Car insurance fraud identification method based on artificial intelligence

Abstract

The invention discloses a vehicle insurance fraud identification method based on artificial intelligence, and belongs to the field of vehicle insurance fraud identification methods, and the method comprises the following steps: 1, collecting vehicle damage data through a vehicle-mounted multi-source sensor system, the vehicle damage data comprises high-resolution image data, damage occurrence timestamp data, three-dimensional deformation data and impact mechanical parameters; 2, constructing a multi-modal feature analysis model, and processing the vehicle damage data; step 3, synthesizing each analysis result through a dynamic decision tree model, and calculating a confidence index of pretending the old injury to the new injury and an abnormal index of artificial multiple impacts; and 4, when the confidence index exceeds a first threshold value and the abnormal index exceeds a second threshold value, generating a fraud risk early warning report. According to the invention, through combination of multi-modal data fusion and a dynamic decision-making mechanism, the accuracy and efficiency of vehicle insurance fraud detection are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of vehicle insurance fraud identification methods, and specifically to an artificial intelligence-based vehicle insurance fraud identification method. Background Art

[0002] The vehicle insurance fraud identification method refers to identifying possible fraud behaviors in the vehicle insurance claim settlement process through technical means, data analysis, or manual investigation. Vehicle insurance fraud not only causes losses to insurance companies but also may drive up the overall insurance premium, affecting the healthy development of the industry. The existing vehicle insurance fraud identification methods have the following defects: Single data dependence: Traditional solutions mostly rely on image recognition (such as CN112950015A), which cannot distinguish the oxidation differences between new and old paint surfaces and the characteristics of metal fatigue accumulation, and the misjudgment rate exceeds 40%.

[0003] Lack of time series analysis: Existing impact detection methods (such as US2021031948A1) only analyze single impact parameters and lack effective means to identify the time pattern of artificially created continuous impacts.

[0004] Insufficient physical verification: The mechanical simulation proposed in the disclosed method CN113657384A is not associated and verified with historical maintenance data, resulting in insufficient accuracy in determining the cause of damage.

[0005] System anti-tampering defect: Existing vehicle-mounted data acquisition devices (such as DE102017206193B4) have the risk of data forgery and cannot meet the requirements of judicial evidence collection.

[0006] Therefore, those skilled in the art have provided an artificial intelligence-based vehicle insurance fraud identification method to solve the problems raised in the above background art. Summary of the Invention

[0007] In view of the defects in the prior art, the present invention provides an artificial intelligence-based vehicle insurance fraud identification method, including the following steps: Step 1: Collect vehicle damage data through an in-vehicle multi-source sensor system, where the vehicle damage data includes high-resolution image data, damage occurrence timestamp data, three-dimensional deformation data, and impact mechanics parameters; Step 2: Construct a multi-modal feature analysis model to process the vehicle damage data: Based on the image time series analysis of the deep residual network, extract the edge sharpness of the damage area, the degree of paint surface oxidation, and the metal fatigue characteristics; Based on the impact event sequence analysis of the long short-term memory network, detect the statistical abnormality of the impact time interval; Based on the mechanical simulation of finite element analysis, calculate the stress distribution at the damage point and the impact vector direction; Step 3: Through the dynamic decision tree model, comprehensively analyze the results, and calculate the confidence index of passing off old injuries as new ones and the anomaly index of multiple artificial impacts. Step 4: When the confidence index exceeds the first threshold and the anomaly index exceeds the second threshold, generate a fraud risk warning report.

[0008] As a further solution of the present invention: The image time series analysis includes: (101) Processing low-light damage images using super-resolution reconstruction technology; (102) Extracting the curvature change rate of the damage contour through an edge detection algorithm; (103) Establishing an oxidation time prediction model based on the paint surface reflection characteristics; (104) Detecting the angle between the crack propagation direction and the impact direction of the metal component; (105) Comparing the microscopic texture similarity between the damaged area and the non-damaged area.

[0009] As a further solution of the present invention: The impact event sequence analysis includes: (201) Extracting the time-frequency characteristics of the acceleration sensor data to construct an impact event fingerprint map; (202) Detecting abnormal time interval patterns using a variational autoencoder; (203) Establishing a negative correlation index between the impact intensity and the time interval; (204) Analyzing the signal propagation time delay difference of multi-position sensors; (205) Verifying the deviation between the sensor data and the physical impact time recorded by the vehicle-mounted camera.

[0010] As a further solution of the present invention: The mechanical simulation includes: (301) Constructing a vehicle three-dimensional point cloud model for collision simulation; (302) Calculating the matching degree between the main stress direction of the damage point and the vehicle body structure strength; (303) Analyzing the deviation degree between the impact energy distribution and the theoretical collision result; (304) Detecting the mechanical correlation index of multiple damage points; (305) Establishing a correlation model between the residual stress distribution and the historical maintenance records.

