A method and system for hierarchical data transfer in auto insurance claims

By separating the trusted domain and the abnormal conflict domain in the link modeling, the problems of unverifiable source and single-source anomalies in the process of transferring vehicle insurance claim image materials are solved, and interpretable review and transfer are realized, reducing false triggering and error migration.

CN123071091APending Publication Date: 2026-11-10QILU UNIVERSITY OF TECHNOLOGY (SHANDONG ACADEMY OF SCIENCES) +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611576406.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-13
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In the current process of transferring auto insurance claim image materials, it is difficult to distinguish between unverifiable sources, single-source anomalies, and cross-evidence conflicts, which leads to distortion of transfer thresholds, wrong release or wrong interception, and lacks an effective risk signal separation calculation and review transfer mechanism.

Method used

The system adopts a model that separates the link credibility domain from the anomaly conflict domain. By generating standardized evidence items, it performs anomaly attribution prediction, calculates the link credibility score, anomaly conflict score, uncertainty score, and cross-evidence conflict degree, performs cascading early stop judgment, and outputs an interpretable audit trajectory.

Benefits of technology

It achieves separate calculation of source credibility, single-source anomalies, and cross-evidence conflicts, reduces system false triggers, prevents abnormal signals from being scored repeatedly, avoids the migration of audit actions erroneously, and forms an interpretable governance closed loop.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123071091A_ABST
    Figure CN123071091A_ABST
Patent Text Reader

Abstract

This invention provides a method and system for hierarchical processing of vehicle insurance claim images, belonging to the field of vehicle insurance claim image material processing technology. This invention parses metadata, recoded features, acquisition logs, uploaded messages, task identifiers, device fingerprints, and business fields from multi-source material nodes in claim cases to generate standardized evidence items containing abnormal signal identifiers and abnormal attribution identifiers. Abnormal signals are first pre-judged as either single-source abnormal attribution or cross-evidence conflict attribution. Then, a link credibility score C, an abnormal conflict score R, an uncertainty score U, and a cross-evidence conflict degree K are calculated respectively, ensuring that R only filters and includes single-source abnormal signals, K only filters and includes cross-evidence conflict signals, and the same abnormal signal is mutually exclusive in the calculation. Based on U and K, cascaded early stop is executed, and hierarchical processing actions are output based on the partial order of actions, generating an interpretable review trajectory, thereby reducing duplicate scoring and misjudgment and improving the stability of the review process.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of image processing technology for auto insurance claims, specifically to a method and system for hierarchical transfer of auto insurance claim images. Background Technology

[0002] Auto insurance claim documentation typically includes photos of vehicle damage, on-site photos, photos of the repair process, scanned copies of invoices, rescue certificates, and supplementary explanatory images. These materials may originate from controlled applications, web pages, mini-programs, customer photo albums, bulk uploads from repair shops, forwardings or screenshots on social media, etc. The evidentiary value and verifiability of materials vary significantly depending on the source.

[0003] Existing claims image processing technologies typically categorize, group, and display images based on visual content or file type, but they fail to simultaneously verify the continuity of nodes and the correspondence of credentials between the acquisition end, upload end, and platform receiving end. When images are imported from albums, forwarded through social media, screenshotted, or batch-uploaded by repair shops, the original acquisition metadata and upload logs may be stripped away. Existing technologies may still subject these images to the same rules for automatic review as controlled acquisition images, making it difficult to distinguish between materials with identifiable content but unverifiable sources and materials with complete sources but abnormal content. This can easily lead to distortion of the flow threshold and false release or interception.

[0004] Existing image tampering detection or image source tracing technologies primarily target single-image on-the-ground detection of recoding, splicing, watermarking, or vouchers. Legitimate compression, format conversion, and transfer via social media may trigger similar characteristics to tampering. However, inconsistencies between images and accident reports, damage assessments, repairs, invoices, or historical cases cannot be identified solely through single-image detection. Current auto insurance image risk control technologies mostly output risk clues or risk reports, lacking a mechanism to quantify and convert unverifiable sources, single-source anomalies, insufficient evidence, and cross-evidence conflicts into separate review processes. This leads to the same anomaly signal potentially being scored repeatedly, and the review process may shift towards a more lenient direction even after supplementary materials are provided, before the risk has been eliminated.

[0005] Therefore, how to separately calculate the credibility of the source link, the degree of single-source anomaly, the degree of insufficient evidence, and the degree of cross-evidence conflict during the review process after the claims platform receives the images, and how to use cascading early stop, mutual exclusion of abnormal signals, partial order constraints of actions, and interpretable review trajectory stability control to control the subsequent processing flow, is a technical problem that needs to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for hierarchical transfer of vehicle insurance claim images, so as to solve at least one of the technical problems existing in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for hierarchical transfer of vehicle insurance claim images, comprising: Obtain a collection of multi-source material nodes for auto insurance claims that have been received by the auto insurance claims platform and entered the review and processing stage; For the data information of the multi-source material node set and the received analytical image material, a standardized evidence item containing anomaly signal identifier and anomaly attribution identifier is generated; wherein, each standardized evidence item includes at least evidence dimension, evidence direction, evidence strength, source reliability discount factor, associated material node, anomaly signal identifier and anomaly attribution identifier; Perform anomaly attribution prediction on the anomalous signals in the standardized evidence item, predicting the anomalous signals as single-source anomaly attribution or cross-evidence conflict attribution, and backfill the anomaly attribution prediction results into the standardized evidence item; Based on the anomaly attribution prediction results, the multi-source material node set is mapped to the link credibility domain, anomaly conflict domain, uncertainty domain, and cross-evidence conflict domain, respectively, and the link credibility score C, anomaly conflict score R, uncertainty score U, and cross-evidence conflict degree K are calculated respectively; and the anomaly conflict score side and the cross-evidence conflict degree side are mutually exclusive in the calculation of the same anomaly signal. Based on the uncertainty score U and the cross-evidence conflict degree K, a cascaded early stop judgment is performed, and the material node and triggering feature that trigger the early stop are recorded; For image material nodes that have not triggered the cascaded early stop judgment, perform at least one of image evidence detection, business field consistency verification, cross-case reuse detection or subject historical risk verification called when the preset reuse trigger condition is met, and update the detection or verification results to the abnormal conflict domain, the uncertainty domain or the cross evidence conflict domain. Based on the combination of the link credibility score C, abnormal conflict score R, uncertainty score U, and cross-evidence conflict degree K, hierarchical flow actions are output according to a preset action set. Generate an interpretable audit trajectory corresponding to the hierarchical flow action, including mutually exclusive inclusion results and action partial order level.

