Vehicle data trust verification method and system based on multi-modal alignment

Abstract

The application discloses a vehicle data credible verification method and system based on multi-modal alignment, and relates to the technical field of data processing. The vehicle data credible verification method based on multi-modal alignment comprises the following steps: S1, collecting multi-modal vehicle accident data and performing a pretreatment operation; S2, performing accident event merging judgment based on an evidence closure order code, an upload offset rate, an actual phase difference and a standard phase difference; S3, performing repeated counting identification by means of a maximum same-part record quantity, an accident evidence record quantity, a risk value counted piece by piece, a part deduplication risk value and an accident phase closure lower limit value; and S4, performing risk back-stability verification by comprehensively considering an accident phase closure mean value, a same-loss re-counting mean value, an original abnormal record quantity, a merged accident event quantity, an original record risk value and a merged deduplication risk value. The problems of high accident event quantity, high risk score and redundant explanation chain in the existing multi-modal vehicle accident data credible verification are solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for reliable verification of vehicle data based on multimodal alignment. Background Technology

[0002] With the digitalization of vehicle transactions, insurance claims, repair and re-inspection, and testing and evaluation, the sources of vehicle accident data have gradually expanded from single records to include accident reports, damage assessment reports, repair work orders, re-inspection reports, testing reports, and records uploaded to the platform. Records from different sources form multiple pieces of evidence data related to the same vehicle, the same accident, and the same damaged area. The data content typically includes the vehicle identification number, accident report number, damage assessment report number, repair work order number, re-inspection report number, testing report number, the time of the incident, the upload time, the location name, and the recorded risk value. Current vehicle data verification methods typically sort and summarize data based on record time, document number, vehicle identification number, and location field, then treat each record as independent risk evidence in accident record statistics and risk aggregation.

[0003] For example, the invention patent with announcement number CN108597223B discloses a data processing method and system for describing the behavior of intelligent vehicles, including: acquiring environmental perception information and vehicle state information; extracting the behavior parameters and state parameters of the intelligent vehicle; extracting the behavior parameters and state parameters of obstacles; generating execution object fields, driving action fields, executed object fields, and constraint condition fields; calling the driving statement structure model to concatenate the fields into driving behavior statement information; performing semantic recognition analysis on the driving behavior statements to obtain the predicted running trajectory of obstacles and the state machine behavior information of the intelligent vehicle; generating the running trajectory of the intelligent vehicle, accurately planning the running trajectory of the intelligent vehicle, and simplifying the complexity of the system, thereby improving the path planning time of the intelligent vehicle.

[0004] For example, the invention patent with announcement number CN113221580B discloses a semantic rejection method, a semantic rejection device, a vehicle, and a medium. The semantic rejection method includes: obtaining the text of the voice request and the confidence of the corresponding words; generating confidence features based on the text and the corresponding word confidence, the confidence features including the text and the confidence of the corresponding word segments; merging the confidence features of the context to generate target confidence features; and using a trained semantic rejection model to predict the target confidence features to obtain the rejection result. The semantic rejection model is trained based on a preset multimodal model, which improves the accuracy of semantic rejection.

[0005] However, in actual vehicle accident reporting, the same accident typically generates multiple timestamps at different stages, such as the time of reporting, damage assessment, repair commencement, repair completion, re-inspection, inspection, and upload. Different data sources do not use entirely consistent time-field recording methods; some records use the time of the incident, some use the upload time, and some records are supplemented, delayed, or transferred from other platforms. This makes it difficult to accurately reflect the order of events for the same accident simply by sorting by time. Furthermore, different institutions or platforms have varying granularities in describing damaged parts. For example, the same damage might be described as front-end damage in an accident report, front bumper damage in a damage assessment report, front bumper replacement in a repair work order, and front structure re-inspection in a re-inspection or inspection report. Matching only the original part name or document text can easily lead to the identification of the same damage as multiple different damage objects. There may also be incomplete connections between document numbers. For example, there may be no complete connection between the police report number, damage assessment number, repair work order number and re-inspection report number, which may cause the same accident to be split into multiple accident candidate records during the data aggregation process.

[0006] Therefore, in order to address the above problems, there is an urgent need for a reliable verification method and system for vehicle data based on multimodal alignment. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a method and system for reliable verification of vehicle data based on multimodal alignment, which solves the problems of inflated accident event numbers, inflated risk scores, and redundant interpretation chains in existing reliable verification of multimodal vehicle accident data.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention implements the following technical solution: a vehicle data credibility verification method based on multimodal alignment, comprising: S1, collecting multimodal vehicle accident data and performing preprocessing operations on the multimodal vehicle accident data; S2, performing accident event merging judgment based on evidence closure exponent, upload offset rate, actual phase difference and standard phase difference, and performing accident event merging and stage phase rearrangement according to the merging judgment result; S3, performing duplicate counting identification by the maximum number of records of the same location, the number of accident evidence records, the risk value of each record, the location deduplication risk value and the lower limit of accident phase closure, and performing main damage record locking and auxiliary evidence record rewriting according to the duplicate counting identification result; S4, performing risk stabilization verification by comprehensively considering the accident phase closure mean, the same damage recount mean, the number of original abnormal records, the number of merged accident events, the original record risk value and the merged deduplication risk value, and performing insufficient stabilization marking, back-calculation verification and credibility result writing according to the risk stabilization verification result.

[0011] Furthermore, the specific steps for collecting multimodal vehicle accident data are as follows: Collect multimodal vehicle accident data by obtaining the vehicle identification number, document number, associated document number, transaction time, upload time, part name, text description, and image recognition results from accident reports, damage assessment reports, repair work orders, re-inspection reports, and inspection reports. The transaction time includes the time of reporting the accident, damage assessment, repair commencement, repair completion, re-inspection, and inspection. The upload time indicates the time the accident evidence record enters the platform. The image recognition results include the image part recognition result and the damaged area. The record risk value of the accident evidence record is obtained through a rule engine based on the accident handling stage type, part name, text description, image recognition result, and document number concatenation status.

[0012] Further, the specific steps for preprocessing multimodal vehicle accident data are as follows: The vehicle identification number (VIN), document number, and associated document number are cleaned to remove format errors, and a document number chain relationship is established based on the document number and associated document number; Partiality and damage terms are extracted from the text description, and image part recognition results and damage areas are extracted from the image recognition results. Part names, part words, and image part recognition results are mapped together to standard part numbers, and the standard part number matching relationship is determined; a standard phase table for the accident chain is constructed based on reporting, damage assessment, repair, re-inspection, and testing, and a business process phase number is configured for each accident evidence record; using the VIN and document number chain relationship as the record association benchmark, the business process phase number as the time alignment benchmark, and the standard part number as the part alignment benchmark, multimodal alignment processing is performed on the text description extraction results, image recognition results, vehicle identification number, document number, associated document number, business time, upload time, and part name to obtain the aligned accident evidence record; combined with the vehicle identification number... Evidence closure exponent is generated based on the relationship between the parties, the serial relationship of the document number, and the hit relationship of the standard part number. The actual phase difference, standard phase difference, and upload offset rate are generated based on the business time difference, business process phase number, and upload time difference between two accident evidence records. Using the accident candidate number as the initial statistical key, the accident evidence records within the accident candidate group are counted to obtain the number of accident evidence records. The record risk value of each accident evidence record within the accident candidate group is accumulated to obtain the risk value for each record. The record risk values ​​of all accident evidence records under the target vehicle are accumulated to obtain the original record risk value. The maximum record risk value corresponding to the same standard part number within the same accident candidate group is taken, and then the maximum record risk values ​​corresponding to each standard part number are accumulated to obtain the part deduplication risk value. Using the target vehicle as the statistical key, all accident evidence records before merging are counted to obtain the original number of abnormal records. The multimodal vehicle accident data is normalized using the maximum-minimum normalization method, and a vehicle accident evidence database is constructed after storage.

[0013] Furthermore, the specific steps for merging and judging accident events based on the evidence closure index, upload offset rate, actual phase difference, and standard phase difference are as follows: Obtain the evidence closure index, upload offset rate, actual phase difference, and standard phase difference between the a-th accident evidence record and the b-th accident evidence record; subtract the standard phase difference from the actual phase difference to obtain the phase difference; perform a cosine operation on the phase difference and subtract the cosine operation result from a constant to obtain the phase cosine offset term; add the constant to the evidence closure index to obtain the phase closure support term; add the constant, evidence closure index, upload offset rate, phase cosine offset term, and a minimum positive number to obtain the phase closure constraint term; divide the phase closure support term by the phase closure constraint term to obtain the accident phase closure value.