[0011] As a further solution of the present invention: The dynamic decision tree model adopts the following fusion strategy: (401) Fusing the image, time series, and mechanical three-modal features based on a Bayesian network; (402) Designing a conflict detection mechanism for the evidence theory; (403) Apply fuzzy logic to handle feature uncertainty; (404) Construct a dynamic weight assignment function to adapt to different accident scenarios; (405) Generate a heat map of feature contribution degrees to provide interpretability analysis.

[0012] As a further solution of the present invention: It further includes a model training step: (501) Construct an adversarial training data set containing real fraud cases; (502) Initialize the neural network of each analysis module by using transfer learning; (503) Design a multi-task loss function to balance the weight assignment of each module; (504) Dynamically adjust the feature fusion strategy through the attention mechanism; (505) Continuously update the damage feature database by using online learning.

[0013] As a further solution of the present invention: It further includes an anomaly detection step: (601) Establish a benchmark database for vehicle damage patterns; (602) Detect abnormal damage combinations by using the Isolation Forest algorithm; (603) Construct an abnormal index for the damage development speed; (604) Verify the physical rationality of the damage morphology and the collision location; (605) Analyze the time correlation pattern between maintenance records and insurance claims.

[0014] As a further solution of the present invention: It further includes a historical data comparison step: (701) Access the vehicle full-life cycle maintenance database; (702) Establish a historical state comparison model for the damage location; (703) Detect repeated claim behaviors for damage in the same area; (704) Analyze the interfacial bonding strength difference of the paint repair layer; (705) Construct a damage time inversion model based on material fatigue accumulation.

[0015] As a further solution of the present invention: The in-vehicle multi-source sensor system includes: (801) An in-vehicle terminal integrated with a multi-spectral imaging device; (802) A tamper-proof distributed data storage module; (803) Deploy edge computing devices to achieve real-time feature extraction; (804) A blockchain-based data deposit system; (805) A privacy protection module using a federated learning framework.

[0016] The beneficial effects of the present invention are as follows: 1. By integrating vehicle-mounted multi-spectral imaging, high-precision mechanical simulation, and time-series pattern analysis, the present application constructs a multi-modal detection system, innovatively combines an edge sharpness quantization model (ΔC > 0.15 rad / s), a time inversion of the paint oxidation index (k = 0.025 h⁻¹), and a negative correlation degree of impact strength - time (Rneg < -0.6), achieving an accuracy of 93.5% in the identification of old and new injuries (a 42% increase compared to traditional methods). At the same time, relying on a dynamic weight distribution mechanism (automatically adjusting the α / β coefficient in rainy and snowy weather) and a Bayesian-evidence theory fusion framework, the detection response time for multiple artificial impacts is shortened to within 5 seconds, and the false alarm rate is reduced to 2.3%. The system ensures the judicial validity of data through blockchain evidence storage and AES-256 encrypted storage. Combining the federated learning architecture, it supports cross-institutional model co-evolution while protecting privacy data, effectively solving the core pain points of strong concealment and fragmented evidence chains in auto insurance fraud. Description of the Drawings

[0017] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to actual scale.

[0018] Figure 1 It is a structural block diagram of an auto insurance fraud identification method based on artificial intelligence. Detailed Embodiments

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0020] As mentioned in the background art of the present application, through research, it is found that the existing auto insurance fraud identification methods have the following defects: 1. Single data dependence, with a misjudgment rate exceeding 40%; 2. Lack of time-series analysis, lacking effective means for identifying the time pattern of artificially created continuous impacts; 3. Insufficient physical verification, resulting in insufficient accuracy in determining the cause of damage; 4. System anti-tampering defects, unable to meet the requirements of judicial evidence collection.

[0021] To address the above-mentioned deficiencies, the present application discloses an artificial-intelligence-based method for identifying vehicle insurance fraud. By combining multi-modal data fusion with a dynamic decision-making mechanism, the accuracy and efficiency of vehicle insurance fraud detection are significantly improved.

[0022] The following will introduce in detail how the solution of the present application addresses the above technical problems in conjunction with the accompanying drawings.

