Multi-modal data-driven intelligent claim settlement risk control system

Abstract

The application provides a multi-modal data-driven intelligent claim settlement risk control system, relates to the technical field of insurance services, and comprises a multi-modal data acquisition module, a multi-modal data fusion and correlation analysis module, a historical claim analysis module and a decision output module; the multi-modal data acquisition module acquires multi-modal data related to a vehicle accident claim settlement event; the multi-modal data fusion and correlation analysis module constructs a multi-modal fusion model and calculates a correlation index; the historical claim analysis module acquires metadata of a claim settlement event and calculates a claim settlement similarity index based on the metadata; the decision output module pre-sets a correlation threshold and a claim settlement similarity threshold, establishes an inspection mechanism, and determines whether to trigger a manual inspection process. The application realizes multi-dimensional evidence consistency verification and space-time context risk perception capability that cannot be achieved by a conventional system, locks high-risk cases through a double-threshold joint decision mechanism, and automatically triggers manual inspection, effectively solving the limitations of traditional risk control systems, such as dependence on a single data source, difficulty in identifying complex fraud patterns, and lack of dynamic adaptability.

Description

Technical Field

[0001] This invention relates to the field of insurance service technology, specifically to a multimodal data-driven intelligent claims risk control system. Background Technology

[0002] In the process of risk control for claims related to vehicle accidents, verifying whether there are interrelationships between the relevant original data of the claim event can effectively identify logical contradictions and traces of forgery that are difficult to expose by a single data source (such as discrepancies between on-site images and sensor acceleration data, or contradictions between the victim's account and third-party environmental data). This allows for the construction of a closed-loop evidence chain from physical evidence to behavioral descriptions, improving the ability to identify complex fraud methods. On the other hand, conducting lateral analysis of historical normal claim events at the accident scene can establish a regional baseline for normal claims, thereby identifying abnormal cluster risks. This provides crucial background information for determining whether the current event constitutes group fraud or insurance fraud, realizing a shift from single-event review to scenario-based risk prediction.

[0003] In the prior art, CN120125355A discloses a vehicle mutual assistance claims method and system based on multimodal data fusion. This technology includes: integrating visual and depth sensing data and repair records from vehicle accidents; extracting component types and collision mechanical parameters; capturing surface micro-deformation and internal hidden cracks through dual-band scanning; and constructing a comprehensive damage map by spatiotemporally correlating the two. Based on the normal damage patterns of components, a standard damage sequence is generated using an adversarial generative network and compared layer by layer with accident damage data to extract abnormal regions that deviate from the evolutionary pattern. A fraud risk index is calculated based on its spatial density. When the index exceeds a dynamic threshold, multi-source data collaborative verification of feature consistency is triggered, and a claims result with a fraud label is output. This technical solution can achieve accurate identification of repair fraud and synergistic optimization of claims risk control efficiency.

[0004] However, the aforementioned existing technologies rely on a single data source, depending solely on visual and depth sensor data and maintenance records from vehicle accidents. They lack comprehensive collection and analysis of third-party environmental information, making it difficult to verify the consistency of images, text, sensor data, and third-party environmental data. Furthermore, existing technologies assume all collected evidence is completely reliable and valid, ignoring potential contradictions and low correlations within the original data. This prevents the construction of a closed chain of evidence from multiple dimensions to identify complex fraudulent methods. Simultaneously, conventional solutions often neglect spatiotemporal context analysis of historical normal claims events at the accident scene, lacking awareness of the clustering risks implied by population density and historical accident frequency in the affected area, resulting in insufficient ability to predict group fraud or insurance fraud.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a multimodal data-driven intelligent claims risk control system to solve the problems mentioned in the background. This invention calculates the correlation index R, which reflects the correlation of each modality of data, and the similarity index CSI between the current claims event and historical normal claims events at the corresponding accident scene. This ultimately achieves multi-dimensional evidence consistency verification and spatiotemporal context risk perception capabilities that are difficult for conventional systems to attain. Finally, through a dual-threshold joint decision-making mechanism, it accurately identifies high-risk cases and automatically triggers manual investigation while ensuring normal claims efficiency. This effectively solves the limitations of traditional risk control systems, such as reliance on a single data source, difficulty in identifying complex fraud patterns, and lack of dynamic adaptability.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A multimodal data-driven intelligent claims risk control system includes:

[0009] Multimodal data acquisition module: Collects multimodal data related to vehicle accident claims, including image data, text data, third-party data, and sensor data. The third-party data includes weather conditions, road congestion index, and historical vehicle accident statistics.

