Insurance claim settlement automatic processing method and system based on artificial intelligence

Abstract

The invention relates to the technical field of insurance, and particularly discloses an insurance claim settlement automatic processing method and system based on artificial intelligence. According to the method, claim settlement request data is automatically screened through artificial intelligence and divided into structured data (insurance policy numbers, amounts and the like) and unstructured data (photos, texts and the like), a basic insurance difference value between insurance information and historical records is calculated based on the structured data, accident site process nodes are extracted from the unstructured data, and the accident site process nodes are extracted from the unstructured data. The analysis processing step is used for normatively obtaining a flow node difference value, comparing traffic and alarm records with a declaration text to obtain a text difference value, calculating a comprehensive abnormal risk difference value by integrating the three difference values, if the comprehensive abnormal risk difference value exceeds a threshold value, turning to manual auditing, and otherwise, automatically settling claims. And the automation efficiency and the insurance fraud risk accurate identification are considered.

Description

Technical Field

[0001] This invention relates to the field of insurance technology, and in particular to an automated insurance claims processing method and system based on artificial intelligence. Background Technology

[0002] As a key part of insurance services, insurance claims directly affect the operational efficiency, risk control capabilities, and customer satisfaction of insurance companies. Traditional insurance claims processes rely heavily on manual processing, requiring claims specialists to manually review various materials submitted by applicants. Secondly, with the rapid growth of the insurance business, insurance fraud has also emerged during the insurance claims process. However, the current identification of insurance fraud risks mainly relies on the experience of individual claims specialists, and manual review is inevitably prone to oversights. For meticulously planned insurance fraud, especially those cases that exploit information asymmetry and create false information at multiple stages, there is a lack of systematic and quantitative detection methods, resulting in a high rate of missed detection of insurance fraud risks. Therefore, an automated insurance claims processing method based on artificial intelligence is needed to solve the above problems. Summary of the Invention

[0003] The purpose of this invention is to provide an automated insurance claims processing method and system based on artificial intelligence, so as to solve the technical problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an automated insurance claims processing method based on artificial intelligence, comprising: The system obtains the applicant's insurance claim request data and filters the insurance claim request data based on artificial intelligence to obtain filtered insurance claim request data, wherein the filtered insurance claim request data includes structured data and unstructured data. Based on the structured data, obtain basic insurance data and historical claims record data, and obtain basic insurance difference values ​​based on the basic insurance data and the historical claims record data; Accident scene data and text description data are obtained based on the unstructured data, and on-site processing flow node data are obtained based on the accident scene data, and the difference value of each processing flow node is obtained based on the on-site processing flow node data. Traffic record data and police report data are obtained based on the text description data, and text difference values ​​are obtained based on the traffic record data and the police report data; Calculate the comprehensive abnormal risk difference value based on the basic insurance difference value, the difference values ​​of multiple processing node values, and the text difference value; Determine whether the overall abnormal risk difference value is greater than a preset threshold; If the comprehensive abnormal risk difference value is greater than the preset threshold, it is determined that the applicant has the risk of insurance fraud, and a manual verification instruction is generated. The insurance claim request data is manually reviewed according to the manual verification instruction. If the overall abnormal risk difference value is not greater than the preset threshold, the corresponding claim amount is obtained based on the basic insurance data, and the insurance claim is automatically processed for the applicant based on the claim amount.

[0005] Preferably, the step of filtering the insurance claim request data based on artificial intelligence to obtain the filtered insurance claim request data includes: The insurance claim request data is subjected to data standardization preprocessing to obtain initial normalized data, wherein the data standardization preprocessing includes missing value imputation and outlier removal; The initial normalized data is input into a pre-trained multimodal data classification model to obtain data completeness scores and data type labels, wherein the data type labels include structured data labels and unstructured data labels; Determine whether the data completeness score is lower than a preset completeness threshold; If the data completeness score is lower than the preset completeness threshold, a data completion instruction is generated and sent to the applicant's terminal; If the data completeness score is not lower than the preset completeness threshold, the initial normalized data is divided into structured data and unstructured data according to the data type label, and the structured data and the unstructured data are combined to form the data for screening insurance claim requests.

[0006] Preferably, the step of obtaining the basic insurance difference value based on the basic insurance data and the historical claims record data includes: Based on the aforementioned basic insurance data, key insurance parameters are obtained, including policy duration, type of insurance, insured amount, and insured's age. Based on the historical claims record data, cosine similarity matching is performed on the policy effective duration, the type of insurance, the insured amount, and the age of the insured to obtain multiple historical related matching claims record data; The first K principal components are extracted from multiple historical related matching claim records to obtain multiple principal component claim data, wherein the principal component claim data includes the first historical claim count, the first average claim amount, and the first most recent claim time interval; Based on the aforementioned basic insurance data, the applicant's second historical claim count, second average claim amount, and second most recent claim interval are obtained. Based on the second historical claim count, second average claim amount, second most recent claim interval, multiple first historical claim counts, multiple first average claim amounts, and multiple first most recent claim intervals, a basic insurance difference value is obtained.

[0007] Preferably, the step of obtaining on-site processing flow node data based on the accident scene data, and obtaining the difference value of each processing flow node based on the on-site processing flow node data, includes: Image recognition and time-series analysis are performed on the accident scene data to obtain key processing flow nodes and on-site processing step sequences. The key processing flow nodes include alarm time point, traffic police arrival time point, on-site investigation completion time point, and vehicle towing time point. Obtain standard incident handling data, and based on the standard incident handling data, obtain the standard time interval and processing step constraint sequence between each key processing flow node; Based on the alarm time point, the traffic police arrival time point, the on-site investigation completion time point, and the vehicle towing time point, multiple time intervals between each other are obtained sequentially. Among them, the multiple time intervals include the alarm-traffic police arrival time interval, the traffic police arrival-on-site investigation completion time interval, and the on-site investigation completion-vehicle towing time interval. The multiple standard time intervals are compared sequentially with the alarm-traffic police arrival time interval, the traffic police arrival-on-site investigation completion time interval, and the on-site investigation completion-vehicle towing time interval to obtain multiple time node difference values; The step deviation coefficient is obtained based on the on-site processing step sequence and the processing step constraint sequence, and the difference value of each processing flow node is obtained based on the step deviation coefficient and the difference values ​​of multiple time nodes.

[0008] Preferably, the step of obtaining the text difference value based on the traffic record data and the police report record data includes: The traffic record data is parsed and key information is extracted to obtain a first accident description element set, wherein the first accident description element set includes the first accident time, the first accident location, the first vehicle identifier, and the first liability determination result; The reported incident data is parsed and key information is extracted to obtain a second accident description element set, which includes the second accident time, the second accident location, the second vehicle identifier, and the second liability determination result. The first set of accident description elements and the second set of accident description elements are matched to obtain multiple matching element pairs. Multiple local differences are obtained based on multiple matching element pairs, and an average local difference is obtained based on the multiple local differences. A standard local difference is calculated based on the multiple local difference degrees and the average local difference degree, and a local difference coefficient is calculated based on the standard local difference degree and the average local difference degree, and the local difference coefficient is used as the text difference value.