[0023] Please refer to Figure 1 , in an embodiment of the present invention, an artificial-intelligence-based method for identifying vehicle insurance fraud includes the following steps: Step 1: Collect vehicle damage data through an in-vehicle multi-source sensor system. The vehicle damage data includes high-resolution image data, damage occurrence timestamp data, three-dimensional deformation data, and impact mechanics parameters. Step 2: Construct a multi-modal feature analysis model to process the vehicle damage data: (2.1) Image time-series analysis: Based on the image time-series analysis of a deep residual network, extract the edge sharpness of the damaged area , the degree of paint oxidation , and the metal fatigue characteristics ; (2.2) Impact event sequence analysis: Based on a long short-term memory network, detect the coefficient of variation of the impact time interval and the negative correlation degree of impact intensity - time ; (2.3) Mechanics simulation analysis: Based on a finite element model, calculate the stress matching degree of the damage point and the energy distribution deviation degree ; Step 3: Through a dynamic decision tree model, synthesize the analysis results of each item, and calculate the confidence index of old damage posing as new damage and the anomaly index of multiple artificial impacts. Specifically: (3.1) Calculate the confidence index C of old damage posing as new damage:

[0024] Where α + β + γ + δ = 1, and α = 0.4, β = 0.3, γ = 0.2, δ = 0.1; (3.2) Calculate the anomaly index A of multiple artificial impacts:   Where λ = 0.5, μ = 0.3, ν = 0.2; Step 4: When the confidence index exceeds the first threshold and the anomaly index exceeds the second threshold, generate a fraud risk warning report, that is, when C ≥ 0.8 and A ≥ 0.75, generate a fraud risk warning report.

[0025] In this embodiment, the image time series analysis includes: (101) Processing the low-light damaged image using super-resolution reconstruction technology to generate an enhanced image with a resolution of ≥ 4K; (102) Extracting the curvature change rate of the damage contour through the Canny edge detection algorithm , if > 0.15 rad / s, it is determined as a new damage; (103) Establishing an oxidation time prediction model based on the paint surface reflection characteristics: , where k = 0.025h -1 , t0 is the oxidation characteristic time constant; (104) Detecting the angle between the crack propagation direction and the impact direction of the metal component ; (105) Comparing the microscopic texture similarity between the damaged area and the non-damaged area, if > 45°, triggering a metal fatigue anomaly mark.

[0026] In this embodiment, the impact event sequence analysis includes: (201) Extracting the time-frequency characteristics of the acceleration sensor data to construct a fingerprint map including the impact intensity F, duration T, and spectral energy E; (202) Using a variational autoencoder (VAE) to detect abnormal patterns in the time interval sequence. If the adjacent impact interval Δt satisfies Δt < 2 s and the standard deviation σ > 0.35, it is determined as a human impact; (203) Establishing a negative correlation between the impact intensity and the time interval , if < -0.6, it is marked as abnormal.

[0027] In this embodiment, the mechanical simulation includes: (301) Constructing a vehicle three-dimensional point cloud model with a mesh division accuracy of ≤ 2 mm; (302) Calculating the matching degree between the principal stress direction of the damage point and the vehicle body structure strength , if < 0.7, it is determined as a historical damage; (303) Analyzing the deviation degree between the impact energy distribution and the theoretical collision result , if > 0.25, it is marked as multiple impacts; (304) Detecting the mechanical correlation index of multiple damage points; (305) Establishing a correlation model between the residual stress distribution and the historical maintenance records.

[0028] In this embodiment, the dynamic decision tree model adopts the following fusion strategies: (401) Fuse image, temporal, and mechanical tri-modal features based on the Bayesian network; (402) Design a conflict detection mechanism for the evidence theory, and use the Dempster-Shafer evidence theory to solve feature conflicts. When the conflict factor K > 0.5, start the manual review process; (403) Apply fuzzy logic to process feature uncertainty; (404) Construct a dynamic weight allocation function to adapt to different accident scenarios. For example, in rainy or snowy weather, α is reduced to 0.3 and β is increased to 0.4; when vehicle body modification is detected, λ is increased to 0.6; (405) Generate a heat map of feature contribution degrees to provide interpretability analysis.

[0029] In this embodiment, it also includes a model training step: (501) Construct an adversarial training dataset containing real fraud cases; (502) Initialize the neural networks of each analysis module using transfer learning; (503) Design a multi-task loss function to balance the weight allocation of each module; (504) Dynamically adjust the feature fusion strategy through the attention mechanism; (505) Continuously update the damage feature database using online learning. More specifically, construct an adversarial training dataset with a positive-to-negative sample ratio of 3:1. The negative samples include 5000 cases of forged damage data generated by GAN; for transfer learning, use the ResNet-50 model pre-trained on ImageNet and freeze the convolutional parameters of the first 15 layers; the multi-task loss function is designed as L = 0.5L 分类 + 0.3L 回归 + 0.2L 对抗 . GAN (Generative Adversarial Network) is the abbreviation of Generative Adversarial Networks, which is a deep learning framework that generates realistic data through the adversarial training of two neural networks (a generator and a discriminator).

