Insurance claim settlement business post-governor method and equipment

Abstract

The invention discloses an insurance claim settlement business governor method and device, and relates to the field of intelligent insurance claim settlement, and the method comprises the steps: connecting a claim settlement core system, an image platform and a third-party data source, obtaining insurance claim settlement data, and carrying out the preprocessing of the data, and obtaining standardized data; performing parallel reasoning on the standardized data by using a multi-model collaborative screening engine, calculating a real-time weight through a dynamic weight scheduler based on performance indexes of each model in an insurance type-scene dimension and insurance type prior risk information, obtaining a comprehensive risk score in combination with a reasoning result, and judging a high-risk case; the key information of the high-risk case is uploaded to an alliance chain for evidence storage, and an evidence storage transaction identifier is recorded; a checking result is received through the manual checking workbench and fed back to the online self-learning module, incremental training updating and performance verification are carried out on the model when a triggering condition is met, and closed-loop optimization is formed.

Description

Technical Field

[0001] This application relates to the field of intelligent insurance claims technology, and in particular to a method and equipment for post-claims supervision. Background Technology

[0002] The current surge in insurance claims volume poses a significant challenge to the traditional manual post-claims monitoring model. Key issues include: a severe shortage of post-claims monitoring human resources, with manual spot checks covering only about 30% of cases; poor consistency in manual review, resulting in high false positive and false negative rates; a lack of joint verification capabilities for complex cases involving multiple companies and insurance types; a simplistic model selection logic that cannot dynamically adjust strategies based on case type; the inability to effectively feed manual verification results back into the model, leading to rapid performance degradation; and the risk of key data tampering, making regulatory traceability difficult.

[0003] Current mainstream solutions primarily employ a single model for risk identification, such as addressing duplicate claims without comprehensive post-claim supervision; or they use a single expert model set, lacking dynamic weight adjustment mechanisms and blockchain-based evidence storage; or they use large language models, incurring huge computational resource costs and failing to solve data trust issues. These solutions generally suffer from limited coverage, insufficient adaptability, and poor data security.

[0004] How to improve the accuracy of audits while ensuring comprehensive coverage; how to build a multi-model collaborative mechanism to adapt to the needs of different insurance types and scenarios; how to achieve continuous self-learning and dynamic optimization of models; how to ensure the immutability of key data and regulatory traceability; and how to balance system performance with computing resource consumption. Effective solutions to these problems are of great significance for improving the quality of post-claims supervision, reducing operating costs, and meeting regulatory requirements. Summary of the Invention

[0005] This application provides a method and equipment for post-claims supervision to solve the above-mentioned problems.

[0006] On the one hand, this application provides a post-claims supervision method, including the following steps: Step S1: Connect with the core claims system, image platform and third-party data sources to obtain insurance claims settlement data and preprocess it to obtain standardized data; Step S2: Use a multi-model collaborative screening engine to perform parallel reasoning on standardized data. Based on the performance indicators of each model in the insurance type-scenario dimension and the prior risk information of the insurance type, calculate the real-time weights through a dynamic weight scheduler, and combine the reasoning results to obtain a comprehensive risk score and determine high-risk cases. Step S3: Upload key information of high-risk cases to the consortium blockchain for evidence storage and record the evidence storage transaction identifier; Step S4: Receive the verification results through the manual verification workbench and feed them back to the online self-learning module. When the triggering conditions are met, perform incremental training and update of the model and performance verification to form a closed-loop optimization.

[0007] In one implementation of this application, step S1 includes the following preprocessing steps: using statistical methods or rule-based strategies to complete missing information; and correcting or removing abnormal information that exceeds a reasonable range.

[0008] In one implementation of this application, the multi-model collaborative screening engine includes: A basic screening model is used to identify common risks such as missing data, abnormal amounts, and contradictory timelines. Special case models are used to identify long-tail risks in scenarios involving multiple overlapping insurance types, cross-company relationships, and personal injury / disability. The fraudulent claims model is used to identify fraud risk signals such as duplicate invoices, high-frequency claims, and blacklist matching.

[0009] In one implementation of this application, the process of calculating real-time weights through a dynamic weight scheduler is as follows: the dynamic weight scheduler receives the performance indicators of each model and the case insurance information, combines the prior risk information of the insurance type, and calculates the real-time weights through a normalization function.