[0009] Preferably, the step of calculating the comprehensive abnormal risk difference value based on the basic insurance difference value, the difference values ​​of multiple processing node values, and the text difference value includes: The basic insurance difference value, the difference values ​​of multiple processing nodes, and the text difference value are normalized to obtain the basic insurance normalized difference value, the normalized difference values ​​of multiple processing nodes, and the text normalized difference value. The basic abnormal risk difference value is obtained by weighting the basic insurance normalized difference value, the normalized difference values ​​of multiple processing nodes, and the text normalized difference value. Obtain the risk adjustment factor for the current claims business scenario; The adjusted abnormal risk difference value is obtained based on the risk adjustment factor and the basic abnormal risk difference value, and the adjusted abnormal risk difference value is used as the comprehensive abnormal risk difference value.

[0010] This application also provides an automated insurance claims processing system based on artificial intelligence, including: The first acquisition module is used to acquire the applicant's insurance claim request data, and to filter the insurance claim request data based on artificial intelligence to obtain filtered insurance claim request data, wherein the filtered insurance claim request data includes structured data and unstructured data. The second acquisition module is used to acquire basic insurance data and historical claims record data based on the structured data, and to acquire basic insurance difference values ​​based on the basic insurance data and the historical claims record data. The third acquisition module is used to acquire accident scene data and text description data based on the unstructured data, acquire on-site processing flow node data based on the accident scene data, and acquire the difference value of each processing flow node based on the on-site processing flow node data. The fourth acquisition module is used to acquire traffic record data and police report record data based on the text description data, and to acquire text difference values ​​based on the traffic record data and the police report record data; The calculation module is used to calculate the comprehensive abnormal risk difference value based on the basic insurance difference value, the difference values ​​of multiple processing flow nodes, and the text difference value; The judgment module is used to determine whether the comprehensive abnormal risk difference value is greater than a preset threshold; If the comprehensive abnormal risk difference value is greater than the preset threshold, it is determined that the applicant has the risk of insurance fraud, and a manual verification instruction is generated. The insurance claim request data is manually reviewed according to the manual verification instruction. If the overall abnormal risk difference value is not greater than the preset threshold, the corresponding claim amount is obtained based on the basic insurance data, and the insurance claim is automatically processed for the applicant based on the claim amount.

[0011] Preferably, the first acquisition module includes: The first acquisition unit is used to perform data standardization preprocessing on the insurance claim request data to obtain initial normalized data, wherein the data standardization preprocessing includes missing value imputation and outlier removal; The training unit is used to input the initial normalized data into a pre-trained multimodal data classification model to obtain data completeness scores and data type labels, wherein the data type labels include structured data labels and unstructured data labels. The judgment unit is used to determine whether the data completeness score is lower than a preset completeness threshold; If the data completeness score is lower than the preset completeness threshold, a data completion instruction is generated and sent to the applicant's terminal; If the data completeness score is not lower than the preset completeness threshold, the initial normalized data is divided into structured data and unstructured data according to the data type label, and the structured data and the unstructured data are combined to form the data for screening insurance claim requests.

[0012] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0013] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0014] The beneficial effects of this application are as follows: This invention automatically filters claim request data through artificial intelligence, dividing it into structured data (policy number, amount, etc.) and unstructured data (photos, text, etc.). Based on the structured data, it calculates the basic insurance difference value between the insurance information and historical records. It extracts accident scene process nodes from the unstructured data, analyzes the standardization of processing steps to obtain process node difference values, and compares traffic and alarm records with the declaration text to obtain text difference values. It calculates the comprehensive abnormal risk difference value by combining the three difference values. If it exceeds the threshold, it is transferred to manual review; otherwise, it is automatically processed. This method achieves multi-dimensional risk quantification, taking into account both automation efficiency and accurate identification of insurance fraud risks. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a method flow according to an embodiment of this application.

[0016] Figure 2 This is a schematic diagram of the system structure according to an embodiment of this application.

[0017] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of this application.

[0018] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] like Figures 1-3 As shown, this application provides an automated insurance claims processing method based on artificial intelligence, including: S1. Obtain the applicant's insurance claim request data, and filter the insurance claim request data based on artificial intelligence to obtain filtered insurance claim request data, wherein the filtered insurance claim request data includes structured data and unstructured data; S2. Obtain basic insurance data and historical claims record data based on the structured data, and obtain the basic insurance difference value based on the basic insurance data and the historical claims record data; S3. Obtain accident scene data and text description data based on the unstructured data, obtain on-site processing flow node data based on the accident scene data, and obtain the difference value of each processing flow node based on the on-site processing flow node data. S4. Obtain traffic record data and police report record data based on the text description data, and obtain text difference values ​​based on the traffic record data and the police report record data; S5. Calculate the comprehensive abnormal risk difference value based on the basic insurance difference value, the difference values ​​of multiple processing nodes, and the text difference value; S6. Determine whether the comprehensive abnormal risk difference value is greater than a preset threshold; If the comprehensive abnormal risk difference value is greater than the preset threshold, it is determined that the applicant has the risk of insurance fraud, and a manual verification instruction is generated. The insurance claim request data is manually reviewed according to the manual verification instruction. If the overall abnormal risk difference value is not greater than the preset threshold, the corresponding claim amount is obtained based on the basic insurance data, and the insurance claim is automatically processed for the applicant based on the claim amount.

