interpolation method and a preset filling algorithm to obtain completed data;

[0034] According to the preset intelligent driving insurance business requirements, the completed data is subjected to field reconstruction and feature generation processing to determine the processed data;

[0035] The processed data is subjected to consistency check processing to determine the pre-processed target data.

[0036] Preferably, the performing fusion matching and analytical modeling on the pre-processed target data to determine the intelligent driving responsibility judgment model includes:

[0037] Using a preset spatiotemporal alignment engine, multimodal feature fusion is performed on the preprocessed target data to obtain a fused multimodal behavior feature sequence;

[0038] According to the preset accident causal path map, responsibility factor modeling and attribution analysis are performed on the multimodal behavior feature sequence to obtain the intelligent driving responsibility judgment model.

[0039] Preferably, the use of a preset spatiotemporal alignment engine to perform multimodal feature fusion on the preprocessed target data to obtain a fused multimodal behavior feature sequence includes:

[0040] Performing multi-scale behavioral information modeling on the pre-processed target data to obtain initial spatiotemporal features including millisecond-level microscopic features, second-level mesoscopic features, and minute-level macroscopic trends;

[0041] Based on the initial spatiotemporal features, a pre-constructed hierarchical spatiotemporal coding structure is used to perform fine-grained coding processing on the data of each behavioral level in the initial spatiotemporal features to obtain a behavioral feature sequence with a hierarchical structure;

[0042] Based on the behavioral feature sequence, a spatiotemporal attention mechanism is constructed, and correlation modeling is performed on features of different modalities and time scales in the behavioral feature sequence to obtain a spatiotemporal correlation weight matrix;

[0043] According to the spatiotemporal correlation weight matrix, cross-modal feature alignment and fusion are performed on the behavioral feature sequence to obtain a preliminarily fused multimodal spatiotemporal behavioral feature set;

[0044] According to the multimodal spatiotemporal behavior feature set, a generative adversarial network is used to compensate and reconstruct missing or incomplete spatiotemporal segments to obtain the multimodal behavior feature sequence.

[0045] Preferably, the intelligent driving responsibility determination model is obtained by performing responsibility factor modeling and attribution analysis on the multimodal behavior feature sequence according to the preset accident causal path map, including:

[0046] Extracting events from abnormal behaviors, sudden operations, and system state changes in the multimodal behavior feature sequence to obtain a set of key event nodes for causal modeling;

[0047] Based on the event key node set, events are classified and managed according to behavior categories, temporal relationships and action targets, and a causal event primitive library is constructed;

[0048] Based on the causal event primitive library, a bidirectional temporal convolutional network is used to model the contextual relationships between event nodes, thereby obtaining a causal relationship graph structure between event nodes.

[0049] Based on the causal relationship graph structure, a graph neural network is applied to perform probabilistic reasoning on the causal path to obtain an accident causal path graph with behavioral explanation capabilities;

[0050] According to the accident causal path map, causal responsibility is assigned to each key behavior node to construct the intelligent driving responsibility judgment model.

[0051] In a second aspect, the present invention provides a device for constructing an intelligent driving responsibility judgment model based on data verification processing, the device comprising:

[0052] The raw data acquisition module is used to obtain multi-source heterogeneous raw vehicle operation data in intelligent driving scenarios;

[0053] a quality verification module, configured to perform quality verification on the raw vehicle operation data according to preset verification items and preset verification methods, and determine target data that meets quality requirements, wherein the quality verification includes a value range legitimacy check, a timestamp continuity check, and a logical consistency check;

[0054] A preprocessing module, configured to preprocess the target data and determine the preprocessed target data;

[0055] The model determination module is used to fuse, match, analyze and model the preprocessed target data to determine the intelligent driving responsibility judgment model.

[0056] In a third aspect, the present invention provides an electronic device comprising at least one processor, at least one memory, and computer program instructions stored in the memory, wherein when the computer program instructions are executed by the processor, the method described in the first aspect is implemented.

[0057] In summary, the beneficial effects of the present invention are as follows:

[0058] The embodiment of the present invention provides a method, device and equipment for constructing an intelligent driving responsibility judgment model based on data verification processing. The method includes: obtaining multi-source heterogeneous original vehicle operation data in an intelligent driving scenario; performing quality verification on the original vehicle operation data according to preset verification items and preset verification methods to determine target data that meets quality requirements; preprocessing the target data to determine the preprocessed target data; and performing fusion matching and analysis modeling on the preprocessed target data to determine the intelligent driving responsibility judgment model. The present invention obtains multi-source heterogeneous original vehicle operation data in an intelligent driving scenario, and performs quality verification according to preset verification items and verification methods to ensure the integrity and credibility of the data, eliminate low-quality, abnormal or missing samples from the source, and build a highly reliable data foundation. On this basis, the target data that meets the quality requirements is preprocessed to further regularize the data format and eliminate noise interference, providing structured input for subsequent analysis. By integrating the dimensions of multi-source data through fusion matching technology and combining the analysis and modeling process, an intelligent driving responsibility judgment model with both interpretability and predictive capabilities is effectively constructed. The overall process not only improves the validity and consistency of model input data, but also ensures the accuracy and traceability of model output results, provides high-quality data support and causal reasoning capabilities for responsibility determination in complex scenarios, and significantly breaks through the bottlenecks of traditional models in data quality control and black box decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0060] Figure 1 This is a schematic diagram of the overall process of the method for constructing an intelligent driving responsibility judgment model based on data verification processing in Example 1 of the present invention;

[0061] Figure 2 is a schematic diagram of a process for performing quality verification on the original vehicle operation data in Example 1 of the present invention;

[0062] Figure 3 1 is a flow chart of performing multimodal feature fusion on the preprocessed target data in Example 1 of the present invention;

[0063] Figure 4 1 is a flow chart of performing responsibility factor modeling and attribution analysis on the multimodal behavior feature sequence in Example 1 of the present invention;

[0064] Figure 5 2 is a schematic diagram of the structure of a device for constructing an intelligent driving responsibility judgment model based on data verification processing in Example 2 of the present invention;

[0065] Figure 6 It is a structural diagram of an electronic device in Example 3 of the present invention. DETAILED DESCRIPTION

[0066] The features and exemplary embodiments of various aspects of the present invention will be described in detail below. In order to make the objects, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and Examples. It should be understood that the specific embodiments described herein are only configured to explain the present invention and are not configured to limit the present invention. For those skilled in the art, the present invention can be implemented without the need for some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the present invention.

[0067] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, the elements defined by the phrase "comprising..." do not exclude the presence of other identical elements in the process, method, article, or device comprising the elements.

[0068] Example 1

[0069] See Figure 1 The embodiment of the present invention provides a method, apparatus, device and medium for evaluating the capacity retention rate of a vehicle battery, the method comprising:

[0070] Acquire multi-source heterogeneous raw vehicle operation data in intelligent driving scenarios;

[0071] Specifically, multi-source heterogeneous raw vehicle operation data refers to vehicle perception and control layer data with diverse sources, non-uniform formats, and significant dimensionality differences. For example, during operation, an intelligent driving vehicle simultaneously records data from multiple sensors and control modules, including cameras, radars (lidar, millimeter-wave radar), GPS positioning modules, inertial navigation units (IMUs), and on-board communication bus (CAN). The camera provides image information, the radar outputs distance and obstacle information, the GPS and IMU together constitute positioning data, and the CAN bus records control signals such as the vehicle's real-time speed, braking status, and steering angle. This data varies in acquisition frequency, timestamp accuracy, and data structure, constituting its multi-source heterogeneity. This step aims to provide a true, complete, and time-aligned raw data foundation for subsequent quality verification, preprocessing, and fusion analysis. In the core safety task of intelligent driving, liability determination, the accuracy of the system depends on the ability to accurately restore the full operational state before and after the accident. Therefore, only comprehensive and diverse data collection can ensure that the spatiotemporal context and behavioral chain of key events (such as emergency braking, lane changes, and target recognition errors) are fully captured. Furthermore, acquiring heterogeneous data from multiple sources allows for information complementation in subsequent analysis (for example, radar omissions can be compensated by imagery), significantly enhancing the system's ability to identify and interpret abnormal scenarios. A multi-channel, high-frequency acquisition architecture is employed in its implementation. During vehicle operation, each sensor module connects to the onboard data recording system via an independent communication link. A central data synchronization unit provides unified timestamps and initial caching. To ensure data continuity, mechanisms are required for resuming transmission, handling buffer overloads, and verifying data. Optional local temporary storage and remote cloud synchronization support large-scale scene recording and offline analysis. Furthermore, to enable simultaneous data collection and analysis, some systems embed a lightweight data quality assessment module into the acquisition process to eliminate obvious errors or defects (such as GPS lock loss, image blur), improving the efficiency of subsequent data processing. By efficiently building a spatiotemporal data foundation covering the vehicle's operational status throughout its lifecycle, this system possesses broad applicability and traceability. This complete and diverse data acquisition capability lays a solid foundation for the subsequent "quality-driven-causal explanation" two-wheel model architecture, ensuring that responsibility determinations not only have sufficient data support, but can also be cross-validated between different sources, enhancing the transparency of interpretation and the reliability of results, and significantly improving the system's robustness and legal adaptability in complex and marginal scenarios.