[0008] Secondly, the present invention provides a vehicle insurance claims image hierarchical transfer system, including an image and business data access module, an evidence item standardization module, an anomaly attribution prediction module, a four-domain calculation module, a cascaded early stop module, a detection and update module, a hierarchical transfer decision module, and an interpretation and traceability module. The image and business data access module is used to acquire a set of multi-source material nodes for auto insurance claims cases that have been received by the auto insurance claims platform and entered the review and processing stage; The evidence item standardization module is used to generate standardized evidence items containing abnormal signal identifiers and abnormal attribution identifiers for the data information of the multi-source material node set and the received analytical image material. The anomaly attribution prediction module is used to perform anomaly attribution prediction on the anomaly signals in the standardized evidence item, predict the anomaly signals as single-source anomaly attribution or cross-evidence conflict attribution, and backfill the anomaly attribution prediction result to the standardized evidence item. The four-domain calculation module is used to map the multi-source material node set to the link credibility domain, the anomaly conflict domain, the uncertainty domain, and the cross-evidence conflict domain based on the anomaly attribution prediction result, and to calculate the link credibility score C, the anomaly conflict score R, the uncertainty score U, and the cross-evidence conflict degree K respectively; and to make the anomaly conflict score side and the cross-evidence conflict degree side mutually exclusive in the calculation of the same anomaly signal. The cascaded early stop module is used to perform cascaded early stop judgment based on the uncertainty score U and the cross-evidence conflict degree K, and record the material node and triggering feature that triggers the early stop, and the cascaded early stop is not triggered by the abnormal conflict score alone; The detection and update module is used to perform at least one of the following on image material nodes that have not triggered cascaded early stop: image evidence detection, business field consistency verification, cross-case reuse detection or subject historical risk verification, and update the detection results to the abnormal conflict domain, uncertainty domain or cross evidence conflict domain. The hierarchical transfer decision module is used to output hierarchical transfer actions according to a preset action set based on the combination of the link credibility score C, abnormal conflict score R, uncertainty score U and cross-evidence conflict degree K. The explanation and traceability module is used to generate an explainable audit trajectory corresponding to the hierarchical flow action, which includes mutually exclusive inclusion results and action partial order level.

[0009] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the vehicle insurance claim image hierarchical transfer method based on the separation modeling of link trusted domain and abnormal conflict domain as described in the first aspect.

[0010] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the vehicle insurance claim image hierarchical transfer method based on link trusted domain and abnormal conflict domain separation modeling as described in the first aspect.

[0011] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the vehicle insurance claim image hierarchical transfer method based on the separation modeling of link trusted domain and abnormal conflict domain as described in the first aspect.

[0012] The beneficial effects of this invention are as follows: it separates the modeling of source credibility and content anomalies, reducing system false triggers caused by mixed scoring of missing links, single-source anomalies, and cross-evidence conflicts; it prevents the same anomaly signal from being repeatedly weighted in R and K through anomaly attribution prediction; it preserves uncertain states with low anomalies but insufficient evidence; it uses U and K as the main triggers for early termination instead of R alone; it avoids the erroneous migration of review actions to a more lenient direction through action partial order levels; it forms an interpretable governance closed loop by including anomaly signal identifiers, mutually exclusive inclusion results, and action partial order levels; and it reduces unnecessary calls to high-computing-power image forensics models, cross-modal comparison models, or manual verification queues when evidence is insufficient or cross-evidence conflicts are sufficient to determine the flow direction.

[0013] The advantages of additional aspects of the invention will be more apparent in the following description or may be learned by practice of the invention. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a functional framework diagram of the vehicle insurance claims image hierarchical transfer system described in an embodiment of the present invention.

[0016] Figure 2 This is a diagram illustrating the overall architecture of the vehicle insurance claims image hierarchical transfer method provided in this embodiment of the invention.

[0017] Figure 3 This is a flowchart of the vehicle insurance claim image hierarchical transfer method according to an embodiment of the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0020] After the auto insurance claim image materials have been received by the auto insurance claim platform and entered the review and transfer stage, how to uniformly convert the image source link, image evidence results, business field consistency, cross-case reuse, and subject historical risk into computable evidence items, and differentiate between source link breakage, single-source anomaly, insufficient evidence, and cross-evidence conflict without directly equating non-native images with fraud, and output an interpretable, traceable, and feedback-updable hierarchical transfer status that does not migrate to a more lenient direction? To solve the above technical problems, this invention provides a hierarchical transfer method and system for auto insurance claim images based on the separation of link credibility domain and anomaly conflict domain modeling. This scheme sets anomaly signal identifier and anomaly attribution identifier in the standardized evidence items. First, it pre-judges the single-source anomaly attribution or cross-evidence conflict attribution of the anomaly signal, and then maps them to the link credibility domain, anomaly conflict domain, uncertainty domain, and cross-evidence conflict domain, respectively. C, R, U, and K are calculated respectively, and cascading early stop is performed based on U and K, and transfer action constraints are performed based on the action partial order level L(a), finally outputting an interpretable review trajectory.

[0021] like Figure 1 As shown, in this embodiment, a vehicle insurance claims image hierarchical transfer system is first provided. The system includes an image and business data access module, an evidence item standardization module, an anomaly attribution prediction module, a four-domain calculation module, a cascaded early stop module, a detection and update module, a hierarchical transfer decision module, and an interpretation and traceability module. The image and business data access module is used to acquire a set of multi-source material nodes for car insurance claims that have been received by the car insurance claims platform and entered the review and processing stage. The evidence item standardization module is used to generate standardized evidence items containing abnormal signal identifiers and abnormal attribution identifiers for the data information of the multi-source material node set and the received analytical image material. The anomaly attribution prediction module is used to perform anomaly attribution prediction on the anomaly signals in the standardized evidence item, predict the anomaly signals as single-source anomaly attribution or cross-evidence conflict attribution, and backfill the anomaly attribution prediction result to the standardized evidence item. The four-domain calculation module is used to map the multi-source material node set to the link credibility domain, the anomaly conflict domain, the uncertainty domain, and the cross-evidence conflict domain based on the anomaly attribution prediction result, and to calculate the link credibility score C, the anomaly conflict score R, the uncertainty score U, and the cross-evidence conflict degree K respectively; and to make the anomaly conflict score side and the cross-evidence conflict degree side mutually exclusive in the calculation of the same anomaly signal. The cascaded early stop module is used to perform cascaded early stop judgment based on the uncertainty score U and the cross-evidence conflict degree K, and record the material node and triggering feature that triggers the early stop, and the cascaded early stop is not triggered by the abnormal conflict score alone; The detection and update module is used to perform at least one of the following on image material nodes that have not triggered cascaded early stop: image evidence detection, business field consistency verification, cross-case reuse detection or subject historical risk verification, and update the detection results to the abnormal conflict domain, uncertainty domain or cross evidence conflict domain. The hierarchical transfer decision module is used to output hierarchical transfer actions according to a preset action set based on the combination of the link credibility score C, abnormal conflict score R, uncertainty score U and cross-evidence conflict degree K. The explanation and traceability module is used to generate an explainable audit trajectory corresponding to the hierarchical flow action, which includes mutually exclusive inclusion results and action partial order level.