[0021] As described in steps S1-S6 above, existing methods for identifying insurance fraud risks primarily rely on the experience of individual claims adjusters. However, manual review inevitably introduces oversights. For meticulously planned insurance fraud, especially cases exploiting information asymmetry and creating false information at multiple stages, there is a lack of systematic and quantitative detection methods, leading to a high rate of missed detections. This invention first obtains the applicant's insurance claim request data and then uses artificial intelligence to filter and submit this data, resulting in filtered insurance claim request data. This filtered insurance claim request data includes structured data... The data includes both structured and unstructured data. Claim request data received from the applicant's terminal (such as an insurance company's app or mini-program) encompasses both structured data (such as policy number, insured amount, insured's age, and other directly quantifiable information) and unstructured data (such as accident scene photos / videos, scanned copies of paper documents, and text descriptions, which cannot be directly stored in tables). Because the claim data submitted by applicants often has issues such as missing information (e.g., omitting the "accident location"), anomalies (e.g., claim amount far exceeding the insured amount), or inconsistent formatting (e.g., embedding text descriptions into tables), the accuracy of subsequent analysis is directly affected. This step uses AI-automated filtering to quickly complete / remove problematic data and accurately segment data types, laying the foundation for subsequent differentiated analysis where "structured data focuses on insurance information, and unstructured data focuses on accident details." Secondly, based on the structured data, basic insurance application data and historical claims records are obtained. Then, a basic insurance application difference value is calculated based on these two data. Since insurance fraud often involves characteristics such as "frequent claims within a short period after insurance purchase" and "significant discrepancies between claim amounts and insurance amounts," these characteristics can only be detected by comparing insurance information with historical records. This step, by quantifying the differences, can quickly identify risks related to "abnormal insurance parameters." Then, accident scene data and text description data are obtained based on the unstructured data. On-site processing flow node data is obtained based on the accident scene data, and the difference value of each processing flow node is obtained based on the on-site processing flow node data. Accident scene data (such as on-site photos, videos, and timestamp records) and text description data (such as the text of the accident process filled in by the applicant) are separated from the unstructured data. Next, through image recognition (such as extracting the "traffic police arrive" scene from the video) and time sequence analysis (such as parsing the shooting time in the EXIF ​​information of the photo), key nodes such as the alarm time, the traffic police arrival time, the on-site investigation completion time, and the vehicle towing time are determined, forming a sequence of on-site processing steps. Furthermore, the standardization of the accident processing flow is the key to judging authenticity. Insurance fraudsters often forge "false time nodes" (such as taking on-site photos in advance to forge the alarm time) or "skip key steps" (such as applying for compensation directly without reporting the incident). This step transforms abstract "process standardization" into quantifiable discrepancy values ​​using unstructured on-site data, enabling accurate identification of insurance fraud involving "process anomalies." For example, in a traffic accident claim, the accident scene video shows "10:00 call to police, 11:30 traffic police arrival," while the local standard time interval between calling the police and the arrival of traffic police is 40 minutes. Furthermore, the on-site handling sequence is "taking photos, applying for compensation, calling the police," which conflicts with the standard constraint sequence of "calling the police, traffic police arrival, investigation." Calculations show that the "difference value between calling the police and the arrival of traffic police" and the "step deviation coefficient" for this claim are both high, indicating a process anomaly. Subsequently, traffic record data and police report data are obtained based on the text description data. Text difference values ​​are then calculated based on these data. First, the traffic management department's traffic record system (e.g., surveillance footage, road traffic data) and police report system are connected via API to obtain official record data related to the current accident. Then, Natural Language Processing (NLP) technology is used to parse the traffic records and police report records, extracting the first and second accident description element sets (both containing accident time, location, vehicle identifier, and liability determination results). The two sets of elements are then aligned and matched (e.g., matching "first accident time" with "second accident time"), calculating the local difference degree for each element, and then calculating the local difference coefficient using the average difference degree and standard difference degree as the text difference value. Since insurance fraudsters often commit fraud by tampering with text information such as "accident time" and "liability determination," it is difficult to simultaneously forge third-party official records such as traffic and police reports. This step, by comparing the applicant's associated "third-party text records" with the "declaration text information," transforms "information conflict" into a quantitative difference value, effectively identifying "text-forged" insurance fraud. Next, a comprehensive abnormal risk difference value is calculated based on the basic insurance difference value, the difference values ​​of multiple processing node points, and the text difference value. A single-dimensional difference value cannot fully reflect the risk (e.g., "high difference at processing node points but low difference in text" may indicate operational error rather than insurance fraud). This step integrates multi-dimensional risk points into a single comprehensive indicator, achieving a comprehensive and objective quantification of insurance fraud risk. Finally, it is determined whether the comprehensive abnormal risk difference value is greater than a preset threshold; If the comprehensive abnormal risk difference value is greater than the preset threshold, it is determined that the applicant has the risk of insurance fraud, and a manual verification instruction is generated. The insurance claim request data is manually reviewed according to the manual verification instruction. If the overall abnormal risk difference value is not greater than the preset threshold, the corresponding claim amount is obtained based on the basic insurance data, and the insurance claim is automatically processed for the applicant based on the claim amount. This step is the final connection point between "automated processing" and "risk control": it realizes fully automated claims processing for low-risk applications, greatly improving processing efficiency; and it accurately guides manual review for high-risk applications, which not only avoids the inefficiency of comprehensive manual review, but also helps specialists focus on key risk points and reduce the rate of missed detection through "difference details prompts".

[0022] In one embodiment, step S1, which uses artificial intelligence to filter the insurance claim request data to obtain filtered insurance claim request data, includes: S101. Perform data standardization preprocessing on the insurance claim request data to obtain initial normalized data, wherein the data standardization preprocessing includes missing value imputation and outlier removal; S102. Input the initial normalized data into a pre-trained multimodal data classification model to obtain data completeness scores and data type labels, wherein the data type labels include structured data labels and unstructured data labels; S103. Determine whether the data completeness score is lower than a preset completeness threshold; If the data completeness score is lower than the preset completeness threshold, a data completion instruction is generated and sent to the applicant's terminal; If the data completeness score is not lower than the preset completeness threshold, the initial normalized data is divided into structured data and unstructured data according to the data type label, and the structured data and the unstructured data are combined to form the data for screening insurance claim requests.

[0023] As described in steps S101-S103 above, the present invention first performs data standardization preprocessing on the insurance claim request data to obtain initial standardized data. This data standardization preprocessing includes missing value imputation and outlier removal. The insurance claim request data originates from all claim-related data submitted by the applicant through official insurance company channels (APP, mini-program, offline counter scan upload), covering structured data (such as policy information filled in on forms) and unstructured data (such as uploaded images and text). A "type-based imputation strategy" is adopted to imput numerical fields in structured data (such as...) For fields such as insured amount and insured age, the mean of data from the same insurance type and age group is used for imputation. For categorical fields (such as insurance type and accident type), the mode is used for imputation. For missing key information in unstructured data (such as missing shooting time in on-site photos), it is deduced through correlation data (such as extracting time from alarm records), and blurry images are removed using image clarity algorithms (such as edge detection operators to calculate clarity scores). For logically conflicting data (such as "claims filed before policy activation"), it is directly identified as an anomaly and removed. Missing and outlier values ​​in the original claims data are the main causes of errors in subsequent analysis. This step, through automated preprocessing, can quickly repair missing information that can be completed and remove abnormal data that cannot be used for analysis, ensuring data usability from the source and avoiding the problem of "using incorrect data for incorrect analysis." Secondly, the initial normalized data is input into a pre-trained multimodal data classification model to obtain data completeness scores and data type labels. The data type labels include structured data labels and unstructured data labels. The multimodal data classification model adopts a hybrid architecture of "CNN+Transformer" to adapt to multiple data types. For image-type unstructured data (such as on-site photos and scanned documents), spatial features are extracted through 3 convolutional layers (3×3 kernel size) and 2 pooling layers (max pooling, 2×2 window). Secondly, for text-type data (structured form text)... Unstructured incident description: Semantic features are extracted through the encoder layer of Transformer (6 encoder layers, 512 hidden layer dimensions). Finally, image features and text features are weighted and fused through an attention mechanism. The output layer is divided into two branches: one branch outputs a completeness score of 0-100 through the sigmoid function, and the other branch outputs the data type label (structured / unstructured) through the softmax function. Although the initial normalized data has repaired missing and abnormal data, there are still questions about "whether the data is sufficient to support subsequent analysis" (completeness) and "which analysis module the data should enter" (type). This step uses a multimodal model to achieve "quality scoring" and "type positioning" of mixed data, providing a quantitative basis for subsequent screening decisions. For example, the initial normalized data includes "policy number, insured amount, blurry scene photos, and clear accident description text." After being input into the model, the model determines that the insufficient clarity of the scene photos leads to the loss of key information, resulting in a completeness score of 70. At the same time, "policy number and insured amount" are labeled with "structured data," while "scene photos and accident description text" are labeled with "unstructured data." Finally, it is determined whether the data completeness score is lower than a preset completeness threshold; If the data completeness score is lower than a preset completeness threshold, a data completion instruction is generated and sent to the applicant's terminal. Furthermore, low-completeness data is not entirely useless; most can be improved by completing key information. This step, by accurately identifying completion needs and proactively pushing instructions, reduces the burden on applicants of repeatedly submitting full data while efficiently fixing data defects and improving data utilization. If the data completeness score is not lower than the preset completeness threshold, the initial normalized data is divided into structured data and unstructured data according to the data type label. The structured data and the unstructured data are combined to form the data for screening insurance claim requests. The completeness score needs to be converted into an executable decision signal through the threshold. This step, by setting scientific thresholds, clearly distinguishes between "usable data" and "data requiring completion," avoiding misjudgments due to "forced analysis despite insufficient data completeness," and also preventing inefficiency caused by "excessive pursuit of completeness." Specifically, the system pre-defines a "tag-data type" mapping relationship. Data labeled with "structured data tags" is stored in a relational database, categorized into subcategories such as "policy information" and "applicant information." Data labeled with "unstructured data tags" is stored in a non-relational database, categorized into subcategories such as "image data" and "text data." Furthermore, it adds unique association identifiers (such as claim application numbers) to the two types of data after segmentation, ensuring that subsequent analysis can accurately match structured and unstructured data for the same claim case. Ultimately, this results in filtered insurance claim request data containing "structured data set + unstructured data set + association identifier." In summary, this step achieves accurate classification, storage, and association of the two types of data through label segmentation, providing data support for subsequent multi-dimensional difference value extraction with "on-demand retrieval and precise matching" and avoiding chaotic data retrieval.