[0072] Performing quality verification on the original vehicle operation data according to preset verification items and preset verification methods to determine target data that meets quality requirements;

[0073] Specifically, the preset verification items include verification indicators developed for multiple dimensions such as data content, format, logic, and timing, such as whether data is missing, whether timestamps are continuous, and whether GPS and radar positions are consistent. The preset verification methods include algorithms or mechanisms for implementing these verifications, such as anomaly detection, logical constraint rule verification, and cross-modal consistency checks. The fundamental goal of this step is to filter out high-quality, complete, and authentic data subsets from complex, chaotic, and potentially flawed raw data. This subset serves as the basis for subsequent model training and inference, preventing judgments from deviating from reality due to data errors or noise. In terms of specific implementation, the original vehicle operation data is first classified into two categories: historical data and real-time data, which go through two quality verification channels, offline and online respectively. For historical data, the "offline dynamic sampling engine" is used for adaptive sampling. The vehicle operating condition matrix (such as speed, steering angle), environmental feature matrix (such as weather, lighting) and driving behavior matrix (such as acceleration mode, frequency of sudden braking) are first extracted to form a multi-dimensional feature space, and then vectorized and mapped into a structured first feature data set. Then, the data is modeled for driving scene distribution through clustering algorithms (such as DBSCAN and K-means) to form a scene feature map containing scene labels and distribution density. The system automatically increases the sampling weight based on the sparse scenes in the map, thereby constructing a representative and diverse historical sample set. Abnormal data is then eliminated through spatiotemporal constraint rules (such as whether the speed change matches the displacement trajectory), and finally the first verification qualified data is formed. For real-time data, a streaming verification channel is used for millisecond-level processing. The system first buffers the data in batches according to the acquisition channel, and quickly marks the missing data through the Bloom filter. The system then uses a predicate logic-based rule base (e.g., GPS offset between two consecutive image frames must not exceed a reasonable threshold) to perform spatiotemporal logic verification. Next, a cross-modal attention alignment network analyzes the consistency of recognition results for the same object from multiple sensors (e.g., radar and camera), thereby selecting a data subset with high sensor coherence. Finally, this data undergoes a spatiotemporal consistency recheck to produce a second set of qualified data. The system fuses the first and second qualified data sets to determine a final, high-quality target dataset for subsequent modeling and analysis. This step implements an intelligent quality screening mechanism that utilizes dual channels of historical and real-time data, coordinates static and dynamic features, and employs multiple layers and algorithms. This not only addresses the issues of traditional systems with fixed sampling ratios and static rules that struggle to cover complex scenarios, but also significantly improves the representativeness and accuracy of the input data for the accountability model, providing a stable and reliable input foundation for subsequent causal modeling and accountability inference. This effectively addresses the core bottlenecks of low data credibility and misjudgments caused by input bias in intelligent driving accountability determination.

[0074] Preferably, see Figure 2, the quality verification of the original vehicle operation data is performed according to the preset verification items and the preset verification method, and the target data that passes the verification is determined to include:

[0075] Classifying the original vehicle operation data into historical data and real-time data;

[0076] Specifically, in the intelligent driving responsibility determination system, the collected raw vehicle operation data must first be classified into historical data and real-time data. Historical data refers to stored past operation records suitable for batch offline analysis, while real-time data refers to the online data stream currently being received by the system, requiring rapid processing and immediate response. This classification facilitates the adoption of the most appropriate quality verification strategy, improving data utilization efficiency and overall system stability.

[0077] Performing multiple rounds of adaptive sampling on the historical data through a preset offline dynamic sampling engine to determine first verification qualified data;

[0078] Specifically, a pre-set offline dynamic sampling engine performs multiple rounds of adaptive sampling on historical data. By extracting multi-dimensional vehicle status and behavior characteristics, it constructs feature vectors and clusters scenarios. This dynamically adjusts sampling weights, increasing sampling for low-coverage or scarce scenarios. Finally, a spatiotemporal constraint rule verification mechanism is combined to eliminate logically conflicting or anomalous data, ultimately determining representative and consistent first-pass validation data, providing a high-quality sample foundation for subsequent modeling.

[0079] Preferably, the performing multiple rounds of adaptive sampling on the historical data by a preset offline dynamic sampling engine to determine the first verified qualified data includes:

[0080] Obtaining a multi-dimensional feature space including a vehicle operating condition matrix, an environmental feature matrix, and a driving behavior matrix;

[0081] Specifically, before data modeling, structured multidimensional feature information is first extracted from historical operating data, thereby constructing a multidimensional feature space. The vehicle operating condition matrix typically includes physical quantities reflecting the vehicle's operating state, such as speed, acceleration, steering angle, and gear position. The environmental feature matrix encompasses external environmental information such as weather, light intensity, visibility, and road conditions. The driving behavior matrix focuses on operational variables, such as frequency of sudden acceleration, number of sudden braking, and steering wheel turn amplitude. This step aims to structure the complex raw data, providing a dimensional, unified numerical representation for subsequent vector modeling and scenario clustering. Given the diverse formats and dispersed sources of raw data, a unified feature space ensures logical comparability and provides a foundation for analysis. Time-series segment data is extracted from the raw data, and statistical features (such as mean, standard deviation, and maximum value) or behavioral indicators (such as action transition points and operation amplitude thresholds) are extracted for each time window. These extracted features are categorized by dimension, constructing three matrices. Ultimately, these matrices are merged under a unified timeline or spatial reference system to form a multidimensional feature space. This method converts the original multimodal time series data into a consistent and computable structured expression, greatly improving the efficiency and accuracy of subsequent modeling, helping to reduce redundant information and improve sample discrimination, thereby improving the modeling accuracy and generalization ability of the entire system.

[0082] Performing vectorized mapping processing on the historical data according to the multi-dimensional feature space to obtain a first feature data set integrating vehicle state features;

[0083] Specifically, after constructing the feature space, each sample's data point undergoes vector mapping. This involves converting each row or time period in the multidimensional matrix into a fixed-length, operable feature vector. These vectors form the first feature dataset, which describes the comprehensive operating state of the vehicle at different times or in different scenarios. By converting complex, multidimensional, and irregular raw data into a unified numerical representation, the data can be used in subsequent machine learning tasks such as cluster analysis and classification. The raw data segments are sliced ​​according to a set time window or event trigger point, and then each sample is represented as a fixed-length vector structure using regularized feature mapping methods (such as normalization and principal component analysis). This mapping method can be flexibly adjusted to meet task requirements, supporting sparse vectors, dense vectors, and even deep feature encoding. This transforms data from a "raw, chaotic" state into a "unified, comparable" feature representation. This not only improves algorithm efficiency but also ensures the accuracy and stability of subsequent cluster analysis, serving as a crucial bridge for the system's transition from data acquisition to intelligent modeling.