[0022] Combination Figure 2 , Figure 3 As shown, in this embodiment, the above-described system is used to implement a method for hierarchical transfer of vehicle insurance claim images. Figure 2 The modules for multi-source material access, standardized evidence item generation, anomaly attribution prediction, C / R / U / K four-domain calculation, cascading early stop, hierarchical flow decision-making, action partial order control, and interpretable audit trajectory output correspond to the following modules respectively. Figure 3 The main processing links from S1 to S8 are shown. Figure 3 In this method, S1 corresponds to the acquisition of multi-source material node sets, S2 corresponds to the generation of standardized evidence items, S3 corresponds to the prediction of anomaly attribution, S4 corresponds to the separate calculation of the four domains C, R, U, and K and the inclusion of R / K mutual exclusion, S5 corresponds to the judgment of U / K cascade early stop, S6 corresponds to the detection of materials that have not stopped early and the updating of R, U, and K, S7 corresponds to the output of hierarchical flow actions under the action partial order constraint, and S8 corresponds to the output of interpretable audit trajectory. The method includes the following steps: S1, using the image and business data access module, obtain a set of multi-source material nodes for auto insurance claims cases that have been received by the auto insurance claims platform and entered the review and processing stage. Specifically, obtain a set of multi-source material nodes for auto insurance claims cases that have been received by the auto insurance claims platform and entered the review and processing stage. The set of multi-source material nodes includes at least image material nodes, and at least one of the following associated with the image material nodes: reporting node, damage assessment node, repair node, compensation node, log upload node, platform receiving node, or manual supplementary data collection node.

[0023] S2, using the evidence item standardization module, standardized evidence items containing anomaly signal identifiers and anomaly attribution identifiers are generated for the data information of the multi-source material node set and the received analytical image materials. Specifically, standardized evidence items are generated for at least one of the following: file format information, metadata information, image recoding information, client acquisition logs, upload message information, task identifiers, device fingerprints, platform reception logs, and business fields of the multi-source material node set and the analytical image materials uploaded by the subject. The uploading subject includes at least one of the following: policyholder, insured, complainant, surveyor, repair organization account, rescue organization account, uploading client account, or platform account associated with the case material upload. Each standardized evidence item includes at least one of the following: evidence dimension, evidence direction, evidence strength, source reliability discount coefficient, associated material node, anomaly signal identifier, and anomaly attribution identifier.

[0024] S3, the anomaly attribution prediction module is used to perform anomaly attribution prediction on the anomaly signals in the standardized evidence items, predicting the anomaly signals as either single-source anomaly attribution or cross-evidence conflict attribution, and the anomaly attribution prediction result is backfilled into the standardized evidence items. Specifically, anomaly attribution prediction is performed on the anomaly signals in the standardized evidence items, predicting the anomaly signals as either single-source anomaly attribution or cross-evidence conflict attribution, and the anomaly attribution prediction result is backfilled into the standardized evidence items; wherein, single-source anomaly attribution means that the anomaly signal can be independently triggered by a single image material, a single log, or a single business field, and cross-evidence conflict attribution means that the anomaly signal is triggered by an inconsistency relationship between at least two evidence sources, material nodes, or business fields.

[0025] S4, using the four-domain calculation module based on the anomaly attribution prediction result, the multi-source material node set is mapped to the link credibility domain, anomaly conflict domain, uncertainty domain, and cross-evidence conflict domain, respectively, and the link credibility score C, anomaly conflict score R, uncertainty score U, and cross-evidence conflict degree K are calculated respectively; and the anomaly conflict score side and the cross-evidence conflict degree side are mutually exclusive in the calculation of the same anomaly signal. Specifically, based on the anomaly attribution prediction result, the multi-source material node set is mapped to the link credibility domain, anomaly conflict domain, uncertainty domain, and cross-evidence conflict domain, respectively, and the link credibility score C, anomaly conflict score R, uncertainty score U, and cross-evidence conflict degree K are calculated respectively; wherein, the anomaly conflict score R only receives anomaly signals that are predicted to be attributed to a single-source anomaly, and the cross-evidence conflict degree K only receives anomaly signals that are predicted to be attributed to cross-evidence conflict; the side where the same anomaly signal is not included only retains the reference relationship to that anomaly signal without weighting the score.

[0026] S5, the cascaded early stop module is used to perform cascaded early stop judgment based on the uncertainty score U and the cross-evidence conflict degree K, and record the material node and triggering feature that triggers the early stop, and the cascaded early stop is not triggered by the abnormal conflict score alone. Specifically, the cascaded early stop judgment is performed based on the uncertainty score U and the cross-evidence conflict degree K; when U reaches a preset uncertainty threshold and the key material node is missing or unverifiable, or K reaches a preset conflict threshold, before calling the preset high-cost detection resources, at least one of the following flow actions is output: supplementary acquisition, data source re-acquisition, abnormal verification, or verification upgrade; the preset high-cost detection resources include at least one of the following: high-computing-power image forensics detection model, cross-modal comparison model, or manual verification queue, and the material node and triggering feature that triggers the early stop are recorded; wherein, the increase of the abnormal conflict score R does not trigger the cascaded early stop judgment alone.

[0027] S6, the detection and update module performs at least one of the following on image material nodes that have not triggered cascading early termination: image evidence detection, business field consistency verification, cross-case reuse detection, or subject historical risk verification. The detection results are then updated to the abnormal conflict domain, uncertainty domain, or cross-evidence conflict domain. Specifically, for image material nodes that have not triggered the cascading early termination judgment, at least one of the following is performed: image evidence detection, business field consistency verification, cross-case reuse detection invoked when preset reuse trigger conditions are met, or subject historical risk verification. The detection or verification results are then updated to the abnormal conflict domain, the uncertainty domain, or the cross-evidence conflict domain.

[0028] S7, the hierarchical transfer decision module outputs hierarchical transfer actions according to a preset action set based on the combination of the link credibility score C, abnormal conflict score R, uncertainty score U, and cross-evidence conflict degree K. Specifically, based on the combination of the link credibility score C, abnormal conflict score R, uncertainty score U, and cross-evidence conflict degree K, hierarchical transfer actions are output according to a preset action set; wherein, the preset action set includes at least two of the following: exemption from inspection, low-ratio sampling, downgraded passage, supplementary collection or data source re-collection, abnormal verification, and verification upgrade, and each hierarchical transfer action corresponds to an action partial order level L(a) that increases from lenient to strict handling; wherein, the action partial order level is set to increase sequentially according to exemption from inspection, low-ratio sampling, downgraded passage, supplementary collection or data source re-collection, abnormal verification, and verification upgrade.