[0024] In one embodiment, step S2, which involves obtaining the basic insurance difference value based on the basic insurance data and the historical claims record data, includes: S201. Obtain key insurance parameters based on the aforementioned basic insurance data, including policy duration, type of insurance, insured amount, and age of the insured. S202. Based on the historical claims record data, perform cosine similarity matching on the policy effective duration, the type of insurance, the insured amount, and the age of the insured to obtain multiple historical related matching claims record data. S203. Extract the first K principal components from the multiple historical related matching claim record data to obtain multiple principal component claim data, wherein the principal component claim data includes the first historical claim count, the first average claim amount, and the first most recent claim time interval; S204. Obtain the applicant's second historical claim count, second average claim amount, and second most recent claim time interval based on the aforementioned basic insurance data, and obtain the basic insurance difference value based on the second historical claim count, second average claim amount, second most recent claim time interval, multiple first historical claim counts, multiple first average claim amounts, and multiple first most recent claim time intervals.

[0025] As described in steps S201-S204 above, this invention first obtains key insurance parameters based on the basic insurance data, including policy validity period, insurance type, insured amount, and insured's age. The basic insurance data originates from the insurance company's core business database and is structured data entered and stored by the applicant during the insurance application stage. This data is retrieved through structured data association via "screening insurance claim request data." The key insurance parameters are the core link between "current applications" and "historical cases," and their comprehensiveness and accuracy directly determine the quality of subsequent matching and difference calculations. This step extracts four core related parameters through standardization, covering the "time dimension" (validity period), "product dimension" (insured type), "amount dimension" (insured amount), and "subject dimension" (insured's age), while also achieving a unified parameter format, laying the foundation for multi-dimensional similarity matching. Then, based on the historical claims record data, cosine similarity matching is performed on the policy effective duration, the type of insurance, the insured amount, and the insured's age to obtain multiple historically relevant matching claims record data. This historical claims record data originates from the insurance company's historical claims database and contains structured data of all processed claims cases, covering fields corresponding to the current key parameters such as "policy effective duration, type of insurance, insured amount, and insured's age." Next, feature vectors are constructed by combining the current key insurance parameters (60 days, 2 (type of insurance), 50,000 yuan, 34 years old) with the corresponding parameters of each historical case. Numerical parameters (effective duration, insured amount, age) are normalized (Min-Max method, mapped to the 0-1 interval), while maintaining the classification code (type of insurance) unchanged. Then, the cosine similarity formula is used to calculate the similarity between the current vector and the historical case vector. The formula is as follows: In this context, A represents the current application feature vector (policy effective duration feature vector, insurance type feature vector, insured amount feature vector, and insured age feature vector), and B represents the historical case feature vector. The similarity range is 0-1 (the closer to 1, the higher the matching degree). Traditional single-dimensional matching cannot guarantee the overall similarity between historical cases and the current application. However, cosine similarity can achieve multi-parameter collaborative matching, ensuring that the selected historical cases are close to the current application in the four dimensions of "time-product-amount-subject". This provides a comparable reference benchmark for subsequent difference calculations and avoids misjudgment of differences due to "mismatch of benchmark cases". Secondly, the first K principal components of multiple historically relevant matched claims records are extracted to obtain multiple principal component claims data. These principal component claims data include the first historical claim count, the first average claim amount, and the first most recent claim time interval. For example, the multi-dimensional claims features of 20 matched cases (including more than 10 fields such as claim count, claim amount per claim, claim time, and claim reason). Next, all numerical features are standardized (mean 0, standard deviation 1) to eliminate the influence of dimensions. Covariance matrix calculation: Construct the covariance matrix between features, and... To reflect the linear correlation between features, eigenvalues ​​and corresponding eigenvectors are obtained by performing eigenvalue decomposition on the covariance matrix. Then, the eigenvalues ​​are sorted from largest to smallest, and the top 3 principal components (K=3) are selected, whose cumulative variance contribution rate is over 85% (i.e., the top 3 principal components can retain more than 85% of the information in the original data). Finally, the original claims features of each matched case are mapped to the 3 principal component dimensions. After business interpretation, the 3 principal components correspond to "first historical number of claims", "first average claim amount", and "first most recent claim time interval", respectively, forming principal component claims data. In summary, historical matching cases have multiple and redundant claim feature dimensions (such as a high correlation between "total claim amount" and "average claim amount"). Direct comparison would increase computational complexity and may lead to duplicate assessments. This step uses PCA for dimensionality reduction, simplifying the multi-dimensional features into three key principal components while retaining core information. This reduces the complexity of subsequent difference calculations and focuses on the claim behavior features that have the greatest impact on risk assessment. Subsequently, based on the aforementioned basic insurance data, the applicant's second historical claim count, second average claim amount, and second most recent claim interval are obtained. Then, a basic insurance difference value is obtained based on the second historical claim count, second average claim amount, second most recent claim interval, multiple first historical claim counts, multiple first average claim amounts, and multiple first most recent claim intervals. Specifically, firstly, the corresponding averages are calculated sequentially based on multiple first historical claim counts, multiple first average claim amounts, and multiple first most recent claim intervals to obtain the average first historical claim count, average first average claim amount, and average first most recent claim interval. Then, the basic insurance difference value is calculated based on the second historical claim count, second average claim amount, second most recent claim interval, average first historical claim count, average first average claim amount, and average first most recent claim interval. The calculation formula is as follows: ; in, Based on the difference in basic insurance coverage, This is the second highest number of historical claims. This represents the average number of historical claims. for The weight value, The second average claim amount, This represents the average of the first average claim amount. for The weight value, The second most recent claim interval, This represents the average time interval between the most recent claim; Among them, "frequent claims" is one of the typical characteristics of insurance fraud (such as "maliciously creating minor accidents for repeated claims" and "using multiple policies to obtain claim funds"). It has the strongest correlation with the risk of insurance fraud. Therefore, the difference in the number of claims has the highest weight. Secondly, "the claim amount deviates significantly from the insured amount / average of similar cases" is an important signal of insurance fraud (such as "applying for high claims with low insured amount" and "claim amount far exceeding the actual loss of the accident"). However, some normal claims may also have abnormal amounts due to "special severity of the accident". Therefore, the correlation is slightly lower than that of the number of claims. Finally, "claims made in a short period of time after insurance" and "the interval between two claims is too short" may imply the risk of insurance fraud (such as malicious insurance behavior of "insurance immediately after insurance"). However, it may also be due to "concentrated occurrence of accidents" which shortens the normal interval. Its risk indication is weaker than the first two dimensions. In summary, this step, through multi-layered calculations of "absolute difference - relative difference - weighted summation," accurately compares the applicant's current claims behavior with the routine behavior of similar groups, transforming "whether it deviates from the norm" into a quantifiable difference value. This value directly reflects the degree of abnormality in the applicant's insurance application and claims behavior, providing a core quantitative basis for the "insurance dimension" in subsequent comprehensive risk assessment.