[0084] Using a preset clustering algorithm, perform driving scene cluster analysis on the first feature data set to obtain a scene feature map including scene labels and density information;

[0085] Specifically, after obtaining a unified vector representation, the first feature dataset is unsupervisedly clustered using a pre-defined clustering algorithm (such as K-means, DBSCAN, or HDBSCAN), thereby dividing the driving data into multiple driving scene clusters with similar characteristics. The clustering results not only generate scene labels for each category but also count the frequency of each category in historical data, forming a scene feature map containing "label + density information." By extracting typical driving scenario patterns underlying the data and identifying sparse, easily missed, and important scenes, the algorithm guides subsequent data sampling strategies and achieves balanced scene coverage. Based on similarity metrics between feature vectors (such as Euclidean distance and cosine similarity), highly similar samples are clustered together, while low-density areas are considered sparse scenes. The algorithm adaptively determines the number of clusters, avoiding the need for artificially set unreasonable cluster numbers, and outputs a clearly structured and distinct scene distribution map. Organizing the raw data into semantically structured driving scene groups allows the system to prioritize rare scenes or de-redundant high-density scenes, significantly improving sample representativeness and the model's ability to learn from marginal scenes.

[0086] Dynamically updating the sampling weight matrix of historical data according to the distribution density of each scene in the scene feature map to obtain a target sampling weight matrix;

[0087] Specifically, based on the scene feature map obtained after cluster analysis, the sample quantity density of different scene categories is analyzed; for common scenes that appear frequently, their sampling weights are lowered; and for low-frequency or critical scenes, such as emergency braking, sudden appearance of obstacles, etc., their corresponding sampling probabilities are increased to form a new, dynamically adjusted target sampling weight matrix. By solving the problem that traditional fixed-ratio sampling methods easily ignore sparse or abnormal scenes, it ensures that the selected samples have diversity and full coverage in type distribution, and enhances the robustness of the model to complex situations. Among them, the calculation formula of the target sampling weight matrix is ​​as follows:

[0088]

[0089] Among them, W s It represents the weight value of a certain type of sample or a certain data segment in the dynamic sampling process. The larger the value, the higher the priority of the sample in the current stage. α is the proportional coefficient, which is used to adjust the influence of the skewness term (the degree of deviation from the mean) on the total weight. The larger its value, the more the model tends to strengthen the weight of skewed samples. i is the observed value of the i-th sample, which represents the original data input involved in the sampling calculation. μ is the sample mean, which is used to measure the average level of the current sample set and to calculate the degree of deviation. σ is the sample standard deviation, which reflects the degree of dispersion of the current data. 3appears to construct a standardized form of skewness; n is the number of samples, that is, the number of samples used for the current statistical calculation, and is the scale adjustment factor used in the denominator to average the skewness. β is another proportional coefficient used to adjust the contribution strength of the time decay factor to the sampling weight; e -λt is an exponential decay function, indicating that the weight of historical samples gradually decreases as time t increases. Here, λ is the decay rate; larger values ​​result in faster decay. This term is used to prioritize recent samples and reduce reliance on outdated data. t is a time variable, which can represent a timestamp, the interval between a sample and the current time, or a discrete sequence number at a certain time scale. It is used to dynamically adjust the sampling strategy to align with the current scenario. The sampling coefficient of each scenario cluster is dynamically adjusted by combining its inverse density function; for example, the lower the density, the higher the sampling weight, guiding the algorithm to more actively acquire scarce data. This weight matrix is ​​then used as a sampling input control term to drive the subsequent adaptive sampling mechanism. This strategy forms a density-driven, semantically guided dynamic sampling mechanism, significantly improving scenario coverage, preventing the model from overfitting only to mainstream scenarios, and enhancing the broad-spectrum adaptability of the accident responsibility identification system.

[0090] Performing multiple rounds of adaptive sampling on the historical data according to the target sampling weight matrix to obtain a multiple rounds of sampling data set;

[0091] Specifically, using the target sampling weight matrix constructed above, a multi-round, adaptive random sampling process is performed on historical data. Multi-round sampling refers to the fact that not all samples are determined at once, but rather the sampling distribution is evaluated round by round, with the sampling strategy continuously fine-tuned to iteratively obtain a structurally balanced sample set, referred to as a multi-round sampling dataset. This aims to further improve data representativeness and balance, ensuring that the training data covers common scenarios while not overly concentrating on repetitive scenarios, thereby enhancing the model's generalization and resistance to overfitting. In practice, the target weight matrix is ​​used as the sampling distribution function. After each round of random sampling, its coverage and scene category distribution are evaluated, and the sampling probability for the next round is dynamically adjusted. The system can set stopping conditions (such as reaching a specific KL divergence or scene coverage) to control the convergence of the sampling process. This strategy combines statistical sampling with semantic scene weights to form a progressively optimized sampling process for complex and heterogeneous data. This strategy can obtain more diverse and complete training data with limited resources, significantly enhancing the model's ability to adapt to complex road conditions and extreme behaviors.

[0092] Performing spatiotemporal constraint verification on the multiple rounds of sampling data sets, eliminating data that does not meet the spatiotemporal constraint requirements, and obtaining the first verified qualified data.

[0093] Specifically, after completing the sample collection, the sampled data set is further subjected to a spatiotemporal constraint check, wherein the spatiotemporal constraint check includes a kinematic consistency check rule, and the formula of the kinematic consistency check rule is as follows:

[0094]

[0095] in, Indicates that the constraint is executed for all time points t in the time set T. The time set T can be a continuous sampling time series, such as [0s, 0.1s, 0.2s, ...]; Indicates the existence of velocity v t , acceleration a t and direction angle θ t Triples that describe the dynamic state of the vehicle at time t. t+1 represents the actual position vector of the vehicle at time point t+1, which can be measured by sensors (such as GPS or visual SLAM); pt represents the current position vector of the vehicle at time point t. t Δt represents the time interval Δt, which is determined by the speed v t The displacement increment caused by the uniform displacement part; 1 / 2a t Indicates that within the time interval Δt, the acceleration a t The non-uniform displacement increment caused by the acceleration is the additional displacement caused by the acceleration; ∈ represents the maximum allowable error threshold, which is used to control whether the judgment is approximately equal. If the deviation between the predicted position and the actual position is less than ∈, the current state is considered to meet kinematic consistency. Spatiotemporal constraints refer to whether the sample data is continuous in time series and reasonable in spatial trajectory. For example, if the GPS position jump exceeds the allowable speed range, or the braking event does not match the speed change, it is considered to be spatiotemporal conflict data. This step further eliminates unreliable samples from the sampled data that may be caused by sensor errors, data loss, or recording anomalies, thereby ensuring that the final dataset has a high degree of logical consistency and spatiotemporal realism. During the implementation process, a rule library based on predicate logic is called to combine factors such as timestamps, GPS trajectories, and speed changes for logical reasoning. For example, if the vehicle changes from a standstill to 20 km / h within 100ms, it is unreasonable. At the same time, graph neural networks can also be used to analyze the rationality of the path to capture the semantic constraint relationship between frames. Through this high-precision spatiotemporal logic verification process, a first-class verification qualified data set with complete structure, correct logic, and usable for accountability modeling is finally formed, which greatly improves the data quality, avoids misleading model training and decision output due to individual abnormal samples, and ensures the interpretability and credibility of the system in real scenarios.

[0096] Verifying the real-time data according to a preset real-time data streaming verification channel to determine second verification qualified data;

[0097] Specifically, the system receives multi-source data streams from vehicle sensors and control modules in real time, divides the data into processing batches using a cache batching mechanism, and then performs millisecond-level data integrity checks (such as using Bloom filters to quickly identify duplicate or missing data), spatiotemporal consistency checks based on predicate logic rules, and consistency alignment checks between multimodal sensors. Ultimately, the real-time data that passes all checks is aggregated into a second validation-qualified dataset, which serves as input for the immediate accountability model or anomaly warnings, ensuring that the system can respond quickly and reliably to emergencies based on accurate, complete, and synchronized field data.