[0010] Multimodal data fusion and correlation analysis module: Constructs a multimodal fusion model, which includes multiple branches, each corresponding to data of one modality. It processes images and video frames through convolutional neural networks, text and third-party data through recurrent neural networks, and sensor values ​​through fully connected networks. The high-level semantic features extracted from each branch are then input into a cross-modal interaction module. This module employs a multi-head attention mechanism to guide and enhance features from different modalities, generating a multimodal data representation vector. This multimodal data representation vector includes: image data correlation scores, text data correlation scores, third-party data correlation scores, and sensor data correlation scores. The generated multimodal data representation vector is preprocessed, including denoising, normalization, and timestamp alignment, and a correlation index is calculated.

[0011] Historical claims analysis module: Collects metadata of claims events, including the historical claim frequency of the location of the claim event and the population density of the vehicle accident area, and calculates the claims similarity index based on the metadata;

[0012] Decision output module: Pre-set correlation threshold and claims similarity threshold, establish an audit mechanism, compare the obtained correlation index and claims similarity index with the correlation threshold and claims similarity threshold respectively, and output a decision plan based on the comparison results to determine whether to trigger the manual audit process.

[0013] Furthermore, in the multimodal data acquisition module, the image data includes panoramic and detailed photos of the vehicle accident claim scene, close-ups of the damaged parts of the vehicle, and video frames of the surrounding environment; the text data includes the complainant's oral record, the insurance company's preliminary investigation report, the repair work order description, and a summary of the liability determination letter; the third-party data includes the actual weather conditions and forecasts for the time of the vehicle accident claim, and the real-time traffic flow status and congestion index of the surrounding road sections.

[0014] The third-party data also includes geospatial information data, and the sensor data includes vehicle driving trajectory, speed changes, acceleration timing signals at the moment of collision recorded by the vehicle diagnostic system, and records of abnormal events such as sudden braking or severe vibration collected by the built-in sensors of the user's mobile phone.

[0015] Furthermore, the multimodal data fusion and correlation analysis module generates a data representation vector with quantitative evaluation effect for each modality of data: the image data correlation score extracts the consistency of physical evidence at the accident scene based on visual features; the text data correlation score evaluates the completeness and logical rationality of the report description through semantic analysis; the third-party data correlation score verifies the credibility of the event in combination with the environmental context; and the sensor data correlation score quantifies the abnormal detection results of vehicle dynamic parameters.

[0016] Furthermore, the multimodal data fusion and correlation analysis module enhances cross-modal features through a multi-head attention mechanism, and outputs a set of multimodal data representation vectors that fuse multi-dimensional evidence correlation strength;

[0017] The denoising preprocessing of multimodal data representation vectors is achieved through the following formula:

[0018]

[0019] in:

[0020] The output value after denoising the multimodal data representation vector;

[0021] For multimodal data representation vectors in time... The original input value;

[0022] k is the size of the sliding window;

[0023] i is the index within the window;

[0024] t is the index of the current time point;

[0025] By calculating the arithmetic mean of k consecutive data points within a sliding window, high-frequency random noise is suppressed, low-pass filtering is performed, low-frequency trend information is preserved, and short-term sudden interference is filtered out. The sliding window size k is adaptively set according to the data sampling frequency and noise intensity. When it is necessary to enhance the smoothing effect, the value of k is increased, and when it is necessary to preserve details, the value of k is decreased.

[0026] Furthermore, the formula for calculating the correlation index is as follows:

[0027]

[0028] in:

[0029] R is the correlation index;

[0030] λ is the cross-modal coordination coefficient, used to quantify the consistency among multimodal data;

[0031] Associating scores with image data;

[0032] Associative scores for text data;

[0033] The score is associated with third-party data;

[0034] The score is the correlation score between sensor data;

[0035] The weighting coefficients are respectively the image data association score, text data association score, third-party data association score, and sensor data association score, and the weighting coefficients satisfy the following conditions: ,and .