[0029] Specifically, during the iterative review process of the same case, when any of the abnormal conflict score R, uncertainty score U, or cross-evidence conflict degree K rises to the corresponding threshold range, the hierarchical transfer action of the current round of output is triggered. Satisfying the partial order of actions No less than the previous round of graded circulation movements The partial order of actions, i.e. The hierarchical flow of actions is prohibited from migrating to a more relaxed direction in the action partial order; wherein, the action partial order constraint and the cascade early stop judgment are independent control mechanisms, the action partial order constraint does not replace the cascade early stop judgment, and only the abnormal conflict score R increases without triggering the cascade early stop judgment alone.

[0030] In this embodiment, the partial order classification and explanation of the above actions are shown in Table 1. (The partial order level of the actions is not directly converted from a single real-time score, but is pre-mapped to an ordinal value L(a) by a preset set of actions according to the severity of the handling, the required review resources, and the irreversibility of the case status. In this embodiment, exemption from inspection, low-proportion sampling, downgraded approval, supplementary collection or data source re-collection, anomaly verification, and verification upgrade are set to levels 1 to 6 in sequence; the same action takes the same level in different cases, and if multiple conditions are met at the same time, the highest level is taken.) Table 1

[0031] S8, using the explanation and traceability module, generate an interpretable audit trajectory corresponding to the hierarchical flow action, including mutually exclusive inclusion results and action partial order level. Specifically, output the interpretable audit trajectory corresponding to the hierarchical flow action, which includes at least a trigger domain, trigger material node, trigger evidence item, and anomaly signal identifier. Abnormal attribution identifier R-side entry marker, K-side entry marker, trigger threshold, action partial order level L(a), and flow action basis.

[0032] In this embodiment, the standardized evidence item is represented as a set consisting of evidence dimension, evidence direction, evidence strength, source reliability discount coefficient, associated material type or link node, abnormal signal identifier, and abnormal attribution identifier. The abnormal attribution identifier includes at least single-source abnormal attribution and cross-evidence conflict attribution, and the abnormal signal identifier is used to track mutually exclusive attributions of the same abnormal signal on the R-side or K-side. Specifically, in one embodiment, the system converts link nodes, metadata, channels, system logs, content credentials, watermarks, image forensics results, business verification results, and subject profiles into unified evidence items. The standardized evidence item is represented as... ,in, Indicates the dimensions of evidence. Indicates the direction of the evidence. Indicates the strength of evidence. This represents the source reliability discount factor. Indicates the associated material type or link node. This indicates an abnormal signal identifier and is used to track the mutual exclusion of the same abnormal signal on the R side or K side; Indicates the attribution identifier for the anomaly. The values ​​can include at least single-source anomaly attribution and cross-evidence conflict attribution. The fields for the anomaly attribution prediction module are generated and backfilled into the standardized evidence items before calculating R and K.

[0033] The anomaly attribution prediction module runs before calculating R and K: when an anomaly signal is independently triggered by a recoding feature of a single image file, a single missing log, a single field format anomaly, or a single source reliability degradation. This is identified as a single-source anomaly; when an anomaly signal is triggered by inconsistencies between image materials and report information, damage assessment information, repair project information, invoice information, platform received logs, or historical case materials. The signal is identified as belonging to a cross-evidence conflict. If the same anomalous signal simultaneously meets the conditions for single-source anomalous attribution and cross-evidence conflict attribution, it is preferentially identified as belonging to a cross-evidence conflict. The single-source anomalous attribution side only retains the reference relationship and does not participate in the weighted scoring on the R side. In this embodiment, the meaning, value, and function of the above parameters are shown in Table 2.

[0034] Table 2

[0035] In this embodiment, the anomaly attribution prediction includes: when an anomaly signal is independently triggered by a single image file's recoding feature, a single missing log, a single field format anomaly, or a single source's decreased reliability, the anomaly signal is identified as a single-source anomaly attribution; when an anomaly signal is triggered by an inconsistency between image materials and case report information, damage assessment information, repair project information, invoice information, platform received logs, or historical case materials, the anomaly signal is identified as a cross-evidence conflict attribution; when the same anomaly signal simultaneously meets both the single-source anomaly attribution condition and the cross-evidence conflict attribution condition, it is preferentially identified as a cross-evidence conflict attribution, and the single-source anomaly attribution side only retains the reference relationship and does not enter the R-side weighted scoring.

[0036] In this embodiment, the link credibility score C is used to measure the reliability of the material link from collection, uploading, platform reception to review and use. The more complete the link nodes and the more reliable the source, the higher the link credibility score C; the more critical link breakpoints and hard stripping features, the lower the link credibility score C. The calculation process is as follows: multiply the completeness of each link node by the reliability of the evidence source, and accumulate them according to the link node weights to obtain a basic credibility value; subtract the critical node missing penalty and hard stripping penalty after corresponding weight adjustment from the basic credibility value; then normalize the subtraction result to obtain the link credibility score C. The link credibility score C satisfies the following formula: ; Where j = 1 to J, This represents the integrity of the j-th link node. Indicates the reliability of the evidence source for the j-th link node. Let L represent the weight of the j-th link node, J represent the number of link nodes, L represent the penalty for missing critical nodes, and H represent the hard stripping penalty. This indicates that the critical node lacks penalty weights. Let σ represent the hard stripping penalty weight, and σ represent the normalization function. The process of determining the hard stripping penalty H is as follows: Read the image's acquisition channel identifier, EXIF ​​metadata, file name, pixel size, encoding parameters, as well as the upload account, device fingerprint, and task identifier; generate EXIF ​​missing feature values ​​when the preset EXIF ​​key fields of the controlled natively captured image are simultaneously missing; generate file name export feature values ​​when the file name matches the preset social software export naming rules; generate screenshot feature values ​​when the pixel size and aspect ratio match the preset terminal screen resolution and screenshot boundary features are detected; generate secondary compression or recoding feature values ​​when the JPEG quantization table is inconsistent with the acquisition end encoding template or when double compression period features are detected; generate subject inconsistency feature values ​​when the binding relationship of the upload account, device, and task is inconsistent with the acquisition log; multiply each feature value by its corresponding confidence level and preset weight, and then sum them up, truncating the sum to [0,1] to obtain the hard stripping penalty H.

[0037] In this embodiment, the anomaly conflict score R is used to measure the comprehensive risk of single-source anomalies, single image forensics anomalies, and single business field anomalies in the material. It represents the anomaly intensity under the single-source anomaly attribution and does not include cross-evidence conflicts formed by inconsistent relationships between at least two evidence sources, material nodes, or business fields. Cross-evidence conflicts are characterized separately by the cross-evidence conflict degree K.