[0098] Preferably, verifying the real-time data according to a preset real-time data streaming verification channel to determine the second verified qualified data includes:

[0099] Buffering and batching the real-time data according to the data access channel of the real-time data to obtain batches of real-time data to be verified;

[0100] Specifically, in intelligent driving systems, data access channels refer to the transmission paths used to receive real-time data streams from sensors, control units, or edge devices. Common examples include CAN buses, Ethernet, and wireless links. Due to the high transmission frequency and large data volumes of these channels, direct, piecemeal processing can lead to resource waste and latency. Therefore, a buffered batch processing mechanism is employed. This mechanism first caches the real-time incoming data for a short period of time and then batches it according to time windows or data frame numbers. This step aims to improve the efficiency of real-time data processing, avoiding the frequent I / O overhead associated with processing individual data pieces. It also provides structured input for subsequent batch verification, enhancing overall throughput. An independent buffer is configured for each access channel. Scheduling strategies (such as circular buffering and sliding time windows) are used to aggregate data and form time-stamped batches of real-time data to be verified. A maximum buffering delay threshold (e.g., 100ms) can be set to ensure timely data processing while avoiding frame loss. This step builds a bridge between the real-time data flow and the verification module. Its technical characteristics are that it balances real-time performance with stability, effectively supporting the subsequent parallel processing of high-frequency data, reducing latency accumulation, and improving the system's responsiveness to emergencies.

[0101] Performing millisecond-level data integrity verification on the batch of real-time data to be verified to obtain a first verification data set, wherein the millisecond-level data integrity verification includes using a Bloom filter to perform rapid anomaly marking;

[0102] Specifically, millisecond-level data integrity verification refers to determining whether data is missing, duplicated, or formatted incorrectly in a very short time, ensuring the integrity of the input data in terms of content and structure. To improve verification efficiency, a Bloom filter is used, which is a probabilistic data structure with extremely high space efficiency. It is used to quickly determine whether data already exists and effectively identify duplicates or illegal data frames. The purpose of this step is to ensure the quality control of input data under time-sensitive conditions, prevent logical errors or model misjudgments due to repeated reporting, packet loss, or communication jitter, and ensure that streaming processing can respond at the millisecond level. In implementation, each data batch is quickly hashed through a Bloom filter to compare its feature summary to see if it already exists in the cached hash set, and abnormal frames are marked in combination with timestamps and structural verification rules. At the same time, missing fields or illegal values ​​(such as NaN, excessive speed) are supplemented or eliminated to generate a first verification data set with standardized structure and complete fields. This step achieves rapid pre-cleaning of high-frequency real-time data, quickly filters problem data through a space-efficient data structure, and significantly reduces the redundancy and error accumulation of subsequent processing. It is a basic link to ensure the reliable online operation of the system.

[0103] Performing spatiotemporal logic verification processing on the first verification data set to obtain a second verification data set, wherein the spatiotemporal logic verification processing is based on a pre-built spatiotemporal constraint rule library based on predicate logic;

[0104] Specifically, spatiotemporal logic verification refers to the consistency verification of data based on the logical relationship between vehicle behavior and the physical environment. For example, the speed of the vehicle should be close to 0 when it is stationary, and the change in steering angle should correspond to the deviation of the directional trajectory. Such rules are expressed as a formal rule base through predicate logic. The purpose of this step is to eliminate potential anomalies in the data, such as GPS jumps, IMU drift, data frame errors, etc., through physical rationality and behavioral consistency judgment, to ensure the authenticity of the input data from a logical level. During the implementation process, each data in the first verification data set is first mapped into a spatiotemporal state triple (timestamp, location, state), and then the preset predicate logic rules (such as IF v=0THENa≈0) are called for reasoning and matching, and the data that does not meet the logical relationship is marked or removed. Exclusive rule subsets can be loaded according to different driving scenarios, such as adding traffic light-related logic in urban roads and adding acceleration and deceleration boundary logic in high-speed scenarios. This step solidifies empirical judgments into an explainable and reproducible logical system by introducing a formal logical reasoning mechanism, thereby improving the intelligence and transparency of data quality control, reducing manual intervention, and helping the system achieve true logical self-verification capabilities.

[0105] Using a preset cross-modal attention alignment network, perform multi-sensor consistency verification on the second verification data set to obtain a third verification data set;

[0106] Specifically, in intelligent driving systems, vehicles are typically equipped with multiple sensors (such as cameras, millimeter-wave radars, and lidars). These sensors independently perceive the environment, but in practice need to collectively identify the same target or scene. To this end, the system introduces a cross-modal attention alignment network to align and correlate heterogeneous information, such as images, point clouds, velocities, and object IDs, to determine whether their perceptions are consistent. This is known as multi-sensor consistency verification. This step aims to address potential recognition discrepancies between data from different modalities, such as when a camera identifies a pedestrian but a radar doesn't, or when the two sensors' distance judgments to the same object differ significantly. This consistency verification eliminates perceptually conflicting or unreliable data segments, improving overall data credibility. This functionality is implemented using a fused Transformer or multi-stream convolutional network to fuse feature encodings from different modalities. An attention weighting mechanism is then constructed, focusing on the degree of match between different sensors regarding the same target's location, category, velocity, and other information. Data pairs with consistency differences exceeding a threshold are flagged, ultimately filtering out data with perceptually consistent synchronization to form a third verification dataset. By using deep learning models to semantically fuse heterogeneous perception results, we ensure collaborative consistency between data at the modal level, effectively reducing the risk of misjudgment caused by sensor failure, misidentification or occlusion, and providing data support for building a trusted multimodal judgment model.

[0107] A time and space constraint check is performed on the third verification data set, and data that does not meet the time and space constraint requirements is eliminated to obtain the second verification qualified data.

[0108] Specifically, after completing the modal consistency screening, the third validation dataset undergoes a further spatiotemporal constraint check. This rigorously constrains the data's final structure. For example, the validation data is checked for instances where the trajectory crosses obstacles or the vehicle completes non-physical motion within a given timeframe. These instances violate physical spatiotemporal constraints and should be eliminated. By ensuring the data's physical authenticity, the system prevents subsequent models from learning spurious scenarios and behaviors, thereby ensuring the stability and security of model decisions in real-world implementation. During implementation, the system constructs a three-dimensional trajectory inference module that combines high-precision maps, obstacle distribution, road network structure, and other information to perform fitting and collision detection on the data trajectories. Furthermore, it uses timing verification (such as the maximum speed difference between consecutive positions) to determine behavioral continuity. Finally, all data that does not meet these criteria is eliminated, forming the second validation-qualified dataset. By establishing a comprehensive quality control pipeline, from perception verification and logic verification to spatiotemporal and physical verification, the accuracy and integrity of the input data are maximized, providing the most authentic and reliable data for the responsibility determination model and improving the overall robustness of the system.

[0109] The target data is determined according to the first verified data and the second verified data.

[0110] Preprocessing the target data to determine the preprocessed target data;

[0111] Preferably, the preprocessing of the target data to determine the preprocessed target data includes:

[0112] Perform outlier detection and elimination processing on the target data to determine valid data;

[0113] Specifically, in the data preprocessing process, outlier detection and removal are the primary steps to ensure data quality. Outliers are values ​​that significantly deviate from normal statistical patterns or logical relationships, such as high speed readings when the vehicle is stationary, sudden and significant GPS drift, and sudden reversals of azimuth. These can be caused by sensor failure, communication errors, or environmental interference. Their presence can cause the model to learn incorrect patterns and even lead to misjudgments. This step aims to remove extreme noise or erroneous inputs from the data, ensuring that subsequent completion, feature construction, and modeling are based on reliable data, thereby avoiding degradation of model performance and the accumulation of logical errors. In actual processing, numerical data is automatically screened using a combination of statistical methods (such as Z-scores and IQR boxplots) and time series feature analysis (such as local stationarity testing and sliding window differencing). Domain rules (such as maximum achievable speed and reasonable acceleration limits) are also used to set upper and lower limits for logical judgment. Detected outliers can be selectively removed or flagged for subsequent special processing. The automated anomaly identification mechanism improves the credibility of the data set, reduces the subsequent data repair costs, and provides a more accurate basis for feature engineering, effectively improving the stable response capability of the responsibility determination model to edge scenarios.