[0014] Furthermore, based on the merging judgment results, the specific steps for merging incident events, removing uploaded disturbances, and rearranging stage phases are as follows: When a new incident evidence record is received, the incident phase closure value is compared with the closure threshold. When the incident phase closure value is greater than or equal to the closure threshold, the two incident evidence records are merged and written into the same incident event group with a unified incident event number. The incident evidence record with the larger evidence closure exponent value is used as the main phase record, and the other incident evidence record is used as a stage supplementary record. When the incident phase closure value is less than the closure threshold, the independent incident candidate numbers of the two incident evidence records are retained, and the upload time is no longer used as the basis for incident chain sorting. The actual phase difference is regenerated using the reporting time, damage assessment time, repair start time, repair completion time, re-inspection time, and inspection time. The business stage sequence is corrected according to the accident chain standard phase table. For accident evidence records where the upload time is mistakenly recorded as the business occurrence time, the repair completion time as the re-inspection time, or the inspection upload time as the inspection time, the field source is corrected to allow for phase rearrangement at the execution stage. The document number sequence relationship is retrieved by vehicle identification number, document number, and associated document number to update the evidence closure exponent and recalculate the accident phase closure value. If the re-obtained accident phase closure value is still less than the closure threshold, the two accident evidence records are split into different accident candidate groups.

[0015] Furthermore, the specific steps for identifying duplicate entries using the maximum number of records for the same location, the number of accident evidence records, the risk value for each entry, the risk value for deduplication of locations, and the lower limit of accident phase closure are as follows: Obtain the maximum number of records for the same location, the number of accident evidence records, the risk value for each entry, the risk value for deduplication of locations, and the accident phase closure value within the accident event group; read the accident phase closure values ​​for all record pairs within the accident event group, and take the minimum value as the lower limit of accident phase closure; subtract a constant from the maximum number of records for the same location, and then divide by the sum of the number of accident evidence records and the smallest positive number. The congestion rate of the same location is obtained; the risk value of each item included in the risk value is subtracted from the risk value of the location after deduplication, and then divided by the sum of the risk value of each item included in the risk value and the smallest positive number to obtain the recalculated risk reduction rate; the constant is subtracted from the lower limit of the accident phase closure, and then divided by the sum of the lower limit of the accident phase closure and the smallest positive number to obtain the phase loosening rate; the constant, the phase loosening rate and the smallest positive number are added together to obtain the recalculated suppression term; the congestion rate of the same location and the recalculated risk reduction rate are added together and then divided by the recalculated suppression term to obtain the recalculated pressure input; the constant and the recalculated pressure input are added together and then the natural logarithm is performed to obtain the recalculated value of the same loss.

[0016] Furthermore, the specific steps for locking the main damage record and rewriting the auxiliary evidence record based on the duplicate count identification results are as follows: By comparing the duplicate count value and the duplicate count threshold in real time, when the duplicate count value is greater than or equal to the duplicate count threshold, the corresponding accident event group is marked as a duplicate count event group. The unified accident event number and standard location number of each accident evidence record in the group are read, and multiple accident evidence records falling under the same standard location number are filtered out. For records under the same standard location number, the accident evidence record with the earlier business process phase number is retained as the main damage record. When the business process phase numbers are the same, the accident evidence record with the earlier business process phase number is retained as the main damage record. Accident evidence records with larger phase closure values ​​are designated as primary damage records, while the remaining records are rewritten as auxiliary evidence records. For multiple accident evidence records under the same standard location number where primary damage records have been locked, only the record risk value of the primary damage record is used as the location risk entry point. The location deduplication risk value is updated according to the record risk value of the primary damage record, and the updated location deduplication risk value is used to enter the accident risk stabilization verification process. When the recount value of the same damage is less than the recount threshold, it is determined that no duplicate count has been formed within the current accident event group, and the independent risk entry status of each accident evidence record within the current accident event group is retained.

[0017] Furthermore, the specific steps for risk stabilization verification based on the comprehensive accident phase closure mean, the same-damage recount mean, the number of original abnormal records, the number of merged accident events, the risk value of the original records, and the merged deduplication risk value are as follows: Obtain the accident phase closure value, the same-damage recount value, the number of original abnormal records, the accident event number, and the original record risk value for the target vehicle; perform deduplication counting on the accident event numbers under the target vehicle to obtain the number of merged accident events; take the maximum value of the record risk value corresponding to the same accident event number and the same standard part number under the target vehicle, and then sum all the maximum record risk values ​​to obtain the merged deduplication risk value; calculate the lower bound of the accident phase closure value for all accident event groups under the target vehicle and... The average of the repeated loss values ​​is calculated to obtain the accident phase closure mean and the repeated loss mean. The event compression rate is obtained by subtracting the number of merged accident events from the number of original abnormal records and then dividing by the sum of the number of original abnormal records and the smallest positive number. The event stabilization risk reduction rate is obtained by subtracting the merged deduplication risk value from the risk value of the original records and then dividing by the sum of the risk value of the original records and the smallest positive number. The event compression residual term is obtained by squared the event compression rate minus the repeated loss mean. The risk reduction residual term is obtained by squared the risk reduction rate minus the repeated loss mean. The event compression residual term and the risk reduction residual term are negativeized and then exponentially multiplied by the accident phase closure mean.

[0018] Furthermore, the specific steps for performing insufficient stabilization marking, backcalculation verification, and reliable result writing based on the risk stabilization verification results are as follows: By comparing the accident risk stabilization value and the stabilization threshold in real time, when the accident risk stabilization value is less than the stabilization threshold, the target vehicle is marked as an insufficient stabilization record. The current unified accident event number, main phase record, main damage record, auxiliary evidence record, and merged deduplication risk value are latched. After generating phase closure backcalculation instructions and similar damage recalculation verification instructions, the accident risk stabilization value is recalculated. When the recalculated accident risk stabilization value is still less than the stabilization threshold, the target vehicle is written into the accident risk stabilization pending verification table, retaining the conflict accident event group, anomaly source field, stabilization risk decline rate, accident risk stabilization value, event compression rate, and pause output mark. When the accident risk stabilization value is greater than or equal to the stabilization threshold, the target vehicle is marked as an accident risk stabilization passed record, retaining the unified accident event number, main phase record, main damage record, auxiliary evidence record, and merged deduplication risk value, and is written into the vehicle accident evidence database.

[0019] Furthermore, a second aspect of the present invention provides a vehicle data reliability verification system based on multimodal alignment, applying a vehicle data reliability verification method based on multimodal alignment, comprising: a vehicle accident data acquisition and preprocessing module for acquiring multimodal vehicle accident data and performing preprocessing operations on the multimodal vehicle accident data; an accident chain phase merging analysis module for performing accident event merging judgment based on evidence closure exponent, upload offset rate, actual phase difference and standard phase difference, and performing accident event merging and stage phase rearrangement according to the merging judgment result; and a same-damage recalculation suppression identification module. The system is used to identify duplicate counts by using the maximum number of records for the same location, the number of accident evidence records, the risk value for each record, the risk value for deduplication of the location, and the lower limit of the accident phase closure. Based on the duplicate count identification results, it performs main damage record locking and auxiliary evidence record rewriting. The accident risk stabilization verification module is used to perform risk stabilization verification by comprehensively considering the accident phase closure mean, the mean of repeated damage counts, the number of original abnormal records, the number of accident events after merging, the risk value of the original record, and the risk value for deduplication of the merged record. Based on the risk stabilization verification results, it performs insufficient stabilization marking, back-calculation verification, and reliable result writing.

[0020] The present invention has the following beneficial effects:

[0021] (1) This invention transforms the vehicle identification code consistency relationship, document number serial relationship and standard part number hit relationship into evidence closure index, and calculates the accident phase closure value by combining the upload offset rate, actual phase difference and standard phase difference, so that the accident report, damage assessment report, repair work order, re-inspection report and test report can be merged and judged around the same accident chain, thereby achieving the effect of closure merging of the same accident event across stages, effectively solving the problem that the same accident is split into multiple abnormal events due to sorting only by the recording time in the prior art.

[0022] (2) The present invention calculates the recalculation value of the same damage by using the maximum number of records of the same location, the number of accident evidence records, the risk value of each item, the risk value of deduplication of the location, and the lower limit of the accident phase closure. When the recalculation value of the same damage is greater than or equal to the recalculation threshold, the main damage record is locked and the auxiliary evidence record is rewritten, thereby achieving the effect of suppressing the repeated inclusion of the same damage fact. It effectively solves the problem of repeated accumulation of risk amount around the same damage in the prior art of accident reports, damage assessment reports, repair work orders, re-inspection reports and test reports.

[0023] (3) This invention uses the accident event number and the standard part number as the joint statistical caliber, takes the maximum value of the recorded risk value under the same accident event number and the same standard part number and adds them up to form the part deduplication risk value and the merged deduplication risk value, thereby realizing the effect of merging the accident risk amount according to the event and part, effectively solving the problem that the risk value of the original record deviates from the real accident chain caused by accumulating accident evidence records one by one in the prior art.