[0036] Furthermore, the cross-modal coordination coefficient λ is dynamically generated. The cross-modal coordination coefficient λ dynamically adjusts the overall reliability assessment by quantifying the internal consistency deviation between multimodal data, and the formula for generating the cross-modal coordination coefficient λ is:

[0037]

[0038] in:

[0039] Calibration factor;

[0040] ΔI, ΔT, ΔP, and ΔS represent the deviations of the image data association score, text data association score, third-party data association score, and sensor data association score from the mean, respectively; these are processed using the sum of squares: This amplifies the influence of high-deviation modes and prevents noise interference from being masked by averaging.

[0041] Furthermore, the formula for calculating the claims similarity index is as follows:

[0042]

[0043] in:

[0044] CSI is the Claims Similarity Index;

[0045] β is the exponential calibration coefficient;

[0046] This represents the real-time population density in the area where the vehicle was located at the time of the accident.

[0047] This represents the historical maximum population density in the area where the accident occurred. Used to Normalize to the [0,1] interval to eliminate the influence of dimensions;

[0048] This represents the cumulative number of incidents at the location of the incident within the past 12 months.

[0049] Furthermore, it also includes the process of determining the index calibration coefficient β: based on the distribution of historical claims data, an initial value of β is set to ensure that the CSI falls within the target range in typical scenarios, the mean CSI is calculated in the fraud case database, and if β leads to high-risk cases, β is adjusted down;

[0050] After adjustment, review the CSI of normal cases. If the proportion of samples with CSI > 90 for normal cases is less than 95%, then increase the β coefficient. After deployment, continuously monitor the audit accuracy. When the misjudgment rate increases by 5%, trigger the gradient descent iteration of β, and the gradient descent range is Δβ = ±10%.

[0051] Furthermore, in the decision output module, the system first loads the pre-trained correlation threshold. Similar thresholds to claims And the correlation threshold is Similar thresholds to claims The initial values ​​are determined by statistical quantiles based on the data distribution of historical fraud and non-fraud cases. A dual-threshold joint decision-making mechanism is established, which compares the real-time generated correlation index R and claims similarity index CSI with their corresponding thresholds in parallel to generate the final decision result.

[0052] Furthermore, the dual-threshold joint decision-making mechanism includes:

[0053] like and If the data is deemed highly valid and the risk of insurance fraud is low, the data will be directly processed into the claims process.

[0054] like or If so, a manual inspection process will be triggered;

[0055] The system automatically extracts new audit result data daily to update the threshold parameters. The audit result data includes manual review conclusions and fraud confirmation tags.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] This invention achieves collaborative verification and quantitative correlation analysis of physical evidence at the accident scene, reporting logic, environmental background, and vehicle dynamic parameters by deeply integrating multimodal data (images, text, third-party environmental data, and sensor time-series signals) and introducing a cross-modal multi-head attention mechanism. This significantly improves the accuracy and robustness of insurance fraud identification. Simultaneously, by calculating and combining the correlation index R, which reflects the correlation of each modality of data, and the similarity index CSI between the current claim event and historical normal claim events at the corresponding accident scene, this invention achieves multi-dimensional evidence consistency verification and spatiotemporal context risk perception capabilities that are difficult for conventional systems to reach. Finally, through a dual-threshold joint decision-making mechanism, high-risk cases are accurately identified and manual investigations are automatically triggered while ensuring normal claim efficiency. This effectively solves the limitations of traditional risk control systems that rely on a single data source, struggle to identify complex fraud patterns, and lack dynamic adaptability. Attached Figure Description

[0058] Fig. 1 This is a schematic diagram of the multimodal data-driven intelligent claims risk control system of the present invention;

[0059] Fig. 2 This is an execution flowchart of the multimodal data-driven intelligent claims risk control system of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0061] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0062] Example:

[0063] Please see Figs. 1-2 The present invention provides the following technical solutions:

[0064] A multimodal data-driven intelligent claims risk control system includes:

[0065] Multimodal data acquisition module: Collects multimodal data related to vehicle accident claims, including image data, text data, third-party data, and sensor data. The third-party data includes weather conditions, road congestion index, and historical vehicle accident statistics for the accident area.