[0114] Performing data completion processing on the missing fields in the valid data according to an interpolation method and a preset filling algorithm to obtain completed data;

[0115] Specifically, during the acquisition process, raw data often contains missing fields due to sensor frame loss, communication delays, or data synchronization issues. For example, acceleration values ​​may be missing at a certain moment or image frames may not align with radar data. To ensure the continuity of feature calculation and model training, data completion is performed using interpolation and a preset filling algorithm. This step aims to maximize the temporal integrity and feature continuity of the data, avoid model input gaps or erroneous inferences caused by missing data, and improve overall analysis accuracy. In practice, continuous completion can be performed for time series data using linear interpolation, spline interpolation, or Kalman filtering. For discrete fields, nearest value filling, mean value interpolation, or rule-based estimation based on associated fields (e.g., inferring acceleration from velocity and angle) can be used. For structured nested fields, such as multimodal synchronized frames, a multi-source fusion strategy can be used to complete missing segments. This completion mechanism integrates data type and scenario characteristics to achieve adaptive strategies, improving data continuity and the completeness of valid dimensions. This provides a more stable input structure for subsequent feature construction and modeling algorithms, enhancing the model's ability to learn from edge cases.

[0116] According to the preset intelligent driving insurance business requirements, the completed data is subjected to field reconstruction and feature generation processing to determine the processed data;

[0117] Specifically, in order to meet the application requirements of intelligent driving responsibility determination in insurance business scenarios, the original data fields need to be reconstructed and processed for feature generation according to preset rules. Field reconstruction includes recoding, combining or splitting the original fields, such as converting GPS latitude and longitude + time into the average vehicle speed field; feature generation includes the construction of derived variables, such as the number of lane changes per unit time, acceleration volatility, behavioral pattern frequency, etc. The purpose of this step is to transform the original physical or original perception data into feature variables with stronger business semantic expression, and to enhance the model's ability to interpret and predict specific business goals such as responsibility judgment and risk assessment. In implementation, the necessary fields are first extracted from the completed data and the business logic processing function is set according to the insurance scenario (such as rear-end collision responsibility determination, emergency stop behavior recognition), and the fields are transformed, aggregated or combined; for example, derivative indicators such as whether there is continuous emergency braking within 10 seconds and brake response delay can be calculated, and new fields can be generated for model input. By integrating domain rules and data-driven feature construction methods, the data representation capabilities in the business context are effectively improved, enabling subsequent models to have stronger scenario understanding and result interpretability, and is a key bridge for converting general perception data into insurance decision-making data.

[0118] The processed data is subjected to consistency check processing to determine the pre-processed target data.

[0119] Specifically, after field processing is complete, the system must further perform consistency checks to ensure that all fields match each other in terms of value range, time alignment, unit specifications, business logic, and other aspects, and that there are no conflicts or anomalies. For example, "Emergency Braking = Yes" should be accompanied by a sudden deceleration change, and "Rear-End Collision = Yes" should identify the obstacle within the collision timeframe.

[0120] The purpose of this step is to conduct a semantic consistency review under unified standards after completion and processing, improve the standardization of data structure, prevent logical conflicts introduced by field processing, ensure that data can be stably input into the model, and guarantee the correctness of reasoning.

[0121] In implementation, a rule engine or semantic constraint checking mechanism is used to batch-verify the logic of inter-field associations and cross-check them against business tags (e.g., accident location, responsible party). Partial inconsistencies can be automatically corrected through recalculation or field replacement, while severe conflicts are eliminated. This allows for the selection of data subsets with high business semantic accuracy and structural integrity, serving as pre-processed target data for modeling and deployment. This significantly reduces the risk of model errors due to data anomalies and improves the reliability and robustness of system deployment.

[0122] The pre-processed target data is fused, matched, analyzed and modeled to determine the intelligent driving responsibility judgment model.

[0123] Specifically, after completing data preprocessing, the modeling phase begins. By fusing, matching, and analyzing the preprocessed target data with unified structure and clear semantics, a core intelligent model for responsibility judgment—the intelligent driving responsibility judgment model—is constructed. Fusion matching refers to the unified embedding and alignment of multi-source heterogeneous features (such as spatiotemporal behavior, environmental scenarios, and driving operations) to ensure the coordinated expression of model inputs; analytical modeling includes methods such as supervised learning, graph reasoning, and causal modeling to identify responsible parties and reconstruct the accident causal chain. By building a responsibility judgment model with business explainability, reasoning transparency, and online deployment capabilities, it is possible to accurately identify the behaviors and responsibility relationships of all parties after an accident occurs and provide data support. Implementation methods include various model paths: for example, using multimodal Transformers for sequence modeling, fusing graph convolutional networks to construct causal graphs, or introducing decision trees + causal discrimination mechanisms to improve model interpretability. In addition, the model must support functions such as backtracking reconstruction, behavior recognition, and behavior attribution, and support the automatic generation of "behavior chains + causal chains" from data. This step transforms complex raw data into an executable and deployable liability judgment tool, providing a model output with legal validity and technical explanatory power for scenarios such as autonomous driving insurance, accident identification, and legal arbitration. This is the core value of the entire system.

[0124] Preferably, the performing fusion matching and analytical modeling on the pre-processed target data to determine the intelligent driving responsibility judgment model includes:

[0125] Using a preset spatiotemporal alignment engine, multimodal feature fusion is performed on the preprocessed target data to obtain a fused multimodal behavior feature sequence;

[0126] Preferably, see Figure 3 , the pre-processed target data is subjected to multimodal feature fusion using a preset spatiotemporal alignment engine, and the fused multimodal behavior feature sequence obtained includes:

[0127] Performing multi-scale behavioral information modeling on the pre-processed target data to obtain initial spatiotemporal features including millisecond-level microscopic features, second-level mesoscopic features, and minute-level macroscopic trends;

[0128] Specifically, in intelligent driving behavior modeling, multi-scale behavior information modeling involves extracting features of driving behavior and vehicle state at different time granularities, thereby reconstructing the complete behavioral chain from braking details to driving mode. Taking driving behavior as an example, millisecond-level micro-features can represent sudden acceleration changes or sharp turns, second-level meso-features can represent lane changes and following behaviors, and minute-level macro-trends correspond to driving stability, fatigue, and other behaviors. This step aims to construct a feature system that comprehensively characterizes the dynamics of a driver's spatiotemporal behavior, forming hierarchical inputs from the detailed level (e.g., braking response) to the strategic level (e.g., risk aversion intentions), providing behavioral evidence for subsequent responsibility reasoning. During implementation, based on a sliding time window and event-triggered mechanism, vehicle state parameters (e.g., speed, direction, pedal state) are statistically analyzed, their rates of change calculated, and patterns detected at the specified time granularity to generate multi-resolution behavior time series. Simultaneously, temporal alignment is performed with sensory data such as video and radar to form an initial spatiotemporal feature matrix with a unified structure. Through this modeling process, driving behavior patterns at different time granularities are effectively integrated, with stronger behavior recognition accuracy and cross-scale information capture capabilities, providing clear data structure and semantic boundaries for subsequent hierarchical modeling and attention mechanisms, and enhancing the model's ability to understand dynamic events.

[0129] Based on the initial spatiotemporal features, a pre-constructed hierarchical spatiotemporal coding structure is used to perform fine-grained coding processing on the data of each behavioral level in the initial spatiotemporal features to obtain a behavioral feature sequence with a hierarchical structure;

[0130] Specifically, the so-called hierarchical spatiotemporal coding structure refers to a network structure that hierarchically models the initial spatiotemporal features according to the time scale and semantic granularity. For example, millisecond-level operations are mapped to micro-behavior codes, second-level behaviors are mapped to action units, and minute-level trends are mapped to driving strategy codes. Its core mechanism is to encode separately by layer and retain the dependencies between layers in the structure. The purpose of this step is to convert the original behavior signal into a structured behavior feature sequence acceptable to the machine learning model, and to ensure that each behavior layer has independent representation capabilities and contextual dependencies, laying the foundation for attention modeling and spatiotemporal fusion. In the implementation process, the initial features are fed into a multi-channel encoding module. Different channels correspond to data of different time granularities. Lightweight CNN, BiGRU or Transformer modules are used to perform sequence modeling on each layer, and a sequence of behavior vectors with semantic hierarchy is uniformly output on the same time axis. This process can also introduce residual connections or gating mechanisms to preserve the original change trend. This hierarchical structure not only improves encoding accuracy, but also provides the ability to explain "at which level the anomaly occurs" during the inference stage, greatly improving the accuracy and traceability of behavior recognition and attribution, while providing a semantically clear and uniformly resolved input sequence for subsequent spatiotemporal attention fusion.