[0024] (4) This invention calculates the accident risk stabilization value by comprehensively considering the accident phase closure mean, the same loss recount mean, the number of original abnormal records, the number of accident events after merging, the risk value of the original records and the merging deduplication risk value, and performs insufficient stabilization marking, backcount verification and reliable result writing based on the stabilization threshold, thereby realizing the effect of verifying the availability of vehicle accident risk results after merging and deduplication, effectively solving the problem of lack of result stabilization judgment after the accident event merging and duplicate count suppression in the prior art.

[0025] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0026] Figure 1 This is a flowchart of the vehicle data reliability verification method based on multimodal alignment according to the present invention;

[0027] Figure 2 This is a structural diagram of the vehicle data trust verification system based on multimodal alignment according to the present invention;

[0028] Figure 3 This is a distribution diagram of the phase closure status of the accident chain evidence in this invention;

[0029] Figure 4 This invention relates to a three-dimensional response surface for accident risk stabilization based on phase closure and convergence with loss recalculation suppression. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figures 1-4This invention provides a technical solution: a vehicle data reliability verification method based on multimodal alignment, comprising: S1, collecting multimodal vehicle accident data and performing preprocessing operations on the multimodal vehicle accident data; S2, performing accident event merging judgment based on evidence closure exponent, upload offset rate, actual phase difference and standard phase difference, and performing accident event merging and stage phase rearrangement according to the merging judgment result; S3, performing duplicate counting identification by the maximum number of records of the same location, the number of accident evidence records, the risk value of each record, the location deduplication risk value and the lower limit of accident phase closure, and performing main damage record locking and auxiliary evidence record rewriting according to the duplicate counting identification result; S4, performing risk stabilization verification by comprehensively considering the accident phase closure mean, the same damage recount mean, the number of original abnormal records, the number of merged accident events, the original record risk value and the merged deduplication risk value, and performing insufficient stabilization marking, back-calculation verification and reliable result writing according to the risk stabilization verification result.

[0032] Specifically, the steps for collecting multimodal vehicle accident data are as follows: Collecting multimodal vehicle accident data involves obtaining the vehicle identification number, document number, associated document number, business time, upload time, part name, text description, image recognition results, vehicle sensor readings, and driving trajectory fragments from accident reports, damage assessment reports, repair work orders, re-inspection reports, and inspection reports. The business time includes the time of reporting the accident, damage assessment, repair commencement, repair completion, re-inspection, and inspection. The upload time indicates the time the accident evidence record enters the platform. The text description is generated as a text vector after word segmentation and semantic extraction. The image recognition results include image part recognition results and damage areas. Vehicle sensor readings include mileage readings, speed change readings, and collision-related readings. The driving trajectory fragments include... The data includes the changes in location sequence, time sequence, and trajectory direction before and after the accident; text vectors, image recognition results, vehicle sensor readings, driving trajectory fragments, and structured document fields are used as multimodal data sources, and aligned around vehicle identification codes, document number concatenation relationships, standard part numbers, and business process phase numbers to form accident evidence records; the rule engine obtains the record risk value of the accident evidence record based on the accident handling process type, part name, text description, text vectors, image recognition results, vehicle sensor readings, driving trajectory fragments, and document number concatenation status. The record risk value ranges from 0 to 1. The closer the value is to 1, the greater the degree of damage correlation and risk inclusion intensity corresponding to the accident evidence record; the closer the value is to 0, the smaller the degree of damage correlation and risk inclusion intensity corresponding to the accident evidence record.

[0033] This implementation plan obtains vehicle identification numbers, document numbers, associated document numbers, business time, upload time, part name, text descriptions, and image recognition results uniformly from accident reports, damage assessment reports, repair work orders, re-inspection reports, and test reports. This allows accident evidence records from different sources to be grouped under the same target vehicle based on the vehicle identification number. Furthermore, the synchronous organization of text descriptions, image recognition results, and structured fields establishes a correspondence between text modalities, image modalities, and structured fields. Simultaneously, a rule engine is used to obtain record risk values ​​based on the accident handling stage type, part name, text description, image recognition results, and document number concatenation status. This ensures that each accident evidence record has a unified data caliber and risk quantification basis before entering subsequent accident event merging judgment, duplicate counting identification, and risk stabilization verification. This provides traceable data support for the subsequent generation of evidence closure exponents, actual phase differences, standard phase differences, upload offset rates, individual risk values, and part deduplication risk values.

[0034] Specifically, the preprocessing steps for multimodal vehicle accident data are as follows: The vehicle identification number (VIN), document number, and associated document number are cleaned to standardize character length, letter case, spaces, and separators in the VIN. The encoding format of the document number and associated document number is also standardized. Duplicate uploads and records with missing key fields are removed. A sequential relationship between document numbers is then established based on the document number and associated document number. Location terms and damage terms are extracted from the text descriptions. Image location recognition results and damage areas are extracted from the image recognition results. Location names, location terms, and image location recognition results are mapped to standard location numbers, and the matching relationships of these standard location numbers are determined. These standard location numbers are derived from a vehicle location dictionary and a component hierarchy library. The system first performs precise matching between the part name and the vehicle part dictionary. If no match is found, it performs synonym matching between the part name and the part synonym database. Then, it combines the image part recognition results with the component-level database for cross-validation. When the text and image sides point to the same standard part number, the standard part number matching relationship is determined. Based on the reporting, damage assessment, repair, re-inspection, and testing processes, a standard phase table for the accident chain is constructed. The standard phase table for the accident chain includes the reporting phase, damage assessment phase, repair commencement phase, repair completion phase, re-inspection phase, and testing phase. Phase numbers are configured in the order of reporting phase, damage assessment phase, repair commencement phase, repair completion phase, re-inspection phase, and testing phase. A business process phase number is configured for each accident evidence record. Vehicle recognition... The correlation between vehicle identification code and document number serves as the record association benchmark, the phase number of the business process serves as the time alignment benchmark, and the standard location number serves as the location alignment benchmark. Multimodal alignment processing is performed on the extracted text description, image recognition results, vehicle identification code, document number, associated document number, business time, upload time, and location name to obtain the aligned accident evidence record. An evidence closure exponent is generated by combining the vehicle identification code consistency relationship, the document number correlation relationship, and the standard location number match relationship. The evidence closure exponent takes an integer value between 0 and 3. It increases by one level when the vehicle identification code consistency relationship is established, when the document number correlation relationship is established, and when the standard location number match relationship is established. This exponent is used to characterize the two accident evidence records on the vehicle body. The system assesses the closure strength across three dimensions: document link, document link, and damaged location. Based on the business time difference, business process phase number, and upload time difference between two accident evidence records, it generates the actual phase difference, standard phase difference, and upload offset rate. Using the accident candidate number as the initial statistical key, it counts the accident evidence records within the accident candidate group to obtain the number of accident evidence records. It then sums the record risk values ​​of each accident evidence record within the accident candidate group to obtain the individual risk value. Finally, it sums the record risk values ​​of all accident evidence records under the target vehicle to obtain the original record risk value. The system takes the maximum record risk value corresponding to the same standard location number within the same accident candidate group, and then sums the maximum record risk values ​​corresponding to each standard location number to obtain the location deduplication risk value.Using the target vehicle as the statistical key, all accident evidence records before merging are counted to obtain the original number of abnormal records. The multimodal vehicle accident data is normalized using the max-min normalization method, compressing the data involved in the calculation to a uniform range of zero to one, and then storing the data to construct a vehicle accident evidence database.

[0035] In this implementation plan, by cleaning the format of vehicle identification numbers, document numbers, and associated document numbers, a serial relationship of document numbers is established. Part names, part terms, and image part recognition results are uniformly mapped to standard part numbers, enabling accident reports, damage assessment reports, repair work orders, re-inspection reports, and test reports to form comparable accident evidence records around the same vehicle, the same document link, and the same damage part. By constructing a standard phase table for the accident chain and generating business process phase numbers, evidence closure exponents, actual phase differences, standard phase differences, and upload offset rates, the time caliber differences, upload time interference, and stage sequence offsets between records from different sources can be quantified. By generating the number of accident evidence records, risk values ​​for each record, risk values ​​for original records, risk values ​​for deduplicated parts, and the number of original abnormal records, a unified data foundation is provided for subsequent accident event merging and judgment, duplicate counting identification, and accident risk stabilization verification, thereby reducing the risk of the same accident being split into multiple candidate events and the same damage being counted repeatedly.