[0038] To avoid double counting, the anomaly conflict score R only filters and includes anomaly signals whose anomaly attribution is identified as a single-source anomaly (R is the normalized risk value after filtering each standardized evidence item according to the anomaly signal identifier; "filtering and including" here means that only evidence items that meet the single-source anomaly attribution condition are sent to the weighted accumulation process of R).

[0039] The cross-evidence conflict degree K only filters and includes anomalous signals whose anomalous attribution is classified as cross-evidence conflict attribution (K is the normalized conflict value after filtering each conflict sub-item according to the anomalous signal identifier; "filtering and including" here means that only evidence items that meet the cross-evidence conflict attribution conditions are sent to the weighted accumulation process of K); the side where the same anomalous signal is not included only retains the reference relationship (reference relationship means that the non-included side retains the same anomalous signal identifier g). k Related material node t k Its evidence item version number is used to trace whether the signal has been scored by the other side; the reference does not generate new weights or scores and does not participate in weighted scoring.

[0040] The calculation process for the anomaly conflict score R is as follows: The anomaly strength, source reliability, and anomaly evidence weight of each single-source anomaly evidence item are multiplied and summed. A baseline bias and the subject's historical risk prior, adjusted according to the subject's historical risk prior weights, are added to the sum, and the result is truncated to [0,1]. The calculation formula is: ; Where k=1 to , The number of anomalous evidence items indicating the attribution of a single-source anomaly. For baseline offset, This indicates the strength of the anomalous evidence that is identified as a single-source anomalous attribution. Indicating the reliability of the source of the unusual evidence, Indicates the weight of anomalous evidence. This indicates the prior historical risk of the subject. This represents the prior weight of the subject's historical risk, and the calculation result is truncated to [0,1]. clip This represents a numerical truncation function.

[0041] Among them, the anomaly intensity z k The confidence level s is output by the corresponding single-source detector. k Press z k = clip ((s k -s min ) / (s max -s min Normalized to 0,1), we get , where s min s max The lower and upper limits are set for the detectors; when multiple detectors operate simultaneously, they are weighted according to preset detector weights and truncated to [0,1].

[0042] Source reliability is determined by factors such as the baseline reliability of the data collection channel, the integrity of the source logs, the verifiability of metadata, and the consistency of the upload subject binding. We get the weighted average, where For the l-th factor, and .

[0043] The weight of unusual evidence is a non-normalized value formed by the importance of the evidence dimension, the importance at the current stage of review, and the discriminative power of the evidence. ;when When >0, ;when At that time, each , so that each Non-negative and the sum is not greater than 1.

[0044] The baseline bias is a risk baseline in the absence of single-source anomalous evidence. It is used to represent the prior benchmark for case type and review stage. It is determined by calibration of the benchmark risk rate of historically labeled samples and fixed within a preset range.

[0045] The prior weight η of the subject's historical risk is the influence coefficient of historical risk on the current R, with a value of [0,1]; it is determined by pre-set lookup table or validation set calibration based on the review stage, historical sample size, and historical record reliability.

[0046] Prior historical risks of the main body It is a smoothed risk ratio calculated based on the number of confirmed abnormal cases in the current uploading entity or related entity, relative to the total number of historical cases. For example... , where a and b are preset smoothing constants; their function is to provide background risk information across cases, but they do not replace evidence in the current case.

[0047] The single-source anomaly evidence includes anomalies in single image evidence collection, anomalies in single-source logs, or anomalies in a single business field.

[0048] Without anomaly attribution prediction and R / K mutual exclusion inclusion, the same secondary compression, field missing, or cross-material inconsistency signal may simultaneously raise both R and K, causing the system to misjudge single-source quality issues as cross-evidence conflicts, or mislead cases with insufficient evidence as high-anomaly cases; through Track the same abnormal signal and through Determining a unique scoring side can reduce false triggers caused by duplicate scoring.

[0049] In this embodiment, the calculation process of the cross-evidence conflict degree K is as follows: Standardized evidence items with anomaly attribution identifiers indicating cross-evidence conflict attribution are selected; the conflict degree of various conflict sub-items, such as location conflict, identity field conflict, and cross-case reuse conflict, is calculated respectively; the conflict degree of each conflict sub-item is multiplied by its corresponding weight and then summed; the smaller value between the summed result and 1 is taken. Anomaly signals with anomaly attribution identifiers indicating single-source anomaly attribution are not included in the summation of the cross-evidence conflict degree K. The calculation formula is as follows: ; Where s = 1 to S, and S represents the number of conflicting sub-items. This represents the s-th type of conflicting sub-item. The weight of each conflicting sub-item is represented by the smaller of the weighted sum and 1; each conflicting sub-item only receives anomalous signals that are identified as belonging to cross-evidence conflict.

[0050] In this embodiment, the position conflict, identity field conflict, and cross-case reuse conflict are determined in the following manner. For the position conflict, first check the latitude and longitude value ranges and coordinate systems of the image coordinates and the case registration coordinates, convert the two sets of latitudes and longitudes into radians, calculate the spherical central angle between the two coordinates according to the Haversine relationship and multiply it by the average Earth radius to obtain the spherical distance; when the spherical distance is not greater than the first distance threshold, the position conflict degree is 0 (when D1<d<D2, the position conflict degree is calculated as (d-D1) / (D2-D1)), when it is not less than the second distance threshold, the position conflict degree is 1 (when d<D2 and d>D1, it is calculated as (d-D1) / (D2-D1); when d≤D1, it is 0, so being less than the second distance threshold does not necessarily mean 0, but is determined by the linear result), and linear normalization is performed according to the distance between the two distance thresholds (the two thresholds satisfy D1<D2, D1 is the tolerance boundary, and D2 is the full conflict boundary).

[0051] Wherein, the first distance threshold D1 and the second distance threshold D2 enumerate candidate threshold pairs satisfying D1<D2 through the validation set, and are determined by the maximum value of the comprehensive index of the target detection rate and the false alarm rate. The second distance threshold D2 is calibrated and determined according to the target conflict detection rate and the distance distribution of historical conflict samples, and satisfies 0≤D1<D2.

[0052] For the identity field conflict, first perform character case unification, blank and separator deletion, and confusing character normalization on the license plate, VIN, and bill number, then perform exact matching or calculate the normalized edit distance; map the maintenance items to the standard item codes and then calculate the item set coincidence degree, and determine the identity field conflict degree according to the comparison result between the normalized edit distance or the item set coincidence degree and the preset threshold.

[0053] Wherein, the comparison results include four categories: exact match, lower than , between and and and not lower than ; when it is an exact match or , the field conflict degree is 0, it is 1, the middle interval is calculated linearly according to ; when the item set coincidence degree , it is 0, , it is 1, the middle part is calculated by inverse linear normalization, and a plurality of fields are weighted by preset weights; the comprehensive conflict degree of the identity field can be calculated according to , where are respectively the conflict degrees of the license plate, VIN, bill number and maintenance item, are non-negative weights and the sum is 1.