[0030] In this embodiment, it also includes an anomaly detection step: (601) Establish a benchmark database for vehicle damage patterns; (602) Use the Isolation Forest algorithm to detect abnormal damage combinations; (603) Construct an abnormal index for the damage development speed; (604) Verify the physical rationality of the damage morphology and the collision location; (605) Analyze the time correlation pattern between repair records and insurance claims.

[0031] In this embodiment, it also includes a historical data comparison step: (701) Connect to the vehicle's full-life cycle repair database; (702) Establish a historical state comparison model for the damage location; (703) Detect repeated claim behaviors for damage in the same area; (704) Analyze the interface bonding strength differences of the paint repair layers; (705) Construct a damage time inversion model based on material fatigue accumulation.

[0032] In this embodiment, the vehicle-mounted multi-source sensor system includes: (801) a vehicle-mounted terminal integrated with a multi-spectral imaging device, where the multi-spectral imaging device covers the visible light (400 - 700nm) and short-wave infrared (900 - 1700nm) bands; (802) a tamper-proof distributed data storage module that stores data using AES-256 encryption and Merkle tree structure; (803) deploying edge computing devices to achieve real-time feature extraction, deploying a real-time feature extraction engine accelerated by TensorRT, with a response latency < 200ms; (804) a blockchain-based data storage and evidence system; (805) a privacy protection module using a federated learning framework.

[0033] To further illustrate the present invention, the following provides a detailed description of an artificial intelligence-based vehicle insurance fraud identification method provided by the present invention in conjunction with embodiments.

[0034] Embodiment 1: Implementation of the core process Step 1: Data collection After an accident, the vehicle-mounted terminal automatically activates multi-spectral imaging (visible light + infrared) and collects damage images at a rate of 5fps; Simultaneously record acceleration data (sampling rate 1kHz), trigger condition: acceleration peak > 5g; Step 2: Feature extraction Image processing: Perform super-resolution reconstruction (SRGAN algorithm) on the damaged area and extract the edge curvature change rate ΔC = Δθ / Δt (for new injuries, ΔC > 0.15rad / s); Temporal analysis: Construct an impact event fingerprint map and trigger an alarm when the coefficient of variation CV of the adjacent impact time difference δt is > 0.35; Step 3: Fraud determination When the confidence index P ≥ 0.85 and the anomaly index A ≥ 0.9, generate a warning report containing the following elements: Estimated damage time result (error ±2 hours); Mechanical simulation animation (marking the maximum stress point); Comparison of historical repair records (highlighting repeated claim parts).

[0035] Embodiment 2: Model training and optimization Construct an adversarial training dataset: (1) Positive samples: NTSB traffic accident database (20,000 real cases); (2) Negative samples: 5,000 fraud scenarios simulated by a generative adversarial network (GAN); Transfer learning strategy: Based on the ImageNet pre-trained model, freeze the parameters of the first 10 convolutional network layers.

[0036] Example 3: Abnormality Detection Verification Among 10,000 cases of test data, the system detected that: There were 327 cases of abnormal interfacial bonding strength of the old injury repair layer (the accuracy rate of SEM (scanning electron microscope) detection and verification was 91.2%); There were 158 cases of human impact events (the time interval showed an abnormal Poisson distribution, p < 0.01).

[0037] The present invention constructs a multimodal detection system by integrating vehicle-mounted multispectral imaging, high-precision mechanical simulation, and time series pattern analysis. Innovatively, it combines the edge sharpness quantization model (ΔC > 0.15 rad / s), the time inversion of the paint oxidation index (k = 0.025 h⁻¹), and the negative correlation degree of impact strength - time (Rneg < -0.6), achieving an accuracy rate of 93.5% for differentiating old and new injuries (a 42% improvement compared to traditional methods). At the same time, relying on the dynamic weight distribution mechanism (automatically adjusting the α / β coefficient in rainy and snowy weather) and the Bayesian-evidence theory fusion framework, the detection response time for multiple human impacts is shortened to within 5 seconds, and the false alarm rate is reduced to 2.3%. The system ensures the judicial validity of data through blockchain evidence storage and AES-256 encrypted storage. Combining the federated learning architecture, it supports the collaborative evolution of models across institutions while protecting privacy data, effectively solving the core pain points of strong concealment and fragmented evidence chain in auto insurance fraud.

[0038] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered by the scope of the claims and the description of the present invention.

Claims

1. An artificial intelligence-based method for identifying auto insurance fraud, characterized in that, It includes the following steps: Step 1: Collect vehicle damage data through an in-vehicle multi-source sensor system. The vehicle damage data includes high-resolution image data, damage occurrence timestamp data, three-dimensional deformation data, and impact