[0036] Specifically, the steps for merging and judging accident events based on evidence closure index, upload offset rate, actual phase difference, and standard phase difference are as follows: Obtain the evidence closure index, upload offset rate, actual phase difference, and standard phase difference between accident evidence record a and accident evidence record b; whereby the evidence closure index characterizes the degree of closure of the link formed by the vehicle identification code consistency relationship, document number serialization relationship, and standard part number hit relationship, corresponding to the risk of document chain breakage and part misplacement; the upload offset rate is calculated by converting the upload time difference between the two accident evidence records with the upload time difference stability, and is used to characterize the impact of upload time drift on the event. Therefore, the disturbance caused by the chain sorting; the actual phase difference is normalized by the business time difference between the two accident evidence records according to the reference duration of the corresponding accident handling links, and then converted into a dimensionless phase quantity in radians, used to characterize the phase interval between the two accident evidence records in the actual accident chain progression; the standard phase difference is calculated by the phase number difference of the corresponding business links of the two accident evidence records according to the standard phase table of the accident chain, and is also a dimensionless phase quantity in radians, used to characterize the standard phase interval between reporting, damage assessment, repair commencement, repair completion, re-inspection, and testing; subtracting the standard phase difference from the actual phase difference yields the phase difference quantity, which is in radians. A dimensionless parameter is used to characterize the deviation in the accident chain caused by incorrect entry of business time, misuse of upload time, and misalignment of stage sequence; a cosine operation is performed on the phase difference, and the result of the cosine operation is subtracted from the constant to obtain the phase cosine offset term. The closer the actual phase difference is to the standard phase difference, the smaller the phase cosine offset term; the more obvious the deviation between the actual phase difference and the standard phase difference, the larger the phase cosine offset term; the constant is added to the evidence closure exponent to obtain the phase closure support term, which is used to enhance the merging support when vehicle identity is consistent, document link closure is achieved, and standard part number is matched; the constant, evidence closure exponent, upload offset rate, and phase The phase closure constraint term is obtained by adding the cosine offset term and the minimum positive number. This term is used to simultaneously constrain the merging risk caused by upload time drift, stage phase deviation, and insufficient evidence chain closure. The phase closure support term is divided by the phase closure constraint term to obtain the accident phase closure value. The accident phase closure value is used to determine whether the a-th accident evidence record and the b-th accident evidence record can be closed to the same accident event. The closure threshold is obtained by statistically analyzing the distribution of accident phase closure values ​​of the same accident record pair and different accident record pairs in the historical accident evidence sample. This is used to ensure that the comparison results of the accident phase closure values ​​are consistent with the actual merging boundary in the historical sample.

[0037] The specific formula for calculating the accident phase closure value is as follows:

[0038] ;

[0039] In the formula, This represents the accident phase closure value between accident evidence record a and accident evidence record b, reflecting whether the two accident evidence records can be closed to the same accident event; The evidence closure index reflects the closure strength formed by the sequential relationship of document numbers, the matching relationship of standard part numbers, and the consistency relationship of vehicle identification codes. This indicates the upload offset rate, reflecting the degree of time disturbance in the entry of two accident evidence records into the platform link; This indicates the actual phase difference, reflecting the difference in the progression of the accident chain between the two accident evidence records at the actual business time. It represents the standard phase difference, reflecting the theoretical phase interval of risk assessment, repair, re-inspection, and testing in the accident chain standard phase table; It represents the smallest positive number and reflects the basis for denominator correction.

[0040] Table 1 shows the accident phase closure calculation data under multiple evidence strengths in this embodiment. The first accident evidence record pair has an evidence closure index of 0, an actual phase difference of 0, a standard phase difference of 1, and an upload offset rate of 0.5, resulting in a calculated accident phase closure value of 0.51. The second accident evidence record pair has an evidence closure index of 1, an actual phase difference of 1, a standard phase difference of 1, and an upload offset rate of 1.0, resulting in a calculated accident phase closure value of 0.67. The third accident evidence record pair has an evidence closure index of 2, an actual phase difference of 2, a standard phase difference of 1, and an upload offset rate of 0.8, resulting in a calculated accident phase closure value of 0.70. The fourth accident evidence record pair has an evidence closure index of 3, an actual phase difference of 3, a standard phase difference of 1, and an upload offset rate of 0.5, resulting in a calculated accident phase closure value of 0.68.

[0041] Table 1. Calculation data for accident phase closure under multiple evidence strengths

[0042] Accident evidence record 1 0 0 1 0.5 0.51 2 1 1 1 1.0 0.67 3 2 2 1 0.8 0.70 4 3 3 1 0.5 0.68

[0043] like Figure 3 As shown, this is a distribution diagram of the phase closure status of accident chain evidence provided in this application embodiment. The horizontal axis represents the phase difference, the vertical axis represents the upload offset rate, the color intensity represents the accident phase closure value, and the red dashed line represents the closure threshold. The four sub-graphs correspond to four types of evidence closure level states: no matching relationship, only vehicle consistency, vehicle plus document association, and vehicle plus document plus location matching. (Refer to Table 1 and...) Figure 3As can be seen, the evidence closure index of the first set of evidence records is 0, the actual phase difference is 0, the standard phase difference is 1, the upload offset rate is 0.5, and the accident phase closure value is 0.51, corresponding to a low closure state under no matching relationship; the evidence closure index of the second set of evidence records is 1, the actual phase difference is consistent with the standard phase difference, but the upload offset rate rises to 1.0, and the accident phase closure value is 0.67, corresponding to a mergeable boundary state where the vehicles are consistent but upload disturbances still exist; the evidence closure index of the third set of evidence records is 2, the actual phase difference differs from the standard phase difference, the upload offset rate is 0.8, and the accident phase closure value is 0.70, corresponding to a stable merge state under the support of vehicle and document association; the evidence closure index of the fourth set of evidence records is 3. Although the phase difference between the actual phase difference and the standard phase difference increases, the vehicle, document, and location all form closed support, and the accident phase closure value remains at 0.68, corresponding to a strong evidence closure relationship and an anti-phase disturbance state. Table 1 provides a direct and quantitative comparison of the numerical differences in evidence closure index, actual phase difference, standard phase difference, upload offset rate, and accident phase closure value among the four sets of evidence records. It highlights that the accident phase closure value increases with the enhancement of the evidence closure index and decreases with the increase of upload offset rate and phase difference. This verifies that the accident chain phase merging analysis module can incorporate document serialization relationships, location hit relationships, business phase offset, and upload disturbances into the same merging judgment process, providing intuitive data support for subsequent accident event merging, upload disturbance stripping, stage phase rearrangement, and duplicate count identification.

[0044] In this implementation plan, by uniformly converting the actual phase difference and the standard phase difference into a dimensionless phase quantity in radians, the actual accident chain progression interval between two accident evidence records can be compared with the standard progression interval in the accident chain standard phase table. The evidence closure exponent expresses the degree of closure of the link formed by the consistency of vehicle identification codes, the serialization of document numbers, and the hit relationship of standard part numbers. The upload offset rate expresses the disturbance caused by upload time drift to the accident chain sorting. The phase cosine offset term expresses the phase deviation caused by incorrect filling of business time, misuse of upload time, and misalignment of stage order. Then, the phase closure support term and the phase closure constraint term are used to generate the accident phase closure value, so that whether two accident evidence records belong to the same accident event can be quantitatively determined, thereby providing a basis for subsequent accident event merging, upload time exclusion, stage phase rearrangement, and document number serialization relationship review.

[0045] Specifically, the steps for merging accident events, removing upload disturbances, and rearranging stage phases based on the merging judgment results are as follows: When a new accident evidence record is received, the accident phase closure value is compared with the closure threshold. When the accident phase closure value is greater than or equal to the closure threshold, the two accident evidence records are merged and written into the same accident event group with a unified accident event number. The accident evidence record with the larger evidence closure exponent value is used as the main phase record, and the other accident evidence record is used as a stage supplementary record. When the accident phase closure value is less than the closure threshold, the independent accident candidate numbers of the two accident evidence records are retained. The upload time is no longer used as the basis for sorting the accident chain. Instead, the actual phase difference is regenerated using the reporting time, damage assessment time, repair start time, repair completion time, re-inspection time, and testing time. The business stage order is corrected according to the accident chain standard phase table. When the absolute value of the difference between the business time and the upload time of any accident evidence record is less than or equal to the upload time difference stability, and the business time's participation in the sorting causes a stage reversal between reporting, damage assessment, repair start, repair completion, re-inspection, and testing, the following steps are taken: In the first instance, it was determined that the incident evidence record contained a misuse of the field where the upload time was mistakenly recorded as the business occurrence time; when the repair completion time was later than the re-inspection time, and the absolute value of the difference between the re-inspection time and the repair upload time was less than or equal to the upload time difference stability, the incident evidence record was determined to contain a misuse of the field where the repair completion time was mistakenly recorded as the re-inspection time; when the absolute value of the difference between the detection time and the detection upload time was less than or equal to the upload time difference stability, and the detection time was earlier than the re-inspection time or earlier than the repair completion time, the incident evidence record was determined to contain a misuse of the detection upload time. The field for the detection time was misused. To address this misuse, the reporting time, damage assessment time, repair start time, repair completion time, re-inspection time, and detection time were re-entered according to the accident chain standard phase table, and the phase numbers of the business links were reconfigured to perform phase rearrangement. Then, the document number serialization relationship was checked according to the vehicle identification number, document number, and associated document number to update the evidence closure index and recalculate the accident phase closure value. If the re-obtained accident phase closure value is still less than the closure threshold, the two accident evidence records were split into different accident candidate groups.