[0010] In one implementation of this application, in step S4, the main nodes of the consortium blockchain include insurance main nodes, regulatory nodes, judicial nodes, and third-party data verification nodes; the key information uploaded to the blockchain includes the hash value of case data, which is used to verify the integrity of the data.

[0011] In one implementation of this application, the consortium blockchain is equipped with a smart contract that provides a case tracing interface, obtains a complete array of evidence storage transaction identifiers by inputting a case identifier, and supports connection to a judicial blockchain evidence collection platform to achieve mutual recognition of electronic evidence.

[0012] In one implementation of this application, in step S4, the manual verification workbench supports dual-terminal operation on both Web and mobile application terminals, and supports image annotation, voice annotation, and label annotation functions; the manual verification result is returned through a preset interface, and the returned fields include the operator's identifier and the storage address of the voice annotation.

[0013] In one implementation of this application, the online self-learning module presets self-learning, and the triggering condition is: the cumulative number of manually checked results reaches a preset threshold within a preset time period, or the model recognition accuracy is lower than a preset accuracy threshold; the online self-learning module uses an adaptive optimization algorithm combined with an early stopping strategy for incremental training, and only updates the model weight parameters.

[0014] In one implementation of this application, the model performance verification of the online self-learning module adopts a canary release method, in which the updated model is applied to a preset proportion of traffic and run for a preset duration. If the performance indicators improve to the preset requirements, the original model is replaced; at the same time, the original model is retained for a preset duration for abnormal rollback.

[0015] On the other hand, this application also provides an insurance claims business post-supervision device, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to complete the aforementioned insurance claims business post-supervision method.

[0016] The post-claims monitoring method and equipment provided in this application have the following beneficial effects: 1. Leveraging a multi-model collaborative screening engine and a dynamic weight scheduling mechanism, it can cover general risks, long-tail scenario risks, and fraud risks, avoiding the limitations of traditional single-model screening logic. By combining model performance indicators with the prior risks of different insurance types to calculate real-time weights through a dynamic weight scheduler, it can adapt to the needs of different insurance types and scenarios, improve the accuracy of high-risk case judgment, reduce the problems of missed and false reports in manual review, and improve processing efficiency through parallel reasoning, alleviating the dilemma of insufficient manpower and low coverage in traditional manual post-supervision.

[0017] 2. Based on a consortium blockchain-based evidence storage design, this system leverages a consortium blockchain comprised of multiple nodes, including insurance entities, regulators, and judiciary members, to store key information (including case data hash values) of high-risk cases on the blockchain, effectively preventing the risk of critical data tampering. Furthermore, the case tracing interface provided by smart contracts allows for the rapid acquisition of evidence storage transaction identifiers. It also supports integration with judicial blockchain evidence collection platforms to achieve mutual recognition of electronic evidence, addressing the challenges of difficult regulatory tracing and insufficient evidence credibility in traditional post-supervision processes, thus meeting compliance requirements.

[0018] 3. Based on the closed-loop design of the online self-learning module, when the trigger condition of "the manual verification results meet the standard within a continuous preset time but the model accuracy does not meet the standard" is met, incremental training is used to update only the model weights, and performance is verified by canary release, ensuring the stability of model optimization. This mechanism can effectively feed the manual verification results back into the model, avoiding the problem of rapid performance decay in traditional models, realizing continuous adaptive optimization of the model, reducing manual maintenance costs, and maintaining post-monitoring quality in the long term. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1A flowchart of a post-claims supervision method provided in this application embodiment; Figure 2 This is a schematic diagram of an insurance claims processing monitoring device provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] This application provides a method and equipment for post-claims supervision. The technical solution proposed in this application will be described in detail below with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart illustrating a post-claims supervision method provided in an embodiment of this application. Figure 1 As shown, the method mainly includes the following steps: Step S1: Connect with the core claims system, image platform and third-party data sources to obtain insurance claims settlement data and preprocess it to obtain standardized data; Step S2: Use a multi-model collaborative screening engine to perform parallel reasoning on standardized data. Based on the performance indicators of each model in the insurance type-scenario dimension and the prior risk information of the insurance type, calculate the real-time weights through a dynamic weight scheduler, and combine the reasoning results to obtain a comprehensive risk score and determine high-risk cases. Step S3: Upload key information of high-risk cases to the consortium blockchain for evidence storage and record the evidence storage transaction identifier; Step S4: Receive the verification results through the manual verification workbench and feed them back to the online self-learning module. When the triggering conditions are met, perform incremental training and update of the model and performance verification to form a closed-loop optimization.