[0054] For cross-case reuse conflicts, the images (images are original image material nodes in the current case that have been received by the platform, whose pixel content can be read, and which meet the cross-case reuse triggering conditions, including on-site photos, vehicle damage photos, repair process photos, and invoice images, excluding nodes that only contain thumbnails or whose pixels cannot be read) are scaled and grayscaled to generate a perceptual hash vector (a perceptual hash vector is a fixed-length binary vector used to represent the overall visual structure of an image; in this embodiment, the image is scaled to 32×32 pixels and grayscaled, the two-dimensional discrete cosine transform (DCT) is calculated, the mean μ of the 8×8 low-frequency coefficients in the upper left corner including the DC coefficient is taken, and h is set to...) uv =1(D uv ≥μ), expanded into a 64-bit vector h in a fixed order, where D uv To correspond to the DCT coefficients), historical candidate images are obtained through an approximate nearest neighbor retrieval algorithm (approximate nearest neighbor retrieval is a common existing retrieval technology; in this embodiment, a perceptual hash index can be established using HNSW hierarchical navigable small-world graph or LSH local sensitive hashing; after excluding the current case in the retrieval index stage, the top M candidates with the smallest distance are obtained; this algorithm is only used to improve candidate retrieval efficiency, and the final conflict degree is still calculated by Hamming distance) to calculate the Hamming distance between the perceptual hash vectors of the image to be detected and the candidate images; when the Hamming distance is not greater than the preset reuse distance threshold (reuse distance threshold τ), historical candidate images are obtained. reuse The calibration is determined by a validation set consisting of recoded, compressed, and cropped samples of the same image and samples from different cases. Specifically, this is achieved by enumerating candidate thresholds and selecting the one with the highest detection rate under a preset false alarm rate constraint (different thresholds can be set according to image type). When candidate images are associated with different cases, the result is determined by the inverse normalization of the Hamming distance and the multiplexing distance threshold (the inverse normalization result is c). reuse =1-d H / τ reuse (0≤d) H ≤τ reuse ); d H =0 is 1, 0 <d H <τ reuse When the value is between 0 and 1, d H =τ reus When e is 0, d H >τ reuse If no reuse conflict is formed, it is recorded as 0; if there are multiple candidate images of different cases, the maximum value among them is taken as the cross-case reuse conflict sub-item conflict degree) to determine the cross-case reuse conflict degree.

[0055] The character normalization function is: RemoveSep removes whitespace and separators, Upper is uniformly uppercase, and Map maps O / 0, I / 1, B / 8, etc. to uniform characters based on a preset easily confused character table; the normalized string then participates in the matching.

[0056] Normalized edit distance is defined as , where Lev is the minimum number of insertions, deletions, or replacements required to transform string N(x) into N(y), and |·| is the string length.

[0057] Repair projects search for standard project codes in a pre-defined project dictionary using project name, alias, repair process keywords, and vehicle model compatibility. Projects that cannot be uniquely matched are marked as unknown codes, reducing their source reliability. Let A and B be the sets of valid standard codes for the two sets of repair projects to be compared; unknown codes are not included in the sets. When A or B is empty, the conflicting sub-item is not scored; otherwise, the overlap of the project sets is calculated using J(A,B)=|A∩B| / |A∪B|, and the project conflict degree is 1-J(A,B).

[0058] The thresholds for each field are calibrated from historical labeled samples according to field type, character length and OCR error rate. The low threshold τ0 and high threshold τ1 are determined by enumerating candidate threshold pairs through the validation set and using the maximum value of the comprehensive index of target detection rate and false alarm rate.

[0059] In this embodiment, the uncertainty score U is used to measure whether the current materials are sufficient to support automatic judgment. U increases when key evidence is missing, evidence is unverifiable, or explanatory items are insufficient; U decreases when evidence is sufficient and verifiable. The cascading early termination judgment uses U and K as the main triggering quantities: when U reaches the preset uncertainty score threshold and key material nodes are missing or unverifiable, supplementary collection or data source re-collection is prioritized; when K reaches the preset conflict threshold, anomaly verification or verification escalation is prioritized; when only R reaches the preset anomaly threshold and U and K do not reach the corresponding early termination threshold, cascading early termination is not triggered individually, but rather proceeds to subsequent detection or hierarchical flow decision-making. R is used to indicate the intensity of single-source anomalies and can trigger partial order constraints or subsequent detection updates, but it is not used as a separate triggering quantity for cascading early termination, thereby preventing cases with low conflict but single-source anomalies from prematurely entering high-intensity verification. The preset conflict threshold is denoted as τ. K The risk strategy is pre-configured according to case type, review stage, and manual processing capacity. Specifically, candidate thresholds are enumerated through a validation set, and the one with the highest detection rate is selected under a preset false alarm rate constraint; when K≥τ K This may trigger an abnormal verification or verification upgrade.

[0060] The calculation process for the uncertainty score U is as follows: Determine the expected evidence items based on the case type and current review stage, and assign a "should exist" flag to each; assign a verifiable flag to expected evidence items that already exist and can be verified through source logs, metadata, or related business fields; assign a "should exist" flag, a verifiable flag, and a source reliability discount factor (source reliability discount factor) to each expected evidence item. ),in For the l-th source factor, and The verifiable evidence quantity is obtained by multiplying and summing the results. The expected evidence quantity is then calculated by adding a preset non-zero constant to the sum of the values ​​that should have a label. The ratio of the verifiable evidence quantity to the expected evidence quantity is calculated and truncated to [0,1]. The uncertainty score U is then obtained by subtracting this ratio from 1. The formula is as follows: ; Where i = 1 to N, and N represents the expected number of evidence items. Indicates whether the i-th expected evidence item is verifiable ( It is a binary label; it is set to 1 if the i-th item already exists and is verified by source logs, metadata or business association, otherwise it is set to 0. Indicates whether the i-th expected evidence item should exist ( The label is binary; the i-th item is set to 1 if the evidence template should exist according to the case type and the current review stage, and 0 otherwise. This represents the source reliability discount factor. To prevent constants with zero denominators, the proportions within the parentheses are truncated to [0,1].

[0061] In multiple rounds of review for the same case, if new materials cause the link credibility score C to increase, but none of the abnormal conflict score R, uncertainty score U, or cross-evidence conflict degree K decreases below the corresponding threshold, then the action partial order level of the current hierarchical transfer action is not lower than the action partial order level of the previous round of hierarchical transfer action. The interpretable review trajectory also includes the evidence item version number, material node identifier, abnormal signal identifier, abnormal attribution identifier, R-side inclusion marker, K-side inclusion marker, early stop trigger marker, action partial order level, and the generation time of the hierarchical transfer action.