[0046] In this implementation plan, by comparing the accident phase closure value with the closure threshold, two accident evidence records that can be closed into the same accident event are written into the same accident event group, and the main phase record and stage supplementary record are determined, so that the merging of accident chains has a clear record linking basis. When the accident phase closure value is insufficient, the accident evidence records are not directly merged. Instead, the independent accident candidate number is retained to eliminate the interference of the upload time on the accident chain sorting. The actual phase difference is regenerated using the report time, damage assessment time, repair start time, repair completion time, re-inspection time, and inspection time. The source of field misuse is identified by the upload time difference stability, stage reverse order relationship, and accident chain standard phase table. Then, the field source is corrected, the business link phase number is reconfigured, the document number serial relationship is checked back, and the evidence closure exponent is updated, so that the merging deviation caused by the misuse of upload time, stage order misalignment, and missing document chain can be back-calculated and corrected. If the accident phase closure value still does not meet the closure threshold after back-calculation, the two accident evidence records are split into different accident candidate groups, thereby avoiding the wrong merging of the same accident or the forced merging of different accidents.

[0047] Specifically, the steps for identifying duplicate counts using the maximum number of records for the same location, the number of accident evidence records, the risk value for each record, the risk value for deduplication of locations, and the lower limit of accident phase closure are as follows: Obtain the maximum number of records for the same location, the number of accident evidence records, the risk value for each record, the risk value for deduplication of locations, and the accident phase closure value within the accident event group. The maximum number of records for the same location indicates the degree to which multiple accident evidence records collectively describe the same standard location number. The number of accident evidence records indicates the scale of records participating in duplicate count identification within the accident event group. The risk value for each record indicates the risk accumulated item by item from accident reports, loss assessment reports, repair work orders, re-inspection reports, and testing reports. The quantity and location deduplication risk value represent the risk quantity retained after removing duplicate entries according to the standard location number. The accident phase closure value represents the degree of closure of record pairs within an accident event group at the same accident event level. The accident phase closure value is calculated by reading the accident phase closure value of all record pairs within the accident event group and taking the minimum value as the lower bound of accident phase closure, ensuring that the weakest closure relationship within the accident event group can participate in duplicate counting identification. The maximum number of records in the same location is subtracted from a constant, and then divided by the sum of the number of accident evidence records and the smallest positive number to obtain the same location congestion rate. The same location congestion rate is a dimensionless proportional quantity used to represent the degree of concentration of records under the same standard location number. The risk value for each entry is subtracted from the location deduplication risk value and then divided by the risk value for each entry. The sum of the risk value and the minimum positive number is used to obtain the recount risk reduction rate, a dimensionless proportional quantity used to represent the degree of risk reduction after switching from item-by-item counting to location deduplication. Subtracting the lower limit of accident phase closure from the constant 1, and then dividing by the sum of the lower limit of accident phase closure and the minimum positive number, yields the phase loosening rate, a non-negative dimensionless adjustment quantity used to represent the degree of suppression of duplicate count identification by the weakest closure relationship within an accident event group. Adding the constant 1, the phase loosening rate, and the minimum positive number yields the recount suppression term, which reduces the intensity of duplicate count identification in accident event groups with insufficient lower limit of accident phase closure, preventing the same location records in different accidents from being misjudged as the same damage fact. The sum of the congestion rate and the recount risk decline rate, divided by the recount suppression term, yields the recount pressure input. This input is used to simultaneously express the repeated counting trend caused by the aggregation of records at the same location and the decline in risk. Adding a constant to the recount pressure input and performing a natural logarithmic operation yields the same-damage recount value. This natural logarithmic operation is used to compress the sudden increase in the same-location congestion rate and the recount risk decline rate under extreme record quantities, enhancing the numerical stability of the same-damage recount value across different accident event groups. The recount threshold is determined statistically by the distribution of the same-damage recount values ​​in the repeated count event group and the non-repeated count event group in the historical accident event sample. It is used to distinguish the boundary between repeated counting of the same damage fact and the coexistence of normal multi-location records.

[0048] The specific formula for calculating the recalculation value of the same loss is as follows:

[0049] ;

[0050] In the formula, This represents the repeated damage value for the g-th accident event group, reflecting the degree to which the same damage fact is repeatedly included in risk aggregation; This represents the maximum number of records falling into the same standard location number within the g-th accident event group, reflecting the intensity of the concentrated description of the same damage location; This represents the number of accident evidence records within the g-th accident event group, reflecting the total number of records generated by accident reports, damage assessment reports, repair work orders, re-inspection reports, and test reports within the accident event group; This indicates that each item is included in the risk value, reflecting the cumulative amount obtained when each incident evidence record within the group is treated as an independent risk evidence. The deduplication risk value for a location reflects the cumulative amount obtained after deduplication according to the unified accident event number and standard location number. This represents the lower bound of the accident phase closure, reflecting the weakest homogeneous support within the g-th accident event group, and is derived from the calculation result of the accident phase closure value. This represents a logarithmic pressure mapping, used to compress the combined intensity formed by congestion and risk reduction at the same location; It represents the smallest positive number and reflects the basis for denominator correction.

[0051] This implementation scheme comprehensively characterizes the concentration of records under the same standard location number, the difference in risk between record-by-record and location deduplication, and the weakest closure relationship within the accident event group by reading the maximum number of records for the same location, the number of accident evidence records, the risk value for each record, the location deduplication risk value, and the accident phase closure value within the accident event group. By introducing a lower limit value for accident phase closure to constrain the reliability of merging, the congestion rate of the same location and the recalculation risk decline rate will only form an effective recalculation pressure input when the accident event group has sufficient closure support, thus avoiding misjudging records of the same location in different accidents as the same damage fact. By compressing the numerical surge caused by extreme record numbers through natural logarithmic operations, a stable recalculation value for the same damage is generated, providing a basis for subsequent judgments on whether to lock the main damage record, rewrite auxiliary evidence records, and update the location deduplication risk value, thereby reducing the duplicate inclusion of the same damage fact in risk aggregation.

[0052] Specifically, the steps for locking the main damage record and rewriting the auxiliary evidence record based on the duplicate count identification results are as follows: By comparing the duplicate count value and the duplicate count threshold in real time, when the duplicate count value is greater than or equal to the duplicate count threshold, the corresponding accident event group is marked as a duplicate count event group. The unified accident event number and standard location number of each accident evidence record in the group are read, and multiple accident evidence records falling into the same standard location number are screened. The screening results are used to establish a candidate sequence for the same location according to the business link phase number, the accident phase closure value, and the record risk value. For records under the same standard location number, the accident evidence record with the earlier business link phase number is retained as the main damage record, so that the record closer to the front end of the accident chain in the accident, loss assessment, repair, re-inspection, and testing is used as the risk entry point for the same damage fact. When the business link phase numbers are the same, the accident evidence record with the larger accident phase closure value is retained as the main damage record, and the remaining records are rewritten as auxiliary evidence records. The document number, associated document number, text description, image recognition result, and damage area recorded are used as evidence in the explanatory chain. Auxiliary evidence records are no longer used as independent risk entry points. For multiple accident evidence records under the same standard location number that have completed the locking of the main damage record, only the record risk value of the main damage record is used as the location risk entry point. The location deduplication risk value is updated according to the record risk value of the main damage record. The record-by-record risk value before the update, the updated location deduplication risk value, the main damage record, and the auxiliary evidence record are written into the duplicate count event group. This ensures that the same damage fact retains only one risk entry result in risk aggregation. The updated location deduplication risk value is used to enter the accident risk stabilization verification process. When the duplicate count value of the same damage is less than the duplicate count threshold, it is determined that no duplicate count has been formed in the current accident event group. The independent risk entry status of each accident evidence record in the current accident event group is retained, and it continues to participate in the subsequent accident risk stabilization verification according to the record risk value corresponding to the accident evidence record.

[0053] In this implementation plan, the system determines whether the same damage fact is repeatedly counted within an accident event group by comparing the repeated damage value with the repeated damage threshold. When the repeated damage value is greater than or equal to the repeated damage threshold, multiple accident evidence records under the same standard location number are grouped into the same damage fact range. The main damage record is determined by the business process phase number and the accident phase closure value. The remaining accident evidence records are rewritten as auxiliary evidence records, so that the auxiliary evidence records continue to play an interpretive role but are no longer used as independent risk entry points. By updating the location deduplication risk value only with the record risk value of the main damage record, multiple records under the same standard location number formed by accident reports, damage assessment reports, repair work orders, re-inspection reports, and testing reports no longer accumulate risk amounts repeatedly. When the repeated damage value is less than the repeated damage threshold, the independent risk entry status of each accident evidence record is retained. This achieves the identification of duplicate count event groups, the locking of main damage records, the retention of auxiliary evidence records, and the updating of location deduplication risk values, providing a basis for deduplication-based risk aggregation for subsequent accident risk stabilization verification.