[0023] In this application, step S1 involves preprocessing focused on data quality optimization, divided into two parts: When filling in missing information, numerical data (such as claim amount and consultation duration) is filled using statistical methods such as mean and median, while categorical data (such as cause of accident and department visited) is filled using the mode; alternatively, information can be filled according to insurance type rules, for example, missing vehicle usage characteristics in car insurance is filled using common characteristics for the same vehicle model. In handling abnormal information, data exceeding a reasonable range (above the 95th percentile of claim amount for the same insurance type) is first verified with the data source; if correct, it is corrected to a reasonable range; otherwise, it is discarded. Data with obvious errors, such as time discrepancies (e.g., the accident occurred before the insurance was purchased), is directly discarded to ensure the standardization and usability of subsequent data.

[0024] In this application, the multi-model collaborative screening engine includes: A basic screening model is used to identify common risks such as missing data, abnormal amounts, and contradictory timelines. Special case models are used to identify long-tail risks in scenarios involving multiple overlapping insurance types, cross-company relationships, and personal injury / disability. The fraudulent claims model is used to identify fraud risk signals such as duplicate invoices, high-frequency claims, and blacklist matching.

[0025] In this application, when the dynamic weight scheduler calculates real-time weights, it first receives three types of key information: the F1 score (performance indicator) of each model (basic screening, special cases, and fraudulent claims model) in the "insurance type-scenario" dimension over the past 30 days; the one-hot vector of the case insurance type (case insurance type information); and the pre-set prior risk bias according to the insurance type (e.g., 0.4 for health insurance, 0.2 for property insurance, i.e., prior risk information of the insurance type). Then, it calculates the weights according to the formula… "Calculation, in which" The algorithm balances the impact of performance metrics and prior risks, and then uses softmax normalization to transform the comprehensive value into probabilistic distributed weights. Finally, it outputs a real-time weight vector, which is dynamically updated with each batch of cases, thus resolving the single-model failure problem caused by "risk type shift".

[0026] In this application, in step S4, the main nodes of the consortium blockchain have clearly defined roles: the insurance main node is responsible for initiating high-risk case evidence storage requests and synchronizing its own claims data; the regulatory node supervises the compliance of the entire evidence storage process and verifies whether the data meets regulatory requirements; the judicial node stores the evidence storage information to support possible subsequent judicial evidence collection; and the third-party data verification node (such as the China Banking and Insurance Regulatory Commission) verifies the authenticity of case-related data and assists in verifying the validity of the information. Among the key information uploaded to the blockchain, the case data hash value is generated using the SHA-256 algorithm and can uniquely correspond to the original case data. When it is necessary to verify the integrity of the data later, it is only necessary to recalculate the data hash value and compare it with the value stored on the blockchain. If they match, it indicates that the data has not been tampered with, ensuring data credibility and supporting subsequent auditing and mutual recognition of evidence.

[0027] In this application, within the consortium blockchain (using Hyperledger Fabric technology) of the post-claims supervision system, smart contracts are the core component for achieving data traceability and evidence mutual recognition. The smart contract has a pre-defined case traceability interface. Users input a unique case identifier (such as caseId) to retrieve a complete array of evidence storage transaction identifiers (txHash) for that case. This array records the entire transaction history from the initiation of evidence storage to confirmation, allowing for reverse tracing of each step of the evidence storage process. Simultaneously, the consortium blockchain supports integration with judicial blockchain evidence collection platforms. Because the key information uploaded to the chain includes the SHA-256 hash value of the case data (ensuring the data has not been tampered with), the judicial platform can directly verify the authenticity of the on-chain data after integration. This ensures that the electronic data stored on the consortium blockchain meets judicial evidence standards, achieving cross-platform electronic evidence mutual recognition and satisfying regulatory auditing and judicial evidence collection needs.