[0066] The system can also be extended to financial business scenarios such as internet consumer finance and factoring services. Through multimodal data fusion technology, it can improve the ability to identify the authenticity of borrowers' behavior and the consistency of transaction background, and enhance the generalization ability and robustness of risk control models.

[0067] The image data includes not only panoramic and detailed photos of the accident scene and close-ups of the damaged parts of the vehicle, but also continuous video frames of the surrounding environment, as well as clear images of key time periods before and after the accident recorded by the dashcam and vehicle identification information.

[0068] The text data compiles detailed oral statements from the complainant, preliminary investigation reports from the insurance company, complete descriptions of repair work orders, and summaries of liability determination reports. It also supplements this with driver identification information, policy details, eyewitness testimonies, and the full text of the traffic police's accident report. When processing text data, the system not only analyzes unstructured text such as complainant statements and investigation reports, but also supports parsing smart contract terms related to claims. Combined with intelligent semantic matching technology, it identifies implicit contractual obligations and liability boundaries within the text. Simultaneously, the system can parse electronic contracts or automated claims agreements involving smart contract codes to ensure consistency between claims logic and business rules.

[0069] In addition to real-time weather conditions and forecasts at the time of the accident, real-time traffic flow and congestion index of surrounding road sections, third-party data also integrates geospatial information data such as precise GPS coordinates, map layers, information on surrounding buildings and roads, as well as historical accident statistics and special event information such as construction or traffic control for that road section.

[0070] Sensor data includes vehicle driving trajectory, speed changes, and acceleration timing signals at the moment of collision recorded by the on-board diagnostic system. It also includes abnormal event records such as sudden braking or severe vibration collected by the sensors built into the user's mobile phone. In addition, it also involves airbag deployment records, video clips from on-board cameras, and precise GPS trajectory timestamp data.

[0071] Multimodal Data Fusion and Correlation Analysis Module: The core task of this module is to construct a multimodal fusion model. This model aims to integrate data from different sources to comprehensively assess the credibility of accident claims. The model employs a multi-branch structure, with each branch specializing in processing data from one modality. For image and video frame data, convolutional neural networks are used to extract deep visual features, including the morphology, location, and severity of vehicle damage, as well as object recognition and spatial relationships in the scene, forming a quantitative description of the physical evidence at the accident scene.

[0072] For textual and third-party data, recurrent neural networks are used to capture semantic information and temporal dependencies. Textual data, such as police reports and investigation reports, are analyzed using natural language processing techniques to assess their completeness, logical consistency, and potential contradictions. Third-party data, such as weather and road conditions, are evaluated for consistency with the time and location of the accident through temporal modeling. For sensor data, fully connected networks are used to process numerical time-series signals, such as speed changes, acceleration peaks, and emergency braking events, extracting abnormal patterns and quantifying the degree of anomalies in vehicle dynamic parameters.

[0073] The high-level semantic features extracted from each branch are then fed into a cross-modal interaction module. This module, based on a multi-head attention mechanism, allows features from different modalities to guide and enhance each other. For example, the damage site in an image may be temporally aligned with the collision acceleration recorded by sensors, and the environmental description in text may corroborate third-party weather data. Through this interaction, the model generates multimodal data representation vectors, which specifically embody the association scores for each modality: image data association scores reflect the degree of consistency between visual evidence and accident description; text data association scores assess the logical rationality and semantic consistency of the reported information; third-party data association scores measure the degree to which external environmental data supports the accident scenario; and sensor data association scores quantify the anomaly detection results of vehicle dynamic parameters, serving as a basis for determining whether human manipulation occurred.

[0074] The generated multimodal data representation vectors require a series of preprocessing operations, including denoising to eliminate random interference during sensor acquisition, normalization to map the modality scores to a uniform scale for easy comparison, and timestamp alignment to ensure strict temporal matching of data from different sources. After preprocessing, the system calculates a comprehensive correlation index based on the correlation scores of each modality. This index measures the overall credibility and consistency of multimodal evidence in this claims event, providing a quantitative basis for subsequent decision-making.