[0026] In one embodiment, step S3, which involves obtaining on-site processing flow node data based on the accident scene data and obtaining the difference value of each processing flow node based on the on-site processing flow node data, includes: S301. Perform image recognition and time sequence analysis on the accident scene data to obtain key processing flow nodes and on-site processing step sequence, wherein the key processing flow nodes include alarm time point, traffic police arrival time point, on-site investigation completion time point, and vehicle towing time point. S302. Obtain standard accident handling data, and obtain the standard time interval and processing step constraint sequence between each key processing flow node based on the standard accident handling data; S303. Based on the alarm time point, the traffic police arrival time point, the on-site investigation completion time point, and the vehicle towing time point, obtain multiple time intervals between each other in sequence, wherein the multiple time intervals include the alarm-traffic police arrival time interval, the traffic police arrival-on-site investigation completion time interval, and the on-site investigation completion-vehicle towing time interval; S304. The multiple standard time intervals are compared sequentially with the alarm-traffic police arrival time interval, the traffic police arrival-on-site investigation completion time interval, and the on-site investigation completion-vehicle towing time interval to obtain multiple time node difference values; S305. Obtain the step deviation coefficient based on the on-site processing step sequence and the processing step constraint sequence, and obtain the difference value of each processing flow node based on the step deviation coefficient and the difference values ​​of multiple time nodes.

[0027] As described in steps S301-S305 above, the present invention first performs image recognition and time sequence analysis on the accident scene data to obtain key processing flow nodes and on-site processing step sequences. The key processing flow nodes include alarm time point, traffic police arrival time point, on-site investigation completion time point, and vehicle towing time point. The accident scene data comes from unstructured claim data submitted by the applicant, including accident scene photos, videos, and on-site records with timestamps. The data is retrieved through unstructured data association by "screening insurance claim request data". Next, the video data is analyzed by frame sequence analysis. The time points of dynamic nodes such as "traffic police arrival" and "vehicle towing" are determined by the target appearance time sequence (taking the video timestamp of the first appearance of the target). Finally, the extracted time points are sorted in chronological order to obtain key processing flow nodes (alarm time point, traffic police arrival time point, on-site investigation completion time point, vehicle towing time point). Then, based on the target recognition results and the association with the time points, an on-site processing step sequence is constructed (e.g., "alarm, traffic police arrival, on-site investigation, vehicle towing"). In summary, accident scene data is mostly unstructured images / videos, making manual node extraction inefficient and prone to errors. This step, through automated extraction of nodes and steps, ensures the accuracy of key time points (error ≤ 5 minutes) and clarifies the logical sequence of processing steps, providing an accurate and complete "actual process benchmark" for subsequent comparison with standard procedures. Next, standard accident handling data is acquired, and based on this data, standard time intervals and processing step constraint sequences between each key processing node are obtained. This standard accident handling data originates from pre-defined insurance specifications and serves as a "reference benchmark" for determining whether the actual process is abnormal. This step ensures the fairness and relevance of subsequent comparisons by accurately retrieving standard data matching the current accident, avoiding the use of "major accident standards" to judge "minor accident processes" or "suburban road standards" to judge "urban main road processes," thus providing a scientific benchmark for difference calculations. Secondly, based on the alarm time, traffic police arrival time, on-site investigation completion time, and vehicle towing time, multiple time intervals are sequentially obtained between them. These multiple time intervals include the alarm-traffic police arrival time interval, the traffic police arrival-on-site investigation completion time interval, and the on-site investigation completion-vehicle towing time interval. The time interval calculation rule is as follows: it is automatically calculated using the method of "later node time - previous node time", with the unit uniformly in "minutes". If the later node time is earlier than the previous node time (such as the investigation completion time being earlier than the traffic police arrival time), it is marked as "logical anomaly", and the maximum difference value is assigned in subsequent calculations. The time interval is a core indicator reflecting the timeliness of the process and a key point that is easily tampered with in insurance fraud (such as falsifying the traffic police arrival time to shorten the interval). This step transforms the "process speed" into a quantifiable value by standardizing the calculation of actual time intervals, laying the foundation for subsequent comparison with standard intervals. Simultaneously, multiple standard time intervals are compared sequentially with the alarm-traffic police arrival time interval, the traffic police arrival-on-site investigation completion time interval, and the on-site investigation completion-vehicle towing time interval to obtain multiple time node difference values. Among them, the "relative difference rate" is used to calculate the time node difference value, and the formula is: time node difference value = |actual time interval - standard time interval| / standard time interval, and the result range is 0 to +∞ (the larger the value, the more serious the time deviation). If the actual time interval is 0 or negative (logical anomaly), the time node difference value is directly set to 2.0 (corresponding to "serious anomaly", based on historical data verification, this value far exceeds the normal difference range). In summary, by calculating the relative difference rate, the unfairness in difference assessment caused by "different standard intervals for different nodes" can be eliminated (e.g., a "deviation of 10 minutes" is 33% different from a 30-minute standard interval, but 50% different from a 20-minute standard interval; the relative difference rate can accurately reflect the degree of deviation). This step quantifies the "degree of time deviation" into a difference value, enabling accurate assessment of abnormal process timeliness. Finally, the step deviation coefficient is obtained based on the on-site processing step sequence and the processing step constraint sequence. The step deviation coefficient is calculated as follows: the similarity between the actual step sequence and the constraint sequence is calculated using the "edit distance algorithm": the edit distance is the minimum number of insertion, deletion, and replacement operations required to convert the actual sequence into the constraint sequence, similarity = 1 - (edit distance / constraint sequence length), and step deviation coefficient = 1 - similarity. The result range is 0-1 (the larger the value, the more serious the step logic deviation). If there are missing key steps (such as the absence of the "on-site investigation" step), it is directly set to 1.0. The difference value of each processing node is obtained based on the step deviation coefficient and the difference values ​​of multiple time nodes. Next, each key node corresponds to a "time interval difference value before the node" (e.g., "traffic police arrival node" corresponds to "alarm-traffic police arrival difference value"). Then, the difference value of each node is calculated using a weighted summation formula: Processing node difference value = (time node difference value × 0.6) + (step deviation coefficient × 0.4), where the time interval has a weight of 0.6 (timeliness has a greater impact on risk assessment), and the step logic has a weight of 0.4 (logic anomalies are the core risk signal). In summary, insurance fraud may manifest only as time anomalies (such as delayed reporting), or only as procedural anomalies (such as skipping an on-site investigation), or both. This step, by integrating "time deviation" and "logical deviation," achieves a comprehensive quantification of the degree of anomaly at each process node, avoiding missed risk assessments due to single-dimensional evaluation.