[0062] In this embodiment, for image material nodes that have not triggered the cascading early termination judgment, the system performs at least one of the following: image evidence detection, business field consistency verification, cross-case reuse detection or subject historical risk verification invoked when preset reuse trigger conditions are met, and updates the detection or verification results to the abnormal conflict domain, uncertainty domain, or cross-evidence conflict domain. The interpretable review trajectory output by the system includes at least the trigger domain, trigger material node, trigger evidence item, abnormal signal identifier, abnormal attribution identifier, R-side inclusion marker, K-side inclusion marker, trigger threshold, early termination trigger marker, action partial order level L(a), and flow action basis. In this embodiment, the output trajectory fields and their meanings are shown in Table 3.

[0063] Table 3

[0064] In one specific embodiment, the platform receives a set of vehicle damage photos, police report fields, damage assessment records, repair items, and uploaded logs. The system first generates standardized evidence items for each material and configures anomaly signal identifiers and anomaly attribution identifiers for each abnormal signal. If an image exhibits secondary compression characteristics but does not contradict other materials, the abnormal signal is counted in R but not in K; if the license plate, location, or repair item in the same image is inconsistent with the police report fields or historical case materials, the abnormal signal is counted in K but not repeatedly counted in R. When a critical platform receives missing logs, causing U to reach a threshold, the system outputs supplementary collection or data source re-collection before calling a higher-powered detection model; when K reaches a conflict threshold, the system outputs anomaly verification or verification upgrade. If C increases after subsequent material supplementation but K still does not fall below the threshold, the action order level remains no lower than the previous round, thus avoiding a direct regression from anomaly verification to exemption from inspection.

[0065] In a set of example parameters, the first round of review yielded C=0.42, R=0.31, U=0.82, and K=0.18. The preset uncertainty threshold was 0.75, and the conflict threshold was 0.70. Because U reached the threshold and the platform received missing logs, the system output supplementary collection or data source re-collection, with an action partial order level. =4. After supplementary collection, the platform's received logs were completed and the reliability of the source was improved, resulting in C=0.68, R=0.33, U=0.41, and K=0.64. At this point, neither U nor K reached the cascading early stop threshold, and the system entered the subsequent detection and hierarchical flow decision-making process; since K=0.64 falls within the preset near-threshold range [0.8τ], K ,τ K Output anomaly verification. =5. If, after the third round of newly added maintenance invoices, it is found that the maintenance items are inconsistent with the damage assessment records, and the same abnormal signal simultaneously exhibits both single-source field anomaly and cross-evidence conflict characteristics, then m i Priority is given to cross-evidence conflict attribution, with K side accounting for it and R side only retaining the reference; for example, when C=0.70, R=0.34, U=0.36, and K=0.78, the system outputs verification upgrade because K has reached the conflict threshold. =6. In the above process, although C increased in the second round, Not less than After the third round of K increases Continue to improve and meet the requirements. Constraints of a partial order level no lower than that of the previous round of actions.

[0066] To verify the technical effectiveness of the method described in this embodiment, offline playback and hierarchical annotation were performed on anonymized auto insurance claim samples. This verification is used to illustrate the technical correspondence between C / R / U / K separate calculation, R / K mutual exclusion inclusion, U / K cascaded early stop, R not stopping early alone, and action partial order constraints. In the anonymized sample playback verification, a set of hierarchical verification samples covering controlled collection, survey app upload, album import, WeChat or enterprise WeChat transfer, screenshots, offline re-upload, special invoices, and rescue scenarios were used; the samples included 487 auto insurance claim cases and 9,306 image materials, and were annotated according to four structural categories: complete chain without anomalies, broken chain without anomalies, broken chain with anomalies, and complete chain with conflicts.

[0067] In this embodiment, samples with complete chains and no anomalies are used to verify that when the chain credibility score C is high and the anomaly conflict score R, uncertainty score U, and cross-evidence conflict degree K are all low, they can pass inspection without inspection or be sampled at a low proportion. Samples with broken chains but no anomalies are used to verify that when U increases, the system will prioritize supplementary collection, data source re-collection, or downgraded approval, rather than directly equating missing chains with fraud. Samples with broken chains and anomalies are used to verify anomaly verification and verification upgrade under the combined effect of U and K. Samples with complete chains but conflicts are used to verify that high C should not cover high K. That is, when cross-evidence conflicts such as OCR, license plate / VIN, GPS, invoices, or maintenance items reach the threshold, even if the uploaded chain is complete, direct inspection exemption is still prohibited.

[0068] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A method for hierarchical transfer of images in auto insurance claims, characterized in that, include: Obtain a collection of multi-source material nodes for auto insurance claims that have been received by the auto insurance claims platform and entered the review and processing stage; Based on the data information of the multi-source material node set and the received analytical image material, a standardized evidence item containing anomaly signal identifiers and anomaly attribution identifiers is generated; Perform anomaly attribution prediction on the anomalous signals in the standardized evidence item, predicting the anomalous signals as single-source anomaly attribution or cross-evidence conflict attribution, and backfill the anomaly attribution prediction results into the standardized evidence item; Based on the anomaly attribution prediction results, the multi-source material node set is mapped to the link credibility domain, anomaly conflict domain, uncertainty domain, and cross-evidence conflict domain, respectively, and the link credibility score, anomaly conflict score, uncertainty score, and cross-evidence conflict degree are calculated respectively. And make the abnormal conflict score side and the cross-evidence conflict degree side mutually exclusive in counting the same abnormal signal; Based on the uncertainty score and the cross-evidence conflict degree, a cascaded early stop judgment is performed, and the material node and triggering feature that trigger the early stop are recorded; For image material nodes that do not trigger the cascading early stop judgment, perform image evidence detection, business field consistency verification, or cross-case reuse detection called when the preset reuse trigger condition is met, and update the detection or verification results to the abnormal conflict domain, the uncertainty domain, or the cross-evidence conflict domain. Based on the combination of the link credibility score, anomaly conflict score, uncertainty score and cross-evidence conflict degree, hierarchical flow actions are output according to the preset action set. Generate an interpretable audit trajectory corresponding to the hierarchical flow action, including mutually exclusive inclusion results and action partial order level.

2. The method for hierarchical transfer of vehicle insurance claim images according to claim 1, characterized in that, The standardized evidence item is a set consisting of evidence dimension, evidence direction, evidence strength, source reliability discount coefficient, associated material type or link node, abnormal signal identifier and abnormal attribution identifier; wherein, the value of the abnormal attribution identifier includes at least single-source abnormal attribution and cross-evidence conflict attribution, and the abnormal signal identifier is used to track the mutually exclusive inclusion of the same abnormal signal in the abnormal conflict score side or the cross-evidence conflict degree side.