[0075] The multimodal data fusion and correlation analysis module generates a data representation vector with quantitative evaluation effect for each modality of data: the image data correlation score extracts the consistency of physical evidence at the accident scene based on visual features; the text data correlation score evaluates the completeness and logical rationality of the report description through semantic analysis; the third-party data correlation score verifies the credibility of the event in combination with the environmental context; and the sensor data correlation score quantifies the abnormal detection results of vehicle dynamic parameters.

[0076] The multimodal data fusion and correlation analysis module further introduces a multi-head attention mechanism to enhance the high-level semantic features generated by each modality through cross-modal interaction. In this process, features from different modalities guide and calibrate each other: key regions in the image may receive higher attention weight due to text descriptions or sensor triggering moments; key information in the text is strengthened through alignment with image content; and third-party environmental data provides background verification for sensor signals. Through multiple rounds of attention interaction, the module ultimately outputs a set of multimodal data representation vectors that integrates the multi-dimensional evidence correlation strength. This set combines the independent evaluation results of each modality with their collaborative relationships.

[0077] The denoising preprocessing of multimodal data representation vectors is achieved through the following formula:

[0078]

[0079] in:

[0080] The output value after denoising the multimodal data representation vector, i.e., the current time step. The output value is the arithmetic mean of all the original input values ​​within the window, therefore its magnitude is related to the time intervals within the window. There is a positive correlation: when most of the original values ​​within the window increase, It increases accordingly;

[0081] For multimodal data representation vectors in time... The original input value is at time 1. The original input values, where From 0 to These values ​​together determine The value;

[0082] k is the size of the sliding window, which directly determines the smoothness and detail retention. and The smoothing effect is positively correlated: The larger the value, the more historical data points are involved in the averaging, the more thoroughly random noise is suppressed, and the smoother the output curve. However, it may also mask short-term real changes, leading to a lag in signal response; conversely, a smaller value indicates a lower value. The smaller the value, The closer it is to the original data, the faster it can reflect instantaneous fluctuations, but the noise reduction capability is weakened;

[0083] i is the index within the window, used to identify the offset of each historical data point relative to the current time.

[0084] t is the index of the current time point, which determines the end position of the sliding window;

[0085] In the data processing, a sliding window averaging method is used to denoise the multimodal data representation vector. This method suppresses high-frequency random noise introduced by sensor accuracy, environmental interference, or transmission errors by calculating the arithmetic mean of k consecutive data points within the sliding window, thus achieving smooth signal processing. This process allows low-frequency trend information to pass through smoothly while filtering out interference from short-term abrupt changes or abnormal peaks, thereby retaining the core components in the data that reflect the evolution of real events.

[0086] The sliding window size k is adaptively adjusted based on the actual sampling frequency of the data and the current noise level: when the data sampling frequency is high or the noise level is strong, increasing the value of k can achieve a stronger smoothing effect and make the trend clearer; while when it is necessary to retain local details of data changes or transient features of events, decreasing the value of k reduces the smoothing degree and avoids information loss due to over-filtering. This dynamic adjustment mechanism ensures that the denoising processing can adapt to the needs of different data qualities and application scenarios.

[0087] The formula for calculating the correlation index is as follows:

[0088]

[0089] in:

[0090] R is the correlation index;

[0091] λ is the cross-modal coherence coefficient, used to quantify the consistency among multimodal data. When the deviation of each modality score from the mean is small, it indicates that they are coherent and the chain of evidence is complete. Approaching 1, thus amplifying the weighted sum, making It is closer to the level of truth and credibility;

[0092] This is an image data association score, based on visual feature extraction, assessing the consistency of physical evidence at the accident scene, such as damage morphology and environmental traces, with the reported description. A higher score indicates more reliable image evidence.

[0093] The text data association score is used to assess the completeness and logical rationality of the reported information through semantic analysis. A higher score indicates more consistent and uncontradictory text information. It is related to... Positive correlation;

[0094] This score represents the correlation score between third-party data and the incident scenario; a high score indicates that the external data supports the incident scenario. It is related to... Positive correlation;

[0095] The sensor data correlation score is the highest score, indicating a better match between the sensor data and the accident description, and a lower degree of anomaly.

[0096] The weighting coefficients are for image data association score, text data association score, third-party data association score, and sensor data association score, respectively. At the same time, since text information, such as police reports or investigation reports, is often easily affected by subjective factors or is subject to human modification, its reliability is relatively low, so it has the lowest weight.