[0131] Based on the behavioral feature sequence, a spatiotemporal attention mechanism is constructed, and correlation modeling is performed on features of different modalities and time scales in the behavioral feature sequence to obtain a spatiotemporal correlation weight matrix;

[0132] Specifically, the spatiotemporal attention mechanism is a mechanism that automatically learns correlations across different time nodes, spatial dimensions, and modalities. Its essence is to enhance the information representation of key frames, key modalities, or key levels by assigning "attention weights." In multimodal scenarios, different modalities (such as images, radar, and control signals) are often semantically coupled but structurally distinct, and behaviors exhibit distinct characteristics at different time scales. This requires the attention mechanism to effectively align and weight these interactions. This step aims to capture the cross-dependencies between behavioral events in time (pre- and post-process logic), space (the relative relationship between the vehicle and the target), and modality (visual and control information), quantifying these dependencies and forming a spatiotemporal correlation weight matrix. At its implementation level, the system first uses positional encoding to introduce temporal information into the behavioral feature sequence. A multi-head attention mechanism (such as the Transformer architecture) then calculates similarity scores across different modalities and time nodes, constructing a global attention map reflecting the strength of relationships between each feature segment and other segments. The output correlation weight matrix guides the subsequent fusion module to extract the most informative feature combinations. Through this step, the system can more accurately locate "who influences whom" and "which behavioral segments have causal paths", significantly enhancing the modeling accuracy and contextual expression capabilities of the behavior chain, while also improving the modeling accuracy of complex interaction scenarios (such as car-to-car, car-to-people).

[0133] According to the spatiotemporal correlation weight matrix, cross-modal feature alignment and fusion are performed on the behavioral feature sequence to obtain a preliminarily fused multimodal spatiotemporal behavioral feature set;

[0134] Specifically, after obtaining the spatiotemporal correlation weight matrix, the system enters the cross-modal feature alignment and fusion phase. Its core task is to semantically align different modalities (such as visual perception, vehicle state, and control actions) based on the attention results, so that they express consistent information on the same time scale and semantic space. This process not only eliminates temporal misalignment between modalities but also enhances information complementarity. The goal of this step is to form a unified set of behavioral expression vectors that integrate multi-source information, providing a rich contextual and perceptual foundation for subsequent behavioral reconstruction and causal reasoning. During implementation, the system selectively aggregates high-weight regions based on the weight matrix, fuses the corresponding modal feature vectors using weighted summation or gating units, and encodes them using a unified embedding space to form a joint representation of the behavioral event. For example, the control signal of a brake operation is fused with the radar range change and image change features during vehicle deceleration into a single behavioral unit. This fusion mechanism overcomes the limitations of traditional independent modal processing and achieves multimodal collaborative expression. This allows the model to comprehensively perceive driving intentions and external environmental changes, improving the system's recognition rate and anomaly detection capabilities for complex behavioral events.

[0135] According to the multimodal spatiotemporal behavior feature set, a generative adversarial network is used to compensate and reconstruct missing or incomplete spatiotemporal segments to obtain the multimodal behavior feature sequence.

[0136] Specifically, even after thorough data cleaning and fusion, missing spatiotemporal segments due to occlusion, sensor failure, and other factors are still unavoidable in real-world scenarios. Therefore, the system incorporates a Generative Adversarial Network (GAN) mechanism to reconstruct these missing or incomplete behavioral segments by compensating for them. This involves generating potentially missing behavioral representations based on context and other modal information. This step aims to maximize the recovery of key behavioral trajectories, ensuring the integrity and coherence of the behavioral chain and avoiding model cognitive gaps or causal chain disruptions caused by missing segments. In implementation, the system uses the fused feature set as input to the discriminator, constructing an adversarial network consisting of a generator and a discriminator. The generator attempts to complete the missing regions, ensuring consistency in statistical distribution and semantic structure with the real behavioral segments; the discriminator attempts to distinguish between real and generated content. During training, the model is optimized through a game-playing mechanism, gradually bringing the generated results closer to real data. This reconstruction mechanism boasts robust data recovery capabilities, addressing the inevitable perception gaps in real-world driving environments. It provides a continuous and complete behavioral description for responsibility analysis and accident replay, effectively improving the system's fault tolerance and operational usability.

[0137] According to the preset accident causal path map, responsibility factor modeling and attribution analysis are performed on the multimodal behavior feature sequence to obtain the intelligent driving responsibility judgment model.

[0138] Preferably, see Figure 4 , the said multimodal behavior feature sequence is subjected to responsibility factor modeling and attribution analysis based on the preset accident causal path map, and the said intelligent driving responsibility judgment model is obtained, including:

[0139] Extracting events from abnormal behaviors, sudden operations, and system state changes in the multimodal behavior feature sequence to obtain a set of key event nodes for causal modeling;

[0140] Specifically, in driving data, abnormal behaviors (such as sudden braking without warning), unexpected maneuvers (such as forced lane changes), and system state changes (such as automated driving to manual takeover) are often key triggers for accidents or risk evolution. By identifying and extracting these behaviors from multimodal behavioral feature sequences, the system can generate important event segments reflecting the driving evolution process, forming a set of key event nodes. This step aims to extract core behavioral nodes with temporal breakpoint significance or causal potential from the continuous, multi-dimensional behavioral sequence, which serve as the basis for constructing the accident causal chain. In implementation, the system uses a combined rule-based and learning approach, incorporating information such as sudden speed changes, angular rate, and control system switching flags. It automatically marks behavioral mutation points using sliding window detection, threshold discrimination, or behavior classifiers. Furthermore, modal consistency verification (such as sudden changes in vision and radar targets) improves event triggering accuracy. This step significantly reduces the computational complexity and noise impact of subsequent modeling through temporal pruning and focusing on sudden behavior mutations, enabling causal analysis to focus on core events with practical driving significance, improving the efficiency and accountability of the entire system.

[0141] Based on the event key node set, events are classified and managed according to behavior categories, temporal relationships and action targets, and a causal event primitive library is constructed;

[0142] Specifically, the causal event primitive library refers to the structured organization of extracted key event nodes according to certain classification rules (such as behavior type, position in the previous and subsequent causal chains, and affected objects), forming a standardized set of event templates that can be directly called by the modeling algorithm. The behavior categories of events may include braking, acceleration, lane changing, target recognition, etc.; the targets of action may be "previous vehicle", "obstacle", "pedestrian" or "the system itself", etc. The goal of this step is to uniformly encode and organize the original behavior nodes according to semantics and structural logic, providing a highly consistent and reusable input format and knowledge structure for subsequent causal modeling. When implemented, the system will first assign a semantic label to each node and identify the causal type between it and other nodes (such as "follow → sudden brake" is a typical causal pair). The system then classifies according to time sequence and spatial location, and marks the cause attributes of the event (active behavior, passive response, external trigger), ultimately constructing a causal primitive library with a hierarchical structure and semantic labels. This step reduces the semantic ambiguity in complex behavior modeling by expressing key events in a structured and templated manner, enabling subsequent modeling models to quickly identify the type, context, and potential consequences of events, thereby improving the accuracy and efficiency of causal modeling.

[0143] Based on the causal event primitive library, a bidirectional temporal convolutional network is used to model the contextual relationships between event nodes, thereby obtaining a causal relationship graph structure between event nodes.