[0028] In this application, in step S4, the manual verification workbench adapts to multi-scenario operation needs, supporting dual-terminal use on both Web (based on Vue technology) and mobile application (based on Flutter technology), facilitating flexible handling of high-risk cases by staff. Functionally, image annotation can accurately mark suspicious locations in case materials, voice annotation can efficiently record verification opinions without manual input, and tag marking can quickly classify case problem types, significantly improving verification efficiency. The verification results are returned through a preset POST / api / v1 / feedback interface. In the returned fields, the operator identifier is used to clarify the verification responsibility entity, the voice annotation storage address allows for subsequent retrieval of the audio recording, and the case identifier and other information are associated to ensure traceability of the results, providing accurate feedback for the online self-learning module and supporting post-supervision closed-loop optimization.

[0029] In this application, the self-learning mechanism of the online self-learning module first defines two triggering conditions: "continuous preset duration" is 3 consecutive days, "cumulative threshold for the number of manually verified results" is 300, and "model recognition accuracy threshold" is 85%. Self-learning is only initiated when both conditions are met simultaneously. During the training phase, the Adam adaptive optimization algorithm is adopted, coupled with an early stopping strategy (patience=10, i.e., termination if there is no improvement in validation set performance after 10 rounds). Only incremental training is performed—the entire model is not retrained, only the weight parameters of the corresponding model are updated. The learning rate is set to 0.05 and the batch size is 512. This avoids wasting computational resources, quickly adapts to data changes, ensures stable model performance, and provides support for post-supervisory closed-loop optimization.

[0030] In this application, the model performance verification of the online self-learning module uses a canary release approach to ensure update stability. Referring to the details in the patent disclosure, the "preset ratio" is 5% of the traffic and the "preset duration" is 24 hours. The updated model is first applied to this portion of traffic to test its actual adaptability on a small scale. Performance evaluation uses the F1 score as the core indicator. If the overall F1 score improves by more than 1% after running (i.e., the preset requirement is met), the original model is completely replaced. At the same time, the "preset duration for retaining the original model" is set to 7 days. If the new model exhibits anomalies during this period (such as a sudden increase in the false alarm rate), the old model can be switched back immediately to avoid affecting subsequent supervision business. This achieves dynamic model optimization and ensures system reliability, supporting the stable operation of the subsequent supervision closed loop.

[0031] The above is an example of an insurance claims post-monitoring system provided by this application. Based on the same inventive concept, this application also provides an insurance claims post-monitoring device. Figure 2 A schematic diagram of a post-claims monitoring device provided in this application embodiment is shown below. Figure 2 As shown, the device mainly includes: at least one processor 201; and a memory 202 communicatively connected to the at least one processor; wherein the memory 202 stores instructions that can be executed by the at least one processor 201, and the instructions are executed by the at least one processor 201 to enable the at least one processor 201 to complete the aforementioned post-insurance claims business supervision method.

[0032] Below is an example of this application in a specific scenario. The health insurance division of a large property insurance company processes approximately 2,000 health insurance claims daily. Previously, they used a post-reporting supervision model of "manually reviewing 30% of cases," which had three major pain points: First, personal injury cases often involve fraud risks such as cross-hospital visits and duplicate document submissions, which are easily missed by manual review; second, health insurance claims have over 130 data fields, and 15% of cases lack crucial information such as treatment duration and cost details, making manual completion inefficient; third, the model only uses a single logistic regression algorithm, resulting in a 4.2% false positive rate for long-tail cases with "multiple insurance types + high-frequency claims," ​​and the model decays rapidly, with accuracy dropping from 88% to 75% after 6 months. To address this, the company deployed a post-reporting supervision system for insurance claims based on "multi-model collaboration + blockchain evidence storage." Using the claim case of health insurance customer Mr. Wang as an example, the system's application process is fully presented.

[0033] Every day at 3:00 AM, the system connects to the company's core claims system, imaging platform, and third-party hospital HIS system and social security settlement platform via HTTPS / JSON protocol to automatically retrieve the previous day's case settlement data. Among them, Wang's case information includes: case ID (HK20240512008), cause of the accident (accidental fall), hospital visited (orthopedics department of hospital A + rehabilitation department of hospital B), claim amount (18,000 yuan), number of receipts (6), and number of historical accidents (3 times in the past year). However, the field of "duration of treatment at hospital A" is missing, and the "daily nursing fee of 5,000 yuan" in the "details of expenses in rehabilitation department of hospital B" exceeds the upper limit of nursing fees for hospitals of the same level in the same region (2,000 yuan / day).