[0097] Weight of third-party data correlation score Slightly higher than text data, Objective environmental information, such as weather, road conditions, and geographic data, can provide relatively stable background verification, but may still have some lag or statistical bias. The weighting of image data association scores... High: On-site photos and videos can visually present accident traces and damage, possessing strong evidentiary value; however, they can sometimes be falsified or selectively filmed. Weighting of sensor data correlation scores. The highest weight, and no lower than, that of image data: Speed, acceleration, and trajectory recorded by sensors such as those in vehicle diagnostic systems and mobile phones are real-time physical signals, difficult to tamper with, and can most objectively reconstruct the dynamic process of an accident. Therefore, they are given the highest weight to reflect their decisive role. Thus, the weighting coefficient satisfies... ,and This normalizes the correlation index, giving the final calculation results a clear benchmark for comparison, making it easier to compare with preset thresholds in the future.

[0098] The cross-modal cooperability coefficient λ is dynamically generated. The overall reliability assessment is dynamically adjusted by quantifying the internal consistency deviation among multimodal data. The formula for generating the cross-modal cooperability coefficient λ is as follows:

[0099]

[0100] in:

[0101] This is a calibration factor used to control the deviation. The severity of the punishment. It is a positive number, and its value is set according to the actual application scenario and risk preference. and Negative correlation: The larger the deviation, the more likely it is to cause The decline is even greater;

[0102] ΔI, ΔT, ΔP, and ΔS represent the deviations of the image data association score, text data association score, third-party data association score, and sensor data association score from the mean, respectively. They are all related to... They are all negatively correlated, meaning that an increase in deviation from any mode will lead to The contribution decreases, but the degree of contribution is determined by the sum of squares treatment; through sum of squares treatment: This amplifies the influence of high-deviation modes and prevents noise interference from being masked by averaging.

[0103] Historical claims analysis module: Collects metadata of claims events, including the historical claim frequency of the location of the claim event and the population density of the vehicle accident area, and calculates the claims similarity index based on the metadata;

[0104] The formula for calculating the claims similarity index is:

[0105]

[0106] in:

[0107] CSI stands for Claims Similarity Index. The higher the CSI, the lower the probability that the claim for that vehicle accident was insurance fraud.

[0108] β is the exponential calibration coefficient, an adjustable global scaling factor used to map the original product value to a reasonable threshold comparison range;

[0109] This represents the real-time crowd density in the area where the vehicle was located at the time of the accident, expressed as the number of people per unit area. It is positively correlated with CSI.

[0110] This represents the historical maximum population density within the accident area, indicating the peak population density in the same area over a past statistical period. It is used to normalize real-time density to the [0, 1] interval, eliminating the impact of differences in population base across different areas. It is negatively correlated with CSI.

[0111] This represents the cumulative number of claims at the accident location over the past 12 months, indicating the total number of claims filed at that location during that period. This indicator directly reflects the frequency of accidents at that location and is strongly positively correlated with the CSI (Cost Per Indicator).

[0112] The process for determining the index calibration coefficient β is as follows: Based on the distribution of historical claims data, an initial value for β is set to ensure that the CSI falls within the target range in typical scenarios. The mean CSI is calculated in the fraud case database. If β leads to high-risk cases, β is adjusted downward.

[0113] After adjustment, review normal cases for CSI. If the percentage of samples with a value >90 is less than 95%, increase the β coefficient. After deployment, continuously monitor the audit accuracy. When the misjudgment rate increases by 5%, trigger the gradient descent iteration of β, and the gradient descent range is Δβ=±10%.

[0114] Decision output module: Pre-set correlation threshold and claims similarity threshold, establish an audit mechanism, compare the obtained correlation index and claims similarity index with the correlation threshold and claims similarity threshold respectively, and output a decision plan based on the comparison results to determine whether to trigger the manual audit process.

[0115] In the decision output module, upon system startup, two pre-trained key thresholds are first loaded: the correlation threshold θR and the claims similarity threshold θCSI. The initial values ​​of these two thresholds are based on a large amount of historical claims data. By analyzing the distribution characteristics of fraudulent and non-fraudulent cases, the thresholds are scientifically determined using statistical quantiles to ensure they effectively distinguish between normal claims and high-risk cases.