[0054] Specifically, the steps for risk stabilization verification based on the average accident phase closure value, the average value of repeated damage calculations, the number of original abnormal records, the number of accident events after merging, and the risk values ​​of original records and the merged deduplication risk values ​​are as follows: Obtain the accident phase closure value, the average value of repeated damage calculations, the number of original abnormal records, the accident event number, and the risk value of original records for the target vehicle. The number of original abnormal records refers to the number of accident evidence records for the target vehicle that have already formed candidate accident numbers and have a risk value greater than zero before the accident events are merged; this is used to characterize the scale of records included in the accident risk verification before merging. The accident phase closure value and the average damage calculation value are both derived from the preliminary calculation results of each accident event group under the target vehicle. The risk value of the records is evaluated before entering this step. Maximum and minimum normalization are applied to the range of zero to one. The accident event numbers under the target vehicle are deduplicated to obtain the number of merged accident events, representing the scale of the independent accident events retained after merging. The maximum recorded risk value is taken for the same accident event number and the same standard part number under the target vehicle, and then all maximum recorded risk values ​​are summed to obtain the merged and deduplicated risk value, representing the risk amount after the same standard part number within the same accident event is only included once. The lower bound of the accident phase closure and the recalculation value of the same damage are averaged for all accident event groups under the target vehicle to obtain the accident phase closure mean and the recalculation mean of the same damage. The calculation window for the accident phase closure mean is limited to the target vehicle. All accident event groups with generated accident event numbers under the target vehicle are used to characterize the overall closed stability of the accident evidence chain for the target vehicle. The average of repeated damage is used to characterize the average intensity of the same damage fact being repeatedly counted in each accident event group of the target vehicle. The event compression rate is obtained by subtracting the number of merged accident events from the number of original abnormal records and then dividing by the sum of the number of original abnormal records and the smallest positive number, and the event compression rate is limited to a dimensionless proportional quantity. The event stabilization risk reduction rate is obtained by subtracting the merged and deduplicated risk value from the risk value of the original records and then dividing by the sum of the risk value of the original records and the smallest positive number, and the event stabilization risk reduction rate is limited to a dimensionless proportional quantity. The event compression residual term is obtained by squaring the event compression rate minus the average of repeated damage. The event compression rate is used to characterize the deviation between the degree of event compression and the intensity of repeated inclusion. The risk reduction residual term is obtained by subtracting the average of repeated losses from the event compression risk reduction rate and then squaring it. This residual term characterizes the deviation between the degree of risk reduction and the intensity of repeated inclusion. The event compression residual term and the risk reduction residual term are negatively evaluated and then exponentially calculated. This exponential calculation attenuates the residual deviation, ensuring that the closer the event compression rate and the risk reduction rate are to the average of repeated losses, the closer the output is to a stable state. Larger deviations result in a faster decrease in the output. Multiplying this by the event phase closure mean yields the event risk stabilization value. This ensures that the event risk stabilization value is simultaneously constrained by the stability of the event evidence chain closure, the rationality of event compression, and the rationality of risk reduction. Figure 4As shown, this is a three-dimensional response surface for accident risk stabilization based on phase closure merging and duplicate count suppression provided in this application embodiment. The X-axis is the event compression rate, which is the ratio of the difference between the original number of abnormal records and the number of accident events after merging to the original number of abnormal records, reflecting the compression capability of the accident phase merging module for redundant event records. The Y-axis is the stabilization risk reduction rate, which is the ratio of the difference between the original record risk value and the merged deduplication risk value to the original record risk value, reflecting the elimination strength of the duplicate damage suppression module for repeated damage risks. The Z-axis is the accident risk stabilization value, which is jointly determined by the accident phase closure mean and the Gaussian exponential term, characterizing the degree to which the risk score returns to the stable state of the real accident chain after the vehicle completes event merging and duplicate count suppression. The curved surface presents a single-peak ridge shape: when the event compression rate and the stabilization risk decline rate are both at a moderate level and match each other, the accident risk stabilization value reaches the peak region, indicating that the merging effect and deduplication effect are optimal at this time, and the risk output is closest to the actual accident damage chain; after deviating from this moderate matching region, the accident risk stabilization value smoothly decays to an extremely low level, indicating that excessive compression or excessive deduplication will destroy the stability of risk statistics, providing an intuitive numerical basis for setting the stabilization threshold, optimizing module linkage, and diagnosing insufficient risk stabilization samples.

[0055] The specific formula for calculating the accident risk stabilization value is as follows:

[0056] ;

[0057] In the formula, This represents the accident risk stabilization value of the target vehicle, reflecting whether the risk outcome returns to a stable state consistent with the actual accident chain after the target vehicle completes accident event merging and duplicate count suppression. This indicates the number of original abnormal records, reflecting the total number of abnormal records generated by the target vehicle before consolidation, including accident reports, damage assessment reports, repair work orders, re-inspection reports, and test reports. This indicates the number of incidents after merging, reflecting the total number of incidents formed after the target vehicle has undergone incident phase closure merging; It represents the average value of repeated damage calculation, reflecting the average intensity of the same damage fact being repeatedly counted in each accident event group of the target vehicle. It is derived from the calculation result of the repeated damage value. This represents the risk value of the original records, reflecting the cumulative number of vehicles obtained by scoring each original abnormal record; This represents the merged and deduplicated risk value, reflecting the cumulative number of vehicles obtained after deduplication according to the unified accident event number and standard part number; It represents the accident phase closure mean, reflecting the average closure stability of multi-source accident evidence for the target vehicle at the same accident event level, and is derived from the accident phase closure value calculation results; It represents the smallest positive number and reflects the basis for denominator correction.

[0058] In this implementation plan, the accident event numbers, standard part numbers, recorded risk values, accident phase closure values, and repeated damage values ​​under the target vehicle are merged and statistically analyzed to form the number of merged accident events, merged deduplication risk values, accident phase closure mean, and repeated damage mean. This allows the accident event merging results, the risk deduplication results under the same standard part number, and the stability of the accident evidence chain closure to be uniformly verified at the same vehicle dimension. The event compression rate measures the degree of convergence of the original abnormal record number to the number of merged accident events, and the stabilization risk decline rate measures the degree of decline of the original recorded risk value to the merged deduplication risk value. The residuals of both are calculated with the repeated damage mean, so that the accident event compression magnitude and risk decline magnitude can correspond to the intensity of repeated inclusion. Then, the residual deviation is attenuated and penalized using exponential calculation, and the accident risk stabilization value is obtained by combining the accident phase closure mean. This makes the risk stabilization verification not only focus on the reduction of the number of records and the decrease of risk, but also constrain whether the accident evidence chain has closure support, thus providing a quantitative basis for judging whether the accident risk results after merging and deduplication are stable and usable.

[0059] Specifically, the steps for marking insufficient stabilization, backcalculation verification, and writing reliable results based on the risk stabilization verification results are as follows: By comparing the accident risk stabilization value and the stabilization threshold in real time, when the accident risk stabilization value is less than the stabilization threshold, the target vehicle is marked as an insufficient stabilization record. The current unified accident event number, main phase record, main damage record, auxiliary evidence record, and merged deduplication risk value are latched, and phase closure backcalculation instructions and similar damage recalculation verification instructions are generated respectively. Among them, the phase closure backcalculation instruction is used to call the evidence closure exponent, upload offset rate, actual phase difference, and standard phase difference to recalculate the accident phase closure value, and re-execute accident event merging, upload time exclusion, and stage phase rearrangement based on the closure threshold. The similar damage recalculation verification instruction is used to call the maximum number of records of the same part, the number of accident evidence records, the risk value to be included one by one, the part deduplication risk value, and the lower bound of accident phase closure. The system recalculates the similar damage recalculation value and re-executes the main damage record locking, auxiliary evidence record rewriting, and location deduplication risk value update based on the recalculation threshold. After the accident phase closure value, similar damage recalculation value, number of merged accident events, merged deduplication risk value, accident phase closure mean, and similar damage recalculation mean are updated, the accident risk stabilization value is recalculated. When the recalculated accident risk stabilization value is still less than the stabilization threshold, the target vehicle is written into the accident risk stabilization pending calibration table, retaining the conflict accident event group, anomaly source field, stabilization risk decline rate, accident risk stabilization value, event compression rate, and pause output flag. When the accident risk stabilization value is greater than or equal to the stabilization threshold, the target vehicle is marked as an accident risk stabilization passed record, retaining the unified accident event number, main phase record, main damage record, auxiliary evidence record, and merged deduplication risk value, and is written into the vehicle accident evidence database.