[0144] Specifically, after obtaining a structured set of event nodes, the system uses a bidirectional temporal convolutional network (Bi-TCN) to perform causal modeling. This network architecture is capable of extracting both forward and backward time series dependencies, simulating the cause-and-effect relationships between events and constructing a causal graph structure that reflects the event triggering chain. This step aims to model the mutual influence between events, identifying which behaviors are the cause, process driver, or final outcome of an incident, and transforming these relationships into a structured graph for subsequent reasoning. During implementation, the system feeds each event node vector into the Bi-TCN in chronological order. Using a context window, it simultaneously captures forward and backward dependencies and outputs a causal strength score between each pair of events. Through threshold filtering and edge weight normalization, a directed and weighted causal graph structure is generated, where nodes represent events and edges represent "cause / effect" relationships. This model utilizes a convolutional structure to improve the efficiency of capturing inter-event correlations. Its bidirectional perception mechanism enables the model to identify complex reverse causal relationships or feedback chains, providing the system with a clear causal flow graph, enhancing the model's explanatory power while ensuring real-time performance.

[0145] Based on the causal relationship graph structure, a graph neural network is applied to perform probabilistic reasoning on the causal path to obtain an accident causal path graph with behavioral explanation capabilities;

[0146] Specifically, using a graph neural network (GNN), the system can perform path-level probabilistic reasoning based on a causal graph structure. GNNs propagate and aggregate information across nodes and edges in the graph, enabling "importance assessment" and "path likelihood analysis" of nodes throughout the causal chain. This step aims to extract the most representative paths from the constructed causal graph, explaining "which key behaviors triggered the accident," and constructing a causal path map for accidents that can be used for decision-making and explanation. To implement this step, the system inputs the causal graph into a GNN model, such as a GAT (graph attention network) or a GCN (graph convolutional network). A multi-layer propagation mechanism aggregates causal features between nodes, while also adjusting the direction and intensity of propagation by incorporating edge weights. The system can set the target node as an "accident occurrence" node and reverse-trace and score the probability and criticality of each path. This step enables the system to not only identify the direct causes of an accident but also quantify the "degree of responsibility" of each behavior in the causal chain, improving the model's reasoning power and interpretability, and helping intelligent driving provide clear, data-driven decision-making in scenarios such as legal liability determination and insurance claims.

[0147] According to the accident causal path map, causal responsibility is assigned to each key behavior node to construct the intelligent driving responsibility judgment model.

[0148] Specifically, after obtaining the accident causal path graph, the system assigns causal responsibility to key behavioral nodes. This involves assessing each behavior's contribution to the accident outcome and determining its role in the responsibility chain (primary, secondary, or not responsible). This process combines the importance scores, node categories, and triggering times of the behavior nodes to output a final responsibility allocation plan. This step aims to transform the behavioral causal paths identified in the model into a basis for determining responsibility that can be used in legal, insurance, or automated decision-making, thereby forming a practical intelligent driving responsibility assessment model. In practice, the system scores and categorizes nodes based on the node influence scores obtained from the GNN, combined with the accident type, regulatory rule base, and responsibility assessment criteria. For example, actively overtaking and causing the preceding vehicle to make an emergency evasive maneuver is primarily responsible; the following vehicle failing to maintain a safe distance is secondary responsible. Ultimately, the causal paths and responsibilities are mapped into a structured output (e.g., diagram structure + responsibility labels + textual explanations of the behaviors). This responsibility assessment model not only accurately identifies responsibility but also preserves the behavioral evolution chain and causal reasoning process, producing a complete "accident explanation package." This greatly improves the system's transparency, legal compliance, and practical business applicability, and provides efficient, reliable, and intelligent support for the rapid determination of responsibility and fair adjudication of smart driving accidents.

[0149] Example 2

[0150] like Figure 5 As shown, the present invention provides a device for constructing an intelligent driving responsibility judgment model based on data verification processing, the device comprising:

[0151] The raw data acquisition module is used to obtain multi-source heterogeneous raw vehicle operation data in intelligent driving scenarios;

[0152] a quality verification module, configured to perform quality verification on the raw vehicle operation data according to preset verification items and preset verification methods, and determine target data that meets quality requirements, wherein the quality verification includes a value range legitimacy check, a timestamp continuity check, and a logical consistency check;

[0153] A preprocessing module, configured to preprocess the target data and determine the preprocessed target data;

[0154] The model determination module is used to fuse, match, analyze and model the preprocessed target data to determine the intelligent driving responsibility judgment model.

[0155] The second embodiment of the present invention provides a device for constructing an intelligent driving accountability model based on data verification processing, the device comprising: a raw data acquisition module for acquiring multi-source heterogeneous raw vehicle operation data in an intelligent driving scenario; a quality verification module for performing quality verification on the raw vehicle operation data according to preset verification items and preset verification methods, and determining target data that meets quality requirements, wherein the quality inspection includes a value range legitimacy verification, a timestamp continuity verification, and a logical consistency verification; a preprocessing module for preprocessing the target data and determining the preprocessed target data; a model determination module for performing fusion matching and analytical modeling on the preprocessed target data to determine the intelligent driving accountability model. This device obtains multi-source heterogeneous raw vehicle operation data in an intelligent driving scenario, and performs quality verification according to preset verification items and verification methods to ensure the integrity and credibility of the data, eliminate low-quality, abnormal, or missing samples from the source, and build a highly reliable data foundation. On this basis, the target data that meets the quality requirements is preprocessed, the data format is further regularized, noise interference is eliminated, and structured input is provided for subsequent analysis. By integrating multi-source data dimensions through fusion matching technology and combining it with analytical modeling processes, we effectively construct an intelligent driving liability determination model that combines both interpretability and predictive capabilities. This overall process not only improves the validity and consistency of the model's input data, but also ensures the accuracy and traceability of the model's output results. This provides high-quality data support and causal reasoning capabilities for liability determination in complex scenarios, significantly overcoming the bottlenecks of traditional models in data quality control and black-box decision-making.

[0156] Example 3

[0157] In addition, combined Figure 1 The method for constructing an intelligent driving responsibility model based on data verification processing in the embodiment of the present invention described above can be implemented by an electronic device. Figure 6 A schematic diagram of the hardware structure of an electronic device provided by an embodiment of the present invention is shown.

[0158] The electronic device may include a processor 401 and a memory 402 storing computer program instructions.

[0159] Specifically, the processor 401 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiment of the present invention.

[0160] Memory 402 may include a large capacity memory for data or instructions. By way of example and not limitation, memory 402 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 402 may include removable or non-removable (or fixed) media. Where appropriate, memory 402 may be inside or outside the data processing device. In a specific embodiment, memory 402 is a non-volatile solid-state memory. In a specific embodiment, memory 402 includes a read-only memory (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or a flash memory, or a combination of two or more of these.

[0161] The processor 401 reads and executes computer program instructions stored in the memory 402 to implement any one of the nozzle offset automatic calibration methods in the above embodiments.

[0162] In one example, the electronic device may further include a communication interface 403 and a bus 410. Figure 6 As shown, the processor 401 , the memory 402 , and the communication interface 403 are connected via a bus 410 and communicate with each other.

[0163] The communication interface 403 is mainly used to implement communication between various modules, devices, units and / or equipment in the embodiment of the present invention.

[0164] Bus 410 comprises hardware, software or both, couples the parts of electronic equipment to each other.For example, and not limitation, bus can comprise accelerated graphics port (AGP) or other graphics bus, enhanced industry standard architecture (EISA) bus, front side bus (FSB), hypertransport (HT) interconnection, industry standard architecture (ISA) bus, infinite bandwidth interconnection, low pin count (LPC) bus, memory bus, micro channel architecture (MCA) bus, peripheral component interconnection (PCI) bus, PCI-Express (PCI-X) bus, serial advanced technology attachment (SATA) bus, video electronics standard association local (VLB) bus or other suitable bus or two or more of these combinations.In suitable cases, bus 410 can comprise one or more buses.Although the embodiment of the present invention describes and shows specific bus, the present invention considers any suitable bus or interconnection.

[0165] The above is a detailed introduction to the method, device and equipment for constructing an intelligent driving responsibility judgment model based on data verification processing provided by an embodiment of the present invention.

[0166] It should be understood that the present invention is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, a detailed description of known methods is omitted. In the above embodiments, several specific steps are described and illustrated as examples. However, the method of the present invention is not limited to the specific steps described and illustrated. Those skilled in the art may make various changes, modifications, and additions, or change the order of the steps after understanding the spirit of the present invention.