3. The method for hierarchical transfer of vehicle insurance claim images according to claim 1 or 2, characterized in that, The anomaly attribution prediction includes: when an anomaly signal is independently triggered by a single image file's recoding feature, a single missing log, a single field format anomaly, or a single source's decreased reliability, the anomaly signal is identified as a single-source anomaly attribution; when an anomaly signal is triggered by an inconsistency between image materials and case report information, damage assessment information, repair project information, invoice information, platform received logs, or historical case materials, the anomaly signal is identified as a cross-evidence conflict attribution; when the same anomaly signal simultaneously meets both the single-source anomaly attribution condition and the cross-evidence conflict attribution condition, it is identified as a cross-evidence conflict attribution.

4. The method for hierarchical transfer of vehicle insurance claim images according to claim 1, characterized in that, The calculation of the link credibility score includes: multiplying the completeness of each link node by the reliability of the evidence source of the corresponding link node, and accumulating them according to the link node weights to obtain a basic credibility value; determining the penalty for missing key nodes based on the number and importance level of missing key nodes, and determining the penalty for hard stripping based on the type, number and confidence level of the detected hard stripping features; subtracting the penalty for missing key nodes and the penalty for hard stripping after adjustment of the corresponding penalty weights from the basic credibility value, and normalizing the subtraction results to obtain the link credibility score.

5. The method for hierarchical transfer of vehicle insurance claim images according to claim 4, characterized in that, The determination of the hard stripping penalty includes: reading the acquisition channel identifier, EXIF ​​metadata, file name, pixel size, encoding parameters, and upload account, device fingerprint, and task identifier of the image; generating corresponding hard stripping feature values ​​for controlled natively captured images based on the following conditions: simultaneous absence of preset EXIF ​​key fields, file name matching preset social software export rules, pixel size and aspect ratio matching preset terminal screen resolution and detection of screenshot boundary features, JPEG quantization table or compression cycle features indicating secondary compression or re-encoding, and inconsistent binding relationship between upload account, device, and task with acquisition logs; accumulating the hard stripping features according to their confidence level and preset weight, and truncating the accumulated result to [0,1] to obtain the hard stripping penalty H.

6. The method for hierarchical transfer of vehicle insurance claim images according to claim 1, characterized in that, The calculation of the anomaly conflict score includes: screening standardized evidence items whose anomaly attribution is identified as single-source anomaly attribution; multiplying the anomaly strength, source reliability discount coefficient, and anomaly evidence weight of each standardized evidence item and summing them; adding the baseline bias and the subject's historical risk prior adjusted according to the subject's historical risk prior weight to the resulting summed value, and truncating the result to [0,1]; anomaly signals whose anomaly attribution is identified as cross-evidence conflict attribution do not participate in the summation of the anomaly conflict score.

7. The method for hierarchical transfer of vehicle insurance claim images according to claim 1, characterized in that, The calculation of the cross-evidence conflict degree includes: screening standardized evidence items whose abnormal attribution identifiers are cross-evidence conflict attributions; calculating the conflict degree of location conflict sub-items, identity field conflict sub-items, and cross-case reuse conflict sub-items respectively; multiplying the conflict degree of each conflict sub-item by the corresponding conflict sub-item weight and summing them, and comparing the summed result with 1, taking the smaller value to obtain the cross-evidence conflict degree; abnormal signals whose abnormal attribution identifiers are single-source abnormal attributions are not included in the summation of the cross-evidence conflict degree.

8. The method for hierarchical transfer of vehicle insurance claim images according to claim 1, characterized in that, The calculation of the uncertainty score includes: determining expected evidence items based on the case type and the current review stage, and setting a "should exist" flag for each expected evidence item; setting a verifiable flag for expected evidence items that already exist and can be verified through source logs, metadata, or related business fields; multiplying and summing the "should exist" flag, the verifiable flag, and the source reliability discount coefficient for each expected evidence item to obtain the verifiable evidence quantity, and using the sum of the "should exist" flags plus a preset non-zero constant as the expected evidence quantity; calculating the ratio of the verifiable evidence quantity to the expected evidence quantity and truncating it to [0,1], and subtracting this ratio from 1 to obtain the uncertainty score.

9. The method for hierarchical transfer of vehicle insurance claim images according to claim 1, characterized in that, The cascaded early stop judgment includes: when the deterministic score reaches a preset uncertainty threshold and key material nodes are missing or unverifiable, priority is given to outputting supplementary collection or data source re-collection; when the cross-evidence conflict degree reaches a preset conflict threshold, priority is given to outputting anomaly verification or verification upgrade; when only the anomaly conflict score reaches a preset anomaly threshold and the uncertainty score and cross-evidence conflict degree do not reach the corresponding early stop threshold, cascaded early stop is not executed, but subsequent detection or hierarchical flow decision is entered; wherein, when R reaches the preset anomaly threshold, it can trigger the action partial order constraint, but does not trigger the cascaded early stop judgment alone.

10. A vehicle insurance claims image hierarchical transfer system, characterized in that, It includes an image and business data access module, an evidence item standardization module, an anomaly attribution prediction module, a four-domain calculation module, a cascaded early stop module, a detection and update module, a hierarchical transfer decision module, and an interpretation and traceability module; The image and business data access module is used to acquire a set of multi-source material nodes for auto insurance claims cases that have been received by the auto insurance claims platform and entered the review and processing stage; The evidence item standardization module is used to generate standardized evidence items containing abnormal signal identifiers and abnormal attribution identifiers for the data information of the multi-source material node set and the received analytical image material. The anomaly attribution prediction module is used to perform anomaly attribution prediction on the anomaly signals in the standardized evidence item, predict the anomaly signals as single-source anomaly attribution or cross-evidence conflict attribution, and backfill the anomaly attribution prediction result to the standardized evidence item. The four-domain calculation module is used to map the multi-source material node set to the link credibility domain, anomaly conflict domain, uncertainty domain and cross-evidence conflict domain respectively based on the anomaly attribution prediction result, and calculate the link credibility score, anomaly conflict score, uncertainty score and cross-evidence conflict degree respectively. And make the abnormal conflict score side and the cross-evidence conflict degree side mutually exclusive in counting the same abnormal signal; The cascaded early stop module is used to perform cascaded early stop judgment based on the uncertainty score and the cross-evidence conflict degree K, and record the material node and triggering feature that triggers the early stop, and the cascaded early stop is not triggered by the abnormal conflict score alone; The detection and update module is used to perform at least one of the following on image material nodes that have not triggered cascaded early stop: image evidence detection, business field consistency verification, cross-case reuse detection or subject historical risk verification, and update the detection results to the abnormal conflict domain, uncertainty domain or cross evidence conflict domain. The hierarchical transfer decision module is used to output hierarchical transfer actions according to a preset set of actions based on the combination of the link credibility score, abnormal conflict score, uncertainty score and cross evidence conflict degree. The explanation and traceability module is used to generate an explainable audit trajectory corresponding to the hierarchical flow action, which includes mutually exclusive inclusion results and action partial order level.