[0028] In one embodiment, step S4, which involves obtaining text difference values ​​based on the traffic record data and the police report data, includes: S401. Perform text parsing and key information extraction on the traffic record data to obtain a first accident description element set, wherein the first accident description element set includes the first accident time, the first accident location, the first vehicle identifier, and the first liability determination result; S402. Perform text parsing and key information extraction on the reported incident data to obtain a second accident description element set, wherein the second accident description element set includes the second accident time, the second accident location, the second vehicle identifier, and the second liability determination result; S403. Perform element alignment matching between the first accident description element set and the second accident description element set to obtain multiple matching element pairs; S404. Obtain multiple local difference degrees based on multiple matching element pairs, and obtain an average local difference degree based on the multiple local difference degrees; S405. Calculate a standard local difference degree based on the multiple local difference degrees and the average local difference degree, calculate a local difference degree coefficient based on the standard local difference degree and the average local difference degree, and use the local difference degree coefficient as the text difference value.

[0029] As described in steps S401-S405 above, the present invention first performs text parsing and key information extraction on the traffic record data to obtain a first accident description element set. This first accident description element set includes the first accident time, the first accident location, the first vehicle identifier, and the first liability determination result. The traffic record data originates from the local traffic management department's business system and is retrieved through the API interface between the insurance company and the traffic management department. It is unstructured text data (such as OCR text of traffic monitoring event records or scanned copies of handwritten records from traffic management personnel at the accident scene), and can be associated with the text description data through "filtering insurance claim request data". First, for structured text (such as tabular records exported from traffic management systems), extract directly by field; for unstructured text (such as OCR results of handwritten records), first perform text cleaning (remove garbled characters, correct recognition errors), and then standardize the elements: accident time: uniformly formatted as "YYYY-MM-DDHH:MM:SS" (e.g., standardizing "May 10, 2024 at 9:00 AM" to "2024-05-10 09:00:00"); accident location: linked to the Gaode Map API to convert to latitude and longitude coordinates (e.g., "intersection of XX Road and YY Street" converted to "116.404267,39.915119"). In summary, traffic record data is an important and authoritative basis for reflecting the objective situation of accidents. However, the unstructured text format makes it difficult to extract key information efficiently. This step realizes the automated and standardized extraction of key elements, which not only ensures the accuracy and completeness of information extraction, but also transforms unstructured text into comparable structured elements, laying the foundation for accurate matching with subsequent accident reports. Next, the reported accident data is parsed and key information is extracted to obtain a second accident description element set. This second accident description element set includes the second accident time, the second accident location, the second vehicle identifier, and the second liability determination result. The reported accident data originates from the police system or accident reporting system and is retrieved through the API interface between the insurance company and the police department. The text types include unstructured data such as police call records and OCR text of the complainant's interview transcript. It is also located through the text description data association in claim 1, where "insurance claim request data is screened." Since the police report is one of the earliest authoritative records formed after an accident, its consistency with traffic records significantly enhances the verification of the accident's authenticity. This step extracts standardized elements through a unified technical process, avoiding misjudgments of text conflicts due to "differences in element format caused by different extraction rules," thus ensuring the objectivity of subsequent comparisons. Secondly, the first set of accident description elements and the second set of accident description elements are matched to obtain multiple matching element pairs. The matching rule is as follows: a "precise field name matching" strategy is adopted, pairing elements with completely identical field names in the first and second set of elements one by one to form matching element pairs. Specifically, these include: (first accident time, second accident time), (first accident location, second accident location), (first vehicle identifier, second vehicle identifier), (first liability determination result, second liability determination result). Missing element handling: if an element is missing in one type of record (e.g., the traffic record does not indicate the liability determination result), it is marked as "element missing", and the corresponding matching element pair is (valid element, missing). Furthermore, element alignment is a prerequisite for difference comparison. Only by matching elements of the same type can information conflicts be accurately identified. This step ensures that each pair of elements is "accident information of the same dimension" (e.g., both are "accident time") through standardized field name matching, avoiding invalid differences caused by cross-dimensional comparison (e.g., comparing "accident time" and "vehicle identifier"), and ensuring the relevance of the comparison. Simultaneously, multiple local differences are obtained based on multiple matching element pairs, and an average local difference is obtained based on the multiple local differences. Secondly, a standard local variability is calculated based on the multiple local variability values ​​and the average local variability value, wherein the calculation formula is: ; in, Indicates the standard local variability. This represents the l-th local difference. This indicates the number of local differences, where l represents the sequence number of the local differences. Represents the average local variability, where, =1, 2, 3...N, Local variability directly reflects the degree of conflict in a single element dimension, while average local variability reflects the overall conflict level of multiple elements. This step uses a differentiated calculation method to transform the "conflict magnitude" of different types of elements into a uniform range (0-1.0) of variability, which not only ensures the accurate quantification of individual element differences but also provides a basic statistical indicator for subsequent comprehensive variability calculations. The local variability coefficient is calculated based on the standard local variability and the average local variability, wherein the calculation formula is as follows: ; in, Represents the local difference coefficient. Indicates the average local variability. Indicates the standard local variability; And the local difference coefficient is used as the text difference value; While the average local difference can only reflect the "overall conflict level," it cannot reflect the "extreme conflict of a single element" (such as most elements being indistinguishable, but vehicle identifiers being completely inconsistent); the standard local difference can capture this kind of "discrete conflict." This step, by integrating the two, considers both the overall consistency of multiple elements and the extreme anomalies of individual elements, achieving a comprehensive and accurate quantification of text differences and avoiding the risk of missed judgments due to "low overall mean but conflict of key elements."

[0030] In one embodiment, step S5, which calculates the comprehensive abnormal risk difference value based on the basic insurance difference value, the difference values ​​of multiple processing node values, and the text difference value, includes: S501. Normalize the basic insurance difference value, the difference values ​​of multiple processing nodes and the text difference value to obtain the basic insurance normalized difference value, the normalized difference values ​​of multiple processing nodes and the text normalized difference value. S502. The basic insurance normalized difference value, the normalized difference values ​​of multiple processing nodes, and the text normalized difference value are weighted and calculated to obtain the basic abnormal risk difference value, wherein the calculation formula is: ,in, Based on the difference in basic abnormal risk, Based on the normalized difference value of basic insurance coverage, To handle the normalized difference values ​​of process nodes, Let be the mean of the sum of normalized differences from the i-th to the n-th processing node, where i is the number of normalized differences in the processing node, i = 1, 2, ..., n. For text normalization difference values, The weighting value of the normalized difference value for basic insurance coverage. The weight value is the average of the normalized difference values ​​of the i-th to n-th processing nodes. S503, Obtain the risk adjustment factor for the current claims business scenario; S504. Obtain the adjusted abnormal risk difference value based on the risk adjustment factor and the basic abnormal risk difference value, and use the adjusted abnormal risk difference value as the comprehensive abnormal risk difference value.