[0034] The system initiates preprocessing: For the missing "visit duration", it is supplemented using the average visit duration (14 days) of similar cases in the orthopedics department of Hospital A in the same region (supplementation using statistical methods); for "daily nursing fee of 5,000 yuan", it is first verified twice with the HIS system of Hospital B to confirm that it is an input error (actually 500 yuan), and corrected to a reasonable value (abnormal information correction). Finally, 120 complete standardized data fields are generated, laying the foundation for subsequent model inference.

[0035] Standardized data enters the multi-model collaborative screening engine, and three models are used for parallel inference (deployed on a TensorFlowServing cluster, with GPU memory isolation and an inference latency of 85ms): The basic screening model (XGBoost, depth 8, number of trees 600) identified Wang's case as having "the expense details of Hospital B had been revised" and "the number of historical claims reached 3 (exceeding the average of 1.2 times for the same type of insurance)", and gave a risk score of 0.72. Special Case Model (RandomForest, 200 trees, depth 12): Through cross-company data verification (connecting to third-party data sources of China Banking and Insurance Information Technology Co., Ltd.), it was found that Wang had simultaneously applied for rehabilitation expenses for the same accident at another life insurance company (multiple companies involved), and a risk score of 0.81 was given; False claims model (L1 regularized Logistic Regression, C=0.1): By comparing the bill hash database, it was found that a rehabilitation fee bill from Hospital B had the same number as another claim bill from Wang three months ago (duplicate bill), and a risk score of 0.85 was given.

[0036] The dynamic weight scheduler receives the F1 scores of the three models for health insurance scenarios over the past 30 days (basic model 0.88, special case model 0.91, fraudulent claim model 0.93), combines them with the prior risk bias of health insurance (0.4), and calculates the weights according to the formula. Real-time weights are calculated as follows: Basic model 0.28, Special case model 0.35, Fraudulent claims model 0.37. Overall risk assessment score: The case was higher than the high-risk case threshold (0.68), so the system marked Wang's case as a high-risk case.

[0037] The system uploads key information about Wang's case to the Hyperledger Fabric consortium blockchain (including four types of nodes: insurance entity, bank-insurance-trust, local court, and State Administration of Financial Regulatory Commission, using PBFT consensus, with a block time of 2.5 seconds) through a blockchain evidence storage gateway. The uploaded content includes the case ID (HK20240512008), the SHA-256 hash value of the case data package (a7f3d2...), the three-model version number (V2.3), the comprehensive risk score (0.80), the weight vector (0.28, 0.35, 0.37), and the evidence storage timestamp (2024-05-13 08:15:32).

[0038] The consortium blockchain returns a transaction identifier txHash (f9e1c5...), which the system writes to its local PostgreSQL database. Subsequently, regulatory agencies can obtain the complete array of evidence-based transactions through the smart contract interface "getTrace(HK20240512008)" to verify the completeness of case materials and the compliance of the judgment process.

[0039] Manual verification and online self-learning perform the following processes: (1) Manual verification: The system assigned Wang's case to the manual verification workbench of the post-supervision specialist Zhang. Zhang opened the case image through the Web client (Vue framework), marked the position of the duplicate invoice number using the "circle annotation function", recorded "The invoice of Hospital B is duplicated with the historical claim, and needs to be verified by contacting the hospital" through the "voice annotation", and checked the "duplicate claim" label. The verification result was returned through the POST / api / v1 / feedback interface, and the fields included caseId (HK20240512008), operatorNo (ZW2023001), and voiceUrl (http: / / xxx / voice / 202405130830.mp3), and finally confirmed that the case was a "false claim".

[0040] (2) Online self-learning: Within three consecutive days, the system received a total of 320 manually verified results (exceeding the threshold of 300), and the accuracy of the health insurance scenario model recognition dropped to 83% (below the threshold of 85%), triggering self-learning. The system adopted the Adam adaptive optimization algorithm (learning rate 0.05, batch=512), combined with an early stopping strategy (patience=10), and only updated the weight parameters of the three models (no need for full training); after training, the new model was applied to 5% of the traffic (200 cases) for 24 hours in a gray-scale release manner. It was observed that the overall F1 score increased from 0.89 to 0.91 (exceeding the preset requirement by 1%), so the original model was fully replaced, while the old model was retained for 7 days in case of abnormal rollback.