[0167] The functional blocks shown in the above-described block diagram can be implemented as hardware, software, firmware or a combination thereof. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in unit, a function card or the like. When implemented in software, the elements of the present invention are programs or code segments that are used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link by a data signal carried in a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROMs, flash memories, erasable ROMs (EROMs), floppy disks, CD-ROMs, optical disks, hard disks, optical fiber media, radio frequency (RF) links, etc. The code segment can be downloaded via a computer network such as the Internet, an intranet, etc.

[0168] It should also be noted that the exemplary embodiments described herein describe methods or systems based on a series of steps or devices. However, the present invention is not limited to the order of the steps described above. In other words, the steps may be performed in the order described in the embodiments, or in a different order, or several steps may be performed simultaneously.

[0169] The above description is only a specific embodiment of the present invention. Those skilled in the art will clearly understand that for the convenience and brevity of description, the specific working processes of the systems, modules and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. It should be understood that the protection scope of the present invention is not limited to this. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention.

Claims

1. A method for constructing an intelligent driving responsibility judgment model based on data verification processing, characterized in that: The method comprises: Acquire multi-source heterogeneous raw vehicle operation data in intelligent driving scenarios; Performing quality verification on the original vehicle operation data according to preset verification items and preset verification methods to determine target data that meets quality requirements; Preprocessing the target data to determine the preprocessed target data; The pre-processed target data is fused, matched, analyzed and modeled to determine the intelligent driving responsibility judgment model.

2. The method for constructing an intelligent driving responsibility judgment model based on data verification processing according to claim 1 is characterized in that: The quality verification of the original vehicle operation data is performed according to the preset verification items and the preset verification method, and the target data that passes the verification is determined to include: Classifying the original vehicle operation data into historical data and real-time data; Performing multiple rounds of adaptive sampling on the historical data through a preset offline dynamic sampling engine to determine first verification qualified data; Verifying the real-time data according to a preset real-time data streaming verification channel to determine second verification qualified data; The target data is determined according to the first verified data and the second verified data.

3. The method for constructing an intelligent driving responsibility judgment model based on data verification processing according to claim 2 is characterized in that: The step of performing multiple rounds of adaptive sampling on the historical data by a preset offline dynamic sampling engine to determine the first qualified verification data includes: Obtaining a multi-dimensional feature space including a vehicle operating condition matrix, an environmental feature matrix, and a driving behavior matrix; Performing vectorized mapping processing on the historical data according to the multi-dimensional feature space to obtain a first feature data set integrating vehicle state features; Using a preset clustering algorithm, perform driving scene cluster analysis on the first feature data set to obtain a scene feature map including scene labels and density information; Dynamically updating the sampling weight matrix of historical data according to the distribution density of each scene in the scene feature map to obtain a target sampling weight matrix; Performing multiple rounds of adaptive sampling on the historical data according to the target sampling weight matrix to obtain a multiple rounds of sampling data set; Performing spatiotemporal constraint verification on the multiple rounds of sampling data sets, eliminating data that does not meet the spatiotemporal constraint requirements, and obtaining the first verified qualified data.

4. The method for constructing an intelligent driving responsibility judgment model based on data verification processing according to claim 2 is characterized in that: Verifying the real-time data according to a preset real-time data streaming verification channel to determine the second verification qualified data includes: Buffering and batching the real-time data according to a data access channel of the real-time data to obtain batches of real-time data to be verified; Performing millisecond-level data integrity verification on the batch of real-time data to be verified to obtain a first verification data set, wherein the millisecond-level data integrity verification includes using a Bloom filter to perform rapid anomaly marking; Performing spatiotemporal logic verification processing on the first verification data set to obtain a second verification data set, wherein the spatiotemporal logic verification processing is based on a pre-built spatiotemporal constraint rule library based on predicate logic; Using a preset cross-modal attention alignment network, perform multi-sensor consistency verification on the second verification data set to obtain a third verification data set; A time and space constraint check is performed on the third verification data set, and data that does not meet the time and space constraint requirements is eliminated to obtain the second verification qualified data.

5. The method for constructing an intelligent driving responsibility judgment model based on data verification processing according to claim 1 is characterized in that: The preprocessing of the target data to determine the preprocessed target data includes: Perform outlier detection and elimination processing on the target data to determine valid data; Performing data completion processing on the missing fields in the valid data according to an interpolation method and a preset filling algorithm to obtain completed data; According to the preset intelligent driving insurance business requirements, the completed data is subjected to field reconstruction and feature generation processing to determine the processed data; The processed data is subjected to consistency check processing to determine the pre-processed target data.

6. The method for constructing an intelligent driving responsibility judgment model based on data verification processing according to any one of claims 1 to 5, characterized in that: The fusion matching and analysis modeling of the pre-processed target data to determine the intelligent driving responsibility judgment model includes: Using a preset spatiotemporal alignment engine, multimodal feature fusion is performed on the preprocessed target data to obtain a fused multimodal behavior feature sequence; According to the preset accident causal path map, responsibility factor modeling and attribution analysis are performed on the multimodal behavior feature sequence to obtain the intelligent driving responsibility judgment model.

7. The method for constructing an intelligent driving responsibility judgment model based on data verification processing according to claim 6 is characterized in that: The method of using a preset spatiotemporal alignment engine to perform multimodal feature fusion on the preprocessed target data to obtain a fused multimodal behavior feature sequence includes: Performing multi-scale behavioral information modeling on the pre-processed target data to obtain initial spatiotemporal features including millisecond-level microscopic features, second-level mesoscopic features, and minute-level macroscopic trends; Based on the initial spatiotemporal features, a pre-constructed hierarchical spatiotemporal coding structure is used to perform fine-grained coding processing on the data of each behavioral level in the initial spatiotemporal features to obtain a behavioral feature sequence with a hierarchical structure; Based on the behavioral feature sequence, a spatiotemporal attention mechanism is constructed, and correlation modeling is performed on features of different modalities and time scales in the behavioral feature sequence to obtain a spatiotemporal correlation weight matrix; According to the spatiotemporal correlation weight matrix, cross-modal feature alignment and fusion are performed on the behavioral feature sequence to obtain a preliminarily fused multimodal spatiotemporal behavioral feature set; According to the multimodal spatiotemporal behavior feature set, a generative adversarial network is used to compensate and reconstruct missing or incomplete spatiotemporal segments to obtain the multimodal behavior feature sequence.

8. The method for constructing an intelligent driving responsibility judgment model based on data verification processing according to claim 6 is characterized in that: The intelligent driving responsibility judgment model is obtained by performing responsibility factor modeling and attribution analysis on the multimodal behavior feature sequence according to the preset accident causal path map, including: Extracting events from abnormal behaviors, sudden operations, and system state changes in the multimodal behavior feature sequence to obtain a set of key event nodes for causal modeling; Based on the event key node set, events are classified and managed according to behavior categories, temporal relationships and action targets, and a causal event primitive library is constructed; Based on the causal event primitive library, a bidirectional temporal convolutional network is used to model the contextual relationships between event nodes, thereby obtaining a causal relationship graph structure between event nodes. Based on the causal relationship graph structure, a graph neural network is applied to perform probabilistic reasoning on the causal path to obtain an accident causal path graph with behavioral explanation capabilities; According to the accident causal path map, causal responsibility is assigned to each key behavior node to construct the intelligent driving responsibility judgment model.

9. A device for constructing an intelligent driving responsibility judgment model based on data verification processing, characterized in that: The device comprises: The raw data acquisition module is used to obtain multi-source heterogeneous raw vehicle operation data in intelligent driving scenarios; a quality verification module, configured to perform quality verification on the raw vehicle operation data according to preset verification items and preset verification methods, and determine target data that meets quality requirements, wherein the quality verification includes a value range legitimacy check, a timestamp continuity check, and a logical consistency check; A preprocessing module, configured to preprocess the target data and determine the preprocessed target data; The model determination module is used to fuse, match, analyze and model the preprocessed target data to determine the intelligent driving responsibility judgment model.

10. An electronic device, characterized in that: The method comprises at least one processor, at least one memory and computer program instructions stored in the memory, and when the computer program instructions are executed by the processor, the method according to any one of claims 1 to 8 is implemented.

Images