[0031] As described in steps S501-S504 above, this invention first normalizes the basic insurance difference value, the difference values ​​of multiple processing nodes, and the text difference value to obtain the basic insurance normalized difference value, the normalized difference values ​​of multiple processing nodes, and the text normalized difference value. The core of the normalization process is to eliminate dimensional differences, mapping the three different ranges of difference values ​​uniformly to the 0-1 interval, so that "basic insurance difference value 0.5" and "text difference value 0.5" have the same meaning of "relative anomaly level". This step provides a "fair and comparable" data foundation for subsequent weighted fusion, avoiding misjudgment of risk contribution due to different original ranges. Then, the basic insurance normalized difference value, the normalized difference values ​​of multiple processing node data points, and the text normalized difference value are weighted and calculated to obtain the basic abnormal risk difference value. The core of this weighted calculation is to quantify the "contribution weight" of the three types of differences to the risk of insurance fraud, avoiding subjective judgment. This step uses machine learning to mine weights from historical data, ensuring that the basic abnormal risk difference value truly reflects the "actual impact of different dimensions of abnormalities on risk." For example, process abnormalities have the highest weight, reflecting the business understanding that "falsifying accident handling processes is a core method of insurance fraud." Subsequently, the risk adjustment factor for the current claims business scenario is obtained. This scenario is based on four preset dimensions: weather, time, region, and insurance type. For example: Weather dimension: heavy rain / snowstorm (1.2), cloudy (1.0), sunny (0.9); Time dimension: holidays (1.2), weekdays (1.0), late night (1.1); Region dimension: high-incidence accident areas (1.1), ordinary areas (1.0), low-incidence areas (0.9); Insurance type dimension: high-risk insurance (e.g., engineering insurance, 1.2), ordinary insurance (e.g., compulsory traffic accident liability insurance, 1.0), low-risk insurance (e.g., accident insurance, 0.8). Next, the risk adjustment factor is calculated using the "multi-dimensional coefficient product method," multiplying the adjustment coefficients of each dimension under the current scenario to obtain the final risk adjustment factor. The calculation formula is as follows: ,in, As a risk adjustment factor for the current claims business scenario, As a weather factor, For time factor, regional factors, As a risk factor; The core of the risk adjustment factor is "dynamic adaptation to business scenarios," addressing the problem that "fixed standards cannot cope with scenario differences." This step integrates multi-dimensional scenario information to make risk assessment more aligned with actual business logic. For example, the probability of insurance fraud is higher during holidays, so factor 1.2 is used to amplify the base risk value and improve the sensitivity of risk identification; while the credibility of anomalies in low-risk insurance products is lower, so factor 0.8 is used to reduce the base risk value and decrease the probability of misjudgment. Finally, the adjusted abnormal risk difference value is obtained based on the risk adjustment factor and the basic abnormal risk difference value. The calculation steps are as follows: the risk adjustment factor and the basic abnormal risk difference value are multiplied to equal the adjusted abnormal risk difference value, and the adjusted abnormal risk difference value is used as the comprehensive abnormal risk difference value. In summary, this step serves as a crucial fusion point between "basic risk" and "scenario risk," transforming static basic risk values ​​into dynamic risk values ​​tailored to the current scenario through factor adjustments. The final output of the comprehensive abnormal risk difference value achieves the dual objectives of "multi-dimensional difference quantification + dynamic scenario adaptation," providing a precise and comparable core indicator for "threshold determination."

[0032] This application also provides an automated insurance claims processing system based on artificial intelligence, including: The first acquisition module is used to acquire the applicant's insurance claim request data, and to filter the insurance claim request data based on artificial intelligence to obtain filtered insurance claim request data, wherein the filtered insurance claim request data includes structured data and unstructured data. The second acquisition module is used to acquire basic insurance data and historical claims record data based on the structured data, and to acquire basic insurance difference values ​​based on the basic insurance data and the historical claims record data. The third acquisition module is used to acquire accident scene data and text description data based on the unstructured data, acquire on-site processing flow node data based on the accident scene data, and acquire the difference value of each processing flow node based on the on-site processing flow node data. The fourth acquisition module is used to acquire traffic record data and police report record data based on the text description data, and to acquire text difference values ​​based on the traffic record data and the police report record data; The calculation module is used to calculate the comprehensive abnormal risk difference value based on the basic insurance difference value, the difference values ​​of multiple processing flow nodes, and the text difference value; The judgment module is used to determine whether the comprehensive abnormal risk difference value is greater than a preset threshold; If the comprehensive abnormal risk difference value is greater than the preset threshold, it is determined that the applicant has the risk of insurance fraud, and a manual verification instruction is generated. The insurance claim request data is manually reviewed according to the manual verification instruction. If the overall abnormal risk difference value is not greater than the preset threshold, the corresponding claim amount is obtained based on the basic insurance data, and the insurance claim is automatically processed for the applicant based on the claim amount.

[0033] In one embodiment, the first acquisition module includes: The first acquisition unit is used to perform data standardization preprocessing on the insurance claim request data to obtain initial normalized data, wherein the data standardization preprocessing includes missing value imputation and outlier removal; The training unit is used to input the initial normalized data into a pre-trained multimodal data classification model to obtain data completeness scores and data type labels, wherein the data type labels include structured data labels and unstructured data labels. The judgment unit is used to determine whether the data completeness score is lower than a preset completeness threshold; If the data completeness score is lower than the preset completeness threshold, a data completion instruction is generated and sent to the applicant's terminal; If the data completeness score is not lower than the preset completeness threshold, the initial normalized data is divided into structured data and unstructured data according to the data type label, and the structured data and the unstructured data are combined to form the data for screening insurance claim requests.

[0034] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0035] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0036] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0037] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0038] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An automated insurance claims processing method based on artificial intelligence, characterized in that, include: The system obtains the applicant's insurance claim request data and filters the insurance claim request data based on artificial intelligence to obtain filtered insurance claim request data, wherein the filtered insurance claim request data includes structured data and unstructured data. Based on the structured data, obtain basic insurance data and historical claims record data, and obtain basic insurance difference values ​​based on the basic insurance data and the historical claims record data; Accident scene data and text description data are obtained based on the unstructured data, and on-site processing flow node data are obtained based on the accident scene data, and the difference value of each processing flow node is obtained based on the on-site processing flow node data. Traffic record data and police report data are obtained based on the text description data, and text difference values ​​are obtained based on the traffic record data and the police report data; Calculate the comprehensive abnormal risk difference value based on the basic insurance difference value, the difference values ​​of multiple processing node values, and the text difference value; Determine whether the overall abnormal risk difference value is greater than a preset threshold; If the comprehensive abnormal risk difference value is greater than the preset threshold, it is determined that the applicant has the risk of insurance fraud, and a manual verification instruction is generated. The insurance claim request data is manually reviewed according to the manual verification instruction. If the overall abnormal risk difference value is not greater than the preset threshold, the corresponding claim amount is obtained based on the basic insurance data, and the insurance claim is automatically processed for the applicant based on the claim amount.