[0116] The system has a built-in configurable business rule engine that supports dynamic adjustment of risk control logic based on different financial business scenarios (such as auto insurance, consumer credit, and factoring services). During the claims decision-making process, the system constructs a reliable transaction evaluation model by verifying the consistency of multimodal data, ensuring that every claim or transaction is interpretable and traceable on the evidence chain.

[0117] This embodiment also constructs a dual-threshold joint decision-making mechanism. This mechanism compares the real-time calculated correlation index R and claims similarity index CSI with their corresponding thresholds θR and θCSI in parallel. Through this dual verification method, the system can comprehensively evaluate the degree of correlation of multimodal data of claims events and their similarity with historical cases, and finally generate a clear decision result, determining whether the claims case should directly enter the fast claims process or need to be transferred to the manual audit stage for in-depth review.

[0118] The dual-threshold joint decision-making mechanism includes:

[0119] like and If the data is deemed highly valid and the risk of insurance fraud is low, the data will be directly processed into the claims process.

[0120] If any index is below the threshold, that is or If so, a manual inspection process will be triggered;

[0121] After the system completes its daily operations, it automatically extracts newly added audit results data for the day. This data includes the final review conclusions given by human auditors and verified fraud case tags. The system uses this newly added labeled data as training samples, inputting it into a logistic regression model for incremental training. Through model iteration, the parameters of the thresholds θR and θCSI are dynamically updated. This closed-loop feedback mechanism allows the thresholds to continuously and adaptively adjust as fraud methods change and data accumulates, ensuring that the risk control strategy remains in an optimal state.

[0122] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0123] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0124] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0125] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A multimodal data-driven intelligent claims risk control system, characterized in that, include: Multimodal data acquisition module: Collects multimodal data related to vehicle accident claims, including image data, text data, third-party data, and sensor data. The third-party data includes weather conditions, road congestion index, and historical vehicle accident statistics. Multimodal data fusion and correlation analysis module: Constructs a multimodal fusion model, which includes multiple branches, each corresponding to data of one modality. It processes images and video frames through convolutional neural networks, text and third-party data through recurrent neural networks, and sensor values ​​through fully connected networks. The high-level semantic features extracted from each branch are then input into a cross-modal interaction module. This module employs a multi-head attention mechanism to guide and enhance features from different modalities, generating a multimodal data representation vector. This multimodal data representation vector includes: image data correlation scores, text data correlation scores, third-party data correlation scores, and sensor data correlation scores. The generated multimodal data representation vector is preprocessed, including denoising, normalization, and timestamp alignment, and a correlation index is calculated. Historical claims analysis module: Collects metadata of claims events, including the historical claim frequency of the location of the claim event and the population density of the vehicle accident area, and calculates the claims similarity index based on the metadata; Decision output module: Pre-set correlation threshold and claims similarity threshold, establish an audit mechanism, compare the obtained correlation index and claims similarity index with the correlation threshold and claims similarity threshold respectively, and output a decision plan based on the comparison results to determine whether to trigger the manual audit process.

2. The multimodal data-driven intelligent claims risk control system according to claim 1, characterized in that: In the multimodal data acquisition module, the image data includes panoramic and partial detail photos of the vehicle accident claim scene, close-ups of the damaged parts of the vehicle, and video frames of the surrounding environment; the text data includes the complainant's oral record, the insurance company's preliminary investigation report, the repair work order description, and the summary of the liability determination letter; the third-party data includes the actual weather conditions and forecasts for the time of the vehicle accident claim, and the real-time traffic flow status and congestion index of the surrounding road sections. The third-party data also includes geospatial information data, and the sensor data includes vehicle driving trajectory, speed changes, acceleration timing signals at the moment of collision recorded by the vehicle diagnostic system, and records of abnormal events such as sudden braking or severe vibration collected by the built-in sensors of the user's mobile phone.

3. The multimodal data-driven intelligent claims risk control system according to claim 1, characterized in that: The multimodal data fusion and correlation analysis module generates a data representation vector with quantitative evaluation effect for each modality of data: the image data correlation score extracts the consistency of physical evidence at the accident scene based on visual features; the text data correlation score evaluates the completeness and logical rationality of the report description through semantic analysis; the third-party data correlation score verifies the credibility of the event in combination with the environmental context; and the sensor data correlation score quantifies the abnormal detection results of vehicle dynamic parameters.