[0041] In this scenario, the system achieves three core values: First, Wang's fraudulent claim was accurately identified, preventing the company from suffering a loss of 18,000 yuan, and the accuracy rate of identifying problematic health insurance claims increased to 93.2%; Second, manual verification is only required for high-risk cases (accounting for about 20%), reducing the daily verification volume from 1,200 cases to 450 cases, freeing up 58% of manpower; Third, the model adapts to new characteristics of health insurance fraud through self-learning, extending the decay period from 6 months to 14 months, and the consortium blockchain ensures that the case data is tamper-proof, allowing for traceability within 10 minutes during subsequent regulatory audits, fully meeting the T+2 regulatory reporting requirements.

[0042] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0043] It should also be noted that 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 a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0044] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A method for post-claims administration of an insurance business, characterized in that, The method comprises the following steps: Step S1: connecting the claim settlement core system, the image platform and the third-party data source, obtaining insurance claim settlement data and preprocessing to obtain standardized data; Step S2: using a multi-model collaborative screening engine to perform parallel reasoning on the standardized data, calculating real-time weights based on the performance indicators of each model in the risk category-scene dimension and the prior risk information of the risk category, combining the reasoning results to obtain a comprehensive risk score and determine high-risk cases; Step S3: uploading the key information of the high-risk cases to the alliance chain for storage, and recording the storage transaction identifier; Step S4: receiving the checking results through the artificial checking workbench and feeding back to the online self-learning module, and when the triggering condition is met, performing incremental training and updating and performance verification on the model to form a closed-loop optimization.

2. The method of claim 1, wherein, In step S1, the preprocessing process includes: using statistical methods or rule-based strategies to complete the missing information; and correcting or removing abnormal information that exceeds a reasonable range.

3. The method of claim 1, wherein, The multi-model collaborative screening engine comprises: A basic screening model for identifying general risks such as missing information, abnormal amounts, and time logic contradictions; A special case model for identifying long-tail scenario risks such as multi-risk stacking, cross-company association, and personal injury disability; A false claim model for identifying fraud risk signals such as duplicate invoices, high-frequency claims, and blacklist matching.

4. The method of claim 1, wherein, The process of calculating real-time weights through the dynamic weight scheduler is: receiving the performance indicators of each model and the case risk information through the dynamic weight scheduler, combining the prior risk information of the risk category, and calculating real-time weights through a normalization function.

5. The method of claim 1, wherein, In step S4, the main nodes of the alliance chain include insurance subject nodes, regulatory nodes, judicial nodes and third-party data verification nodes; the key information uploaded to the chain includes case data hash values for verifying data integrity.

6. The method of claim 5, wherein, The alliance chain is deployed with a smart contract, which provides a case tracing interface, obtains a complete storage transaction identifier array by inputting a case identifier, and supports interfacing with a judicial blockchain forensics platform to realize electronic evidence mutual recognition.

7. The method of claim 1, wherein, In step S4, the artificial checking workbench supports dual-end operation of the Web end and the mobile application end, supports image circle annotation, voice comment and label labeling functions; the artificial checking results are returned through a preset interface, and the returned fields include operator identifier and voice comment storage address.

8. The method of claim 1, wherein, The online self-learning module is preset to learn, and the triggering condition is: the number of artificial checking results accumulated within a continuous preset time period reaches a preset threshold, or the model recognition accuracy is lower than a preset accuracy threshold; the online self-learning module uses an adaptive optimization algorithm combined with an early stopping strategy for incremental training, and only updates the model weight parameters.

9. The method of claim 8, wherein, The model performance verification of the online self-learning module uses a gray release method, applies the updated model to a preset proportion of traffic for a preset time period, and if the performance indicators improve to meet the preset requirements, the original model is replaced; at the same time, the original model is retained for a preset time period for abnormal rollback.

10. An after-service device for insurance claims, characterized in that The device comprises: at least one processor; and a memory connected in communication with the at least one processor; wherein The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to complete the insurance claim business post-survey method of any one of claims 1-9.

Images