[0060] In this implementation plan, the final usability judgment is made on the results of the target vehicle's accident event merging, same-damage recalculation suppression, and merged deduplication risk value update by comparing the accident risk stabilization value and the stabilization threshold. When the accident risk stabilization value is less than the stabilization threshold, the current merged deduplication risk value is not directly output. Instead, the current unified accident event number, main phase record, main damage record, auxiliary evidence record, and merged deduplication risk value are latched. The accident phase closure value is re-verified through the phase closure back-calculation instruction, and the same-damage recalculation value is re-verified through the same-damage recalculation review instruction. This ensures that the number of merged accident events, merged deduplication risk value, accident phase closure average, and same-damage recalculation average can be re-verified after the update. The new vehicle is added to the accident risk stabilization value calculation. When the recalculated accident risk stabilization value is still less than the stabilization threshold, the target vehicle is written into the accident risk stabilization pending calibration table. Only the conflict accident event group, anomaly source field, stabilization risk reduction rate, accident risk stabilization value, event compression rate, and pause output mark are retained to prevent unstable results from entering the reliable output. When the accident risk stabilization value is greater than or equal to the stabilization threshold, the target vehicle is marked as an accident risk stabilization pass record. The unified accident event number, main phase record, main damage record, auxiliary evidence record, and merged deduplication risk value are retained to achieve closed-loop confirmation of the vehicle accident data merging results, duplicate count suppression results, and risk stabilization results.

[0061] Specifically, this embodiment provides a vehicle data trust verification system based on multimodal alignment, applied to a vehicle data trust verification method based on multimodal alignment. It includes a vehicle accident data acquisition and preprocessing module, used to acquire multimodal vehicle accident data and perform preprocessing operations on the data. Specifically, it obtains vehicle identification numbers, document numbers, associated document numbers, business time, upload time, part names, text descriptions, and image recognition results from accident reports, damage assessment reports, repair work orders, re-inspection reports, and test reports. It also extracts part and damage terms from the text descriptions and processes the image recognition results. The system extracts the image part identification results and damage areas, and uses the vehicle identification code, document number concatenation relationship, standard part number, and business process phase number as the basis for multimodal alignment to form an accident evidence record. Simultaneously, a rule engine obtains the record risk value of the accident evidence record based on the accident handling process type, standard part number, damage words, damage area, and document number concatenation relationship. The accident chain phase merging analysis module is used to perform accident event merging judgment based on the evidence closure exponent, upload offset rate, actual phase difference, and standard phase difference. Based on the merging judgment result, it performs accident event merging and stage phase re-merging. The ranking mechanism, specifically, compares the accident phase closure value with the closure threshold to determine whether two accident evidence records are written into the same accident event group. When the accident phase closure value is insufficient, it stops using the upload time as the basis for accident chain sorting and instead uses the reporting time, damage assessment time, repair start time, repair completion time, re-inspection time, and testing time to regenerate the actual phase difference. It then corrects the field source using the accident chain standard phase table, updates the evidence closure exponent, and re-executes the merging judgment. The same-damage recount suppression identification module is used to determine whether two accident evidence records are written into the same accident event group by comparing the maximum number of records for the same location, the number of accident evidence records, and the number of records entered into the risk assessment. The system identifies duplicate counts based on risk values, location deduplication risk values, and the lower limit of accident phase closure. Based on the duplicate count identification results, it locks the main damage record and rewrites the auxiliary evidence record. Specifically, when the duplicate count value of the same damage is greater than or equal to the duplicate count threshold, it marks the corresponding accident event group as the duplicate count event group, filters multiple accident evidence records under the same standard location number, determines the main damage record based on the business process phase number and the accident phase closure value, rewrites the remaining accident evidence records as auxiliary evidence records, and uses only the record risk value of the main damage record as the location risk entry point to update the location deduplication risk value.The accident risk stabilization verification module is used to perform risk stabilization verification by comprehensively considering the accident phase closure average, the average of similar damage recalculation, the number of original abnormal records, the number of merged accident events, the risk value of the original records, and the merged deduplication risk value. Based on the risk stabilization verification results, it performs insufficient stabilization marking, backcalculation verification, and writing of reliable results. Specifically, when the accident risk stabilization value is less than the stabilization threshold, it latches the current unified accident event number, main phase record, main damage record, auxiliary evidence record, and merged deduplication risk value, and generates phase closure backcalculation instructions and similar damage recalculation verification instructions. When the accident risk stabilization value is greater than or equal to the stabilization threshold, it marks the target vehicle as having passed the accident risk stabilization record and writes the unified accident event number, main phase record, main damage record, auxiliary evidence record, and merged deduplication risk value into the vehicle accident evidence database.

[0062] In this implementation plan, through the collaborative processing of the vehicle accident data acquisition and preprocessing module, the accident chain phase merging analysis module, the same-damage recount suppression and identification module, and the accident risk stabilization verification module, multimodal vehicle accident data from accident reports, damage assessment reports, repair work orders, re-inspection reports, and test reports can first form a unified accident evidence record. Then, based on the accident phase closure value, the same accident event is merged and judged, avoiding the same accident being split into multiple accident candidate groups. By identifying duplicate counts under the same standard part number through the same-damage recount value, the main damage record is locked and the auxiliary evidence record is retained, so that the same damage fact is only retained once for risk entry. By using the accident risk stabilization value, the accident phase closure mean, the same-damage recount mean, the number of original abnormal records, the number of merged accident events, the risk value of the original record, and the merged deduplication risk value are comprehensively verified. This ensures that the merged and deduplicated results can be confirmed for usability before being written into the vehicle accident evidence database, thereby improving the stability, traceability, and interpretability consistency of the vehicle accident data credibility verification results.

[0063] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0064] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A vehicle data reliability verification method based on multimodal alignment, characterized in that, Includes the following steps: S1, collect multimodal vehicle accident data and perform preprocessing operations on the multimodal vehicle accident data; S2, based on the evidence closure exponent, upload offset, actual phase difference and standard phase difference, performs incident event merging judgment, and performs incident event merging and stage phase rearrangement according to the merging judgment result; The specific steps for incident event merging and judgment based on evidence closure exponent, upload offset, actual phase difference and standard phase difference are as follows: Obtain the evidence closure exponent, upload offset, actual phase difference, and standard phase difference between accident evidence record a and accident evidence record b; Subtract the standard phase difference from the actual phase difference to obtain the phase difference quantity; perform a cosine operation on the phase difference quantity, and subtract the cosine operation result from the constant to obtain the phase cosine offset term; add the constant to the evidence closure exponent to obtain the phase closure support term; add the constant, the evidence closure exponent, the upload offset rate, the phase cosine offset term, and the minimum positive number to obtain the phase closure constraint term. Divide the phase closure support term by the phase closure constraint term to obtain the accident phase closure value; The specific steps for performing event merging and phase reordering based on the merging judgment result are as follows: When a new accident evidence record is received, the accident phase closure value is compared with the closure threshold. When the accident phase closure value is greater than or equal to the closure threshold, the two accident evidence records are merged and written into the same accident event group with a unified accident event number. The accident evidence record with the larger evidence closure exponent value is used as the main phase record, and the other accident evidence record is used as a stage supplementary record. When the accident phase closure value is less than the closure threshold, the independent accident candidate numbers of the two accident evidence records are retained. The upload time is no longer used as the basis for sorting the accident chain. Instead, the actual phase difference is regenerated using the report time, damage assessment time, repair start time, repair completion time, re-inspection time, and inspection time. The business stage order is corrected according to the standard phase table of the accident chain. The field source of accident evidence records where the upload time is mistakenly taken as the business occurrence time, the repair completion time is mistakenly taken as the re-inspection time, and the inspection upload time is mistakenly taken as the inspection time is corrected to perform phase rearrangement of the execution stage. The document number serial relationship is checked back by vehicle identification number, document number, and associated document number to update the evidence closure exponent and recalculate the accident phase closure value. If the newly obtained accident phase closure value is still less than the closure threshold, the two accident evidence records will be split into different accident candidate groups. S3 identifies duplicate counts by using the maximum number of records for the same location, the number of accident evidence records, the risk value for each record, the risk value for deduplication of locations, and the lower limit of accident phase closure. Based on the duplicate count identification results, it performs main damage record locking and auxiliary evidence record rewriting. S4 performs risk stabilization verification based on the average of the phase closure of the comprehensive accident, the average of the recalculation of the same loss, the number of original abnormal records, the number of accident events after merging, the risk value of the original record and the risk value of the merged deduplication. Based on the risk stabilization verification results, it performs insufficient stabilization marking, backcalculation verification and reliable result writing.

2. The vehicle data reliability verification method based on multimodal alignment according to claim 1, characterized in that: The specific steps for collecting multimodal vehicle accident data are as follows: Collect multimodal vehicle accident data: Obtain vehicle identification number, document number, associated document number, transaction time, upload time, part name, text description, and image recognition results from accident reports, damage assessment reports, repair work orders, re-inspection reports, and inspection reports. Among them, the transaction time includes the time of reporting the accident, the time of damage assessment, the time of commencement of repair, the time of completion of repair, the time of re-inspection, and the time of inspection. The upload time indicates the time when the accident evidence record enters the platform. The image recognition results include the image part recognition results and the damaged area. The record risk value of the accident evidence record is obtained through the rule engine based on the accident handling stage type, part name, text description, image recognition results, and document number concatenation status.