4. The multimodal data-driven intelligent claims risk control system according to claim 1, characterized in that: The multimodal data fusion and correlation analysis module enhances cross-modal features through a multi-head attention mechanism, and outputs a set of multimodal data representation vectors that fuse multi-dimensional evidence correlation strength. The denoising preprocessing of multimodal data representation vectors is achieved through the following formula: ； in: The output value after denoising the multimodal data representation vector; For multimodal data representation vectors in time... The original input value; k is the size of the sliding window; i is the index within the window; t is the index of the current time point; By calculating the arithmetic mean of k consecutive data points within a sliding window, high-frequency random noise is suppressed, low-pass filtering is performed, low-frequency trend information is preserved, and short-term sudden interference is filtered out. The sliding window size k is adaptively set according to the data sampling frequency and noise intensity. When it is necessary to enhance the smoothing effect, the value of k is increased, and when it is necessary to preserve details, the value of k is decreased.

5. The multimodal data-driven intelligent claims risk control system according to claim 4, characterized in that: The formula for calculating the correlation index is as follows: ； in: R is the correlation index; λ is the cross-modal coordination coefficient, used to quantify the consistency among multimodal data; Associating scores with image data; Associative scores for text data; The score is associated with third-party data; The score is the correlation score between sensor data; The weighting coefficients are respectively the image data association score, text data association score, third-party data association score, and sensor data association score, and the weighting coefficients satisfy the following conditions: ,and .

6. The multimodal data-driven intelligent claims risk control system according to claim 5, characterized in that: The cross-modal cooperability coefficient λ is dynamically generated. The overall reliability assessment is dynamically adjusted by quantifying the internal consistency deviation among multimodal data. The formula for generating the cross-modal cooperability coefficient λ is as follows: ； in: Calibration factor; ΔI, ΔT, ΔP, and ΔS represent the deviations of the image data association score, text data association score, third-party data association score, and sensor data association score from the mean, respectively; these are processed using the sum of squares: Amplify the effects of high-bias modes to prevent noise interference from being masked by averaging.

7. The multimodal data-driven intelligent claims risk control system according to claim 1, characterized in that: The formula for calculating the claims similarity index is as follows: ； in: CSI stands for Claims Similarity Index; β is the exponential calibration coefficient; This represents the real-time population density in the area where the vehicle was located at the time of the accident. This represents the historical maximum population density in the area where the accident occurred. Used to Normalize to the [0,1] interval to eliminate the influence of dimensions; This represents the cumulative number of incidents at the location of the incident within the past 12 months.

8. The multimodal data-driven intelligent claims risk control system according to claim 7, characterized in that, It also includes the process of determining the index calibration coefficient β: based on the distribution of historical claims data, set an initial value for β to ensure that the CSI falls within the target range in typical scenarios, calculate the average CSI in the fraud case database, and adjust β down if β leads to high-risk cases; After adjustment, review the CSI of normal cases. If the proportion of samples with CSI > 90 for normal cases is less than 95%, then increase the β coefficient. After deployment, continuously monitor the audit accuracy. When the misjudgment rate increases by 5%, trigger the gradient descent iteration of β, and the gradient descent range is Δβ = ±10%.

9. The multimodal data-driven intelligent claims risk control system according to claim 7, characterized in that: In the decision output module, the system first loads the pre-trained correlation threshold. Similar thresholds to claims And the correlation threshold is Similar thresholds to claims The initial values ​​are determined by statistical quantiles based on the data distribution of historical fraud and non-fraud cases. A dual-threshold joint decision-making mechanism is established, which compares the real-time generated correlation index R and claims similarity index CSI with their corresponding thresholds in parallel to generate the final decision result.

10. The multimodal data-driven intelligent claims risk control system according to claim 9, characterized in that: The dual-threshold joint decision-making mechanism includes: like and If the data is deemed highly valid and the risk of insurance fraud is low, the data will be directly processed into the claims process. like or If so, a manual inspection process will be triggered; The system automatically extracts new audit result data daily to update the threshold parameters. The audit result data includes manual review conclusions and fraud confirmation tags.

Images