2. The automated insurance claims processing method based on artificial intelligence according to claim 1, characterized in that, The step of filtering the insurance claim request data based on artificial intelligence to obtain the filtered insurance claim request data includes: The insurance claim request data is subjected to data standardization preprocessing to obtain initial normalized data, wherein the data standardization preprocessing includes missing value imputation and outlier removal; The initial normalized data is input into a pre-trained multimodal data classification model to obtain data completeness scores and data type labels, wherein the data type labels include structured data labels and unstructured data labels; Determine whether the data completeness score is lower than a preset completeness threshold; If the data completeness score is lower than the preset completeness threshold, a data completion instruction is generated and sent to the applicant's terminal; If the data completeness score is not lower than the preset completeness threshold, the initial normalized data is divided into structured data and unstructured data according to the data type label, and the structured data and the unstructured data are combined to form the data for screening insurance claim requests.

3. The automated insurance claims processing method based on artificial intelligence according to claim 1, characterized in that, The step of obtaining the basic insurance difference value based on the basic insurance data and the historical claims record data includes: Based on the aforementioned basic insurance data, key insurance parameters are obtained, including policy duration, type of insurance, insured amount, and insured's age. Based on the historical claims record data, cosine similarity matching is performed on the policy effective duration, the type of insurance, the insured amount, and the age of the insured to obtain multiple historical related matching claims record data; The first K principal components are extracted from multiple historical related matching claim records to obtain multiple principal component claim data, wherein the principal component claim data includes the first historical claim count, the first average claim amount, and the first most recent claim time interval; Based on the aforementioned basic insurance data, the applicant's second historical claim count, second average claim amount, and second most recent claim interval are obtained. Based on the second historical claim count, second average claim amount, second most recent claim interval, multiple first historical claim counts, multiple first average claim amounts, and multiple first most recent claim intervals, a basic insurance difference value is obtained.

4. The automated insurance claims processing method based on artificial intelligence according to claim 1, characterized in that, The step of obtaining on-site processing flow node data based on the accident scene data, and obtaining the difference value of each processing flow node based on the on-site processing flow node data, includes: Image recognition and time-series analysis are performed on the accident scene data to obtain key processing flow nodes and on-site processing step sequences. The key processing flow nodes include alarm time point, traffic police arrival time point, on-site investigation completion time point, and vehicle towing time point. Obtain standard incident handling data, and based on the standard incident handling data, obtain the standard time interval and processing step constraint sequence between each key processing flow node; Based on the alarm time point, the traffic police arrival time point, the on-site investigation completion time point, and the vehicle towing time point, multiple time intervals between each other are obtained sequentially. Among them, the multiple time intervals include the alarm-traffic police arrival time interval, the traffic police arrival-on-site investigation completion time interval, and the on-site investigation completion-vehicle towing time interval. The multiple standard time intervals are compared sequentially with the alarm-traffic police arrival time interval, the traffic police arrival-on-site investigation completion time interval, and the on-site investigation completion-vehicle towing time interval to obtain multiple time node difference values; The step deviation coefficient is obtained based on the on-site processing step sequence and the processing step constraint sequence, and the difference value of each processing flow node is obtained based on the step deviation coefficient and the difference values ​​of multiple time nodes.

5. The automated insurance claims processing method based on artificial intelligence according to claim 1, characterized in that, The step of obtaining text difference values ​​based on the traffic record data and the police report data includes: The traffic record data is parsed and key information is extracted to obtain a first accident description element set, wherein the first accident description element set includes the first accident time, the first accident location, the first vehicle identifier, and the first liability determination result; The reported incident data is parsed and key information is extracted to obtain a second accident description element set, which includes the second accident time, the second accident location, the second vehicle identifier, and the second liability determination result. The first set of accident description elements and the second set of accident description elements are matched to obtain multiple matching element pairs. Multiple local differences are obtained based on multiple matching element pairs, and an average local difference is obtained based on the multiple local differences. A standard local difference is calculated based on the multiple local difference degrees and the average local difference degree, and a local difference coefficient is calculated based on the standard local difference degree and the average local difference degree, and the local difference coefficient is used as the text difference value.

6. The automated insurance claims processing method based on artificial intelligence according to claim 1, characterized in that, The step of calculating the comprehensive abnormal risk difference value based on the basic insurance difference value, the difference values ​​of multiple processing node values, and the text difference value includes: The basic insurance difference value, the difference values ​​of multiple processing nodes, and the text difference value are normalized to obtain the basic insurance normalized difference value, the normalized difference values ​​of multiple processing nodes, and the text normalized difference value. The basic abnormal risk difference value is obtained by weighting the basic insurance normalized difference value, the normalized difference values ​​of multiple processing nodes, and the text normalized difference value. Obtain the risk adjustment factor for the current claims business scenario; The adjusted abnormal risk difference value is obtained based on the risk adjustment factor and the basic abnormal risk difference value, and the adjusted abnormal risk difference value is used as the comprehensive abnormal risk difference value.

7. An automated insurance claims processing system based on artificial intelligence, characterized in that, include: The first acquisition module is used to acquire the applicant's insurance claim request data, and to filter the insurance claim request data based on artificial intelligence to obtain filtered insurance claim request data, wherein the filtered insurance claim request data includes structured data and unstructured data. The second acquisition module is used to acquire basic insurance data and historical claims record data based on the structured data, and to acquire basic insurance difference values ​​based on the basic insurance data and the historical claims record data. The third acquisition module is used to acquire accident scene data and text description data based on the unstructured data, acquire on-site processing flow node data based on the accident scene data, and acquire the difference value of each processing flow node based on the on-site processing flow node data. The fourth acquisition module is used to acquire traffic record data and police report record data based on the text description data, and to acquire text difference values ​​based on the traffic record data and the police report record data; The calculation module is used to calculate the comprehensive abnormal risk difference value based on the basic insurance difference value, the difference values ​​of multiple processing flow nodes, and the text difference value; The judgment module is used to determine whether the comprehensive abnormal risk difference value is greater than a preset threshold; If the comprehensive abnormal risk difference value is greater than the preset threshold, it is determined that the applicant has the risk of insurance fraud, and a manual verification instruction is generated. The insurance claim request data is manually reviewed according to the manual verification instruction. If the overall abnormal risk difference value is not greater than the preset threshold, the corresponding claim amount is obtained based on the basic insurance data, and the insurance claim is automatically processed for the applicant based on the claim amount.

8. The automated insurance claims processing system based on artificial intelligence according to claim 7, characterized in that, The first acquisition module includes: The first acquisition unit is used to perform data standardization preprocessing on the insurance claim request data to obtain initial normalized data, wherein the data standardization preprocessing includes missing value imputation and outlier removal; The training unit is used to input the initial normalized data into a pre-trained multimodal data classification model to obtain data completeness scores and data type labels, wherein the data type labels include structured data labels and unstructured data labels. The judgment unit is used to determine whether the data completeness score is lower than a preset completeness threshold; If the data completeness score is lower than the preset completeness threshold, a data completion instruction is generated and sent to the applicant's terminal; If the data completeness score is not lower than the preset completeness threshold, the initial normalized data is divided into structured data and unstructured data according to the data type label, and the structured data and the unstructured data are combined to form the data for screening insurance claim requests.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images