3. The vehicle data reliability verification method based on multimodal alignment according to claim 1, characterized in that: The specific steps for preprocessing multimodal vehicle accident data are as follows: The system performs format cleaning on vehicle identification numbers, document numbers, and associated document numbers, and establishes a serial relationship between document numbers based on these numbers. It extracts location and damage terms from text descriptions, extracts image location recognition results and damage areas from image recognition results, maps location names, location terms, and image location recognition results to standard location numbers, and determines the standard location number matching relationship. Based on reporting, damage assessment, repair, re-inspection, and testing, it constructs a standard phase table for the accident chain, assigning a business process phase number to each accident evidence record. The system uses the serial relationship between vehicle identification numbers and document numbers as the record association benchmark, the business process phase number as the time alignment benchmark, and the standard location number as the reference. Using the digit number as the alignment benchmark, multimodal alignment processing is performed on the extracted text description, image recognition results, vehicle identification number, document number, associated document number, business time, upload time, and location name to obtain the aligned accident evidence record. An evidence closure exponent is generated by combining the vehicle identification number consistency relationship, document number concatenation relationship, and standard location number hit relationship. Based on the business time difference, business process phase number, and upload time difference between two accident evidence records, the actual phase difference, standard phase difference, and upload offset rate are generated respectively. Using the accident candidate number as the initial statistical key, the accident evidence records within the accident candidate group are counted to obtain the number of accident evidence records. The risk values ​​of each piece of accident evidence recorded in the accident candidate group are summed to obtain the risk value for each piece of evidence. The original recorded risk value is obtained by summing up the recorded risk values ​​of all accident evidence records under the target vehicle. For records with the same standard part number within the same accident candidate group, the maximum risk value is taken, and then the maximum risk values ​​of records corresponding to each standard part number are summed to obtain the part deduplication risk value; using the target vehicle as the statistical key, all accident evidence records before merging are counted to obtain the number of original abnormal records; the maximum-minimum normalization method is used to normalize the multimodal vehicle accident data, and a vehicle accident evidence database is constructed after storage.

4. The vehicle data reliability verification method based on multimodal alignment according to claim 1, characterized in that: The specific steps for identifying duplicate counts by using the maximum number of records for the same location, the number of accident evidence records, the risk value for each record, the location deduplication risk value, and the lower limit of the accident phase closure are as follows: Obtain the maximum number of records for the same location within the accident event group, the number of accident evidence records, the risk value for each record, the risk value for deduplication of locations, and the accident phase closure value; The accident phase closure value of all records within the accident event group is read, and the minimum value is taken as the lower limit of the accident phase closure value; the maximum number of records in the same location is subtracted from a constant one, and then divided by the sum of the number of accident evidence records and the smallest positive number to obtain the congestion rate of the same location; the risk value included in each item is subtracted from the location deduplication risk value, and then divided by the sum of the risk value included in each item and the smallest positive number to obtain the recalculated risk reduction rate; Subtract the lower limit of the accident phase closure from the constant 1, and then divide by the sum of the lower limit of the accident phase closure and the minimum positive number to obtain the phase loosening rate; add the constant 1, the phase loosening rate, and the minimum positive number to obtain the recalculation suppression term; add the same location congestion rate and the recalculation risk reduction rate, and then divide by the recalculation suppression term to obtain the recalculation pressure input; add the constant 1 and the recalculation pressure input, and then perform a natural logarithmic operation to obtain the same loss recalculation value.

5. The vehicle data reliability verification method based on multimodal alignment according to claim 1, characterized in that: The specific steps for locking the primary injury record and rewriting the auxiliary evidence record based on the duplicate count identification results are as follows: By comparing the repeated damage value with the repeated damage threshold in real time, when the repeated damage value is greater than or equal to the repeated damage threshold, the corresponding accident event group is marked as a duplicate count event group. The unified accident event number and standard location number of each accident evidence record in the group are read, and multiple accident evidence records falling into the same standard location number are filtered out. For records under the same standard location number, the accident evidence record with the earlier business process phase number is retained as the main damage record. When the business process phase numbers are the same, the accident evidence record with the larger accident phase closure value is retained as the main damage record, and the remaining records are rewritten as auxiliary evidence records. For multiple accident evidence records under the same standard location number that have completed the locking of the main damage record, only the record risk value of the main damage record is used as the location risk entry point. The location deduplication risk value is updated again according to the record risk value of the main damage record, and the updated location deduplication risk value is used to enter the accident risk stabilization verification process. When the repeated count of the same loss is less than the repeated count threshold, it is determined that no duplicate count has been formed within the current accident event group, and the independent risk inclusion status of each accident evidence record within the current accident event group is retained.

6. The vehicle data reliability verification method based on multimodal alignment according to claim 1, characterized in that: The specific steps for risk stabilization verification using the comprehensive accident phase closure mean, the same loss recount mean, the number of original abnormal records, the number of merged accident events, and the risk values ​​of the original records and the merged deduplication risk values ​​are as follows: Obtain the accident phase closure value, similar damage recount value, number of original abnormal records, accident event number, and risk value of the original record for the target vehicle; perform deduplication and counting on the accident event numbers under the target vehicle to obtain the number of merged accident events; For the target vehicle, take the maximum value of the recorded risk value corresponding to the same accident event number and the same standard part number, and then sum all the maximum recorded risk values ​​to obtain the merged and deduplicated risk value; The average values ​​of the lower boundary of the accident phase closure and the recalculation value of the same damage are calculated for all accident event groups under the target vehicle to obtain the average value of the accident phase closure and the average value of the recalculation of the same damage. The event compression rate is obtained by subtracting the number of merged incidents from the number of original abnormal records and then dividing by the sum of the number of original abnormal records and the smallest positive number. The event stabilization risk reduction rate is obtained by subtracting the merged and deduplicated risk value from the risk value of the original records and then dividing by the sum of the risk value of the original records and the smallest positive number. The event compression residual term is obtained by squared the sum of the event compression rate and the average of the repeated losses. The risk decline residual term is obtained by subtracting the average of the same loss from the risk decline rate and then squaring it. The event compression residual and the risk fallback residual are negativeized and then exponentially calculated. Finally, the result is multiplied by the accident phase closure mean to obtain the accident risk stabilization value.

7. The vehicle data reliability verification method based on multimodal alignment according to claim 1, characterized in that: The specific steps for performing insufficient stabilization marking, backcalculation verification, and writing of reliable results based on the risk stabilization verification results are as follows: By comparing the accident risk recovery value and the recovery threshold in real time, when the accident risk recovery value is less than the recovery threshold, the target vehicle is marked as a record of insufficient recovery. The current unified accident event number, main phase record, main damage record, auxiliary evidence record and merged deduplication risk value are locked. After generating phase closure back calculation instruction and same damage recalculation verification instruction respectively, the accident risk recovery value is recalculated. When the recalculated accident risk stabilization value is still less than the stabilization threshold, the target vehicle is written into the accident risk stabilization pending calibration table, retaining the conflict accident event group, anomaly source field, stabilization risk decline rate, accident risk stabilization value, event compression rate, and pause output flag. When the accident risk recovery value is greater than or equal to the recovery threshold, the target vehicle is marked as having passed the accident risk recovery record. The unified accident event number, main phase record, main damage record, auxiliary evidence record, and merged deduplicated risk value are retained and written into the vehicle accident evidence database.

8. A vehicle data trust verification system based on multimodal alignment, applied to the vehicle data trust verification method based on multimodal alignment as described in any one of claims 1-7, characterized in that, include: The vehicle accident data acquisition and preprocessing module is used to acquire multimodal vehicle accident data and perform preprocessing operations on the multimodal vehicle accident data. The accident chain phase merging analysis module is used to perform accident event merging judgment based on evidence closure exponent, upload offset rate, actual phase difference and standard phase difference, and to perform accident event merging and stage phase rearrangement according to the merging judgment result. The duplicate count suppression and identification module is used to identify duplicate counts by using the maximum number of records for the same location, the number of accident evidence records, the risk value for each record, the risk value for deduplication of the location, and the lower limit of the accident phase closure. Based on the duplicate count identification results, the main damage record is locked and the auxiliary evidence record is rewritten. The accident risk stabilization verification module is used to perform risk stabilization verification by comprehensively considering the accident phase closure mean, the same loss recalculation mean, the number of original abnormal records, the number of merged accident events, the risk value of the original records and the merged deduplication risk value. Based on the risk stabilization verification results, it performs stabilization insufficiency marking, backcalculation verification and reliable result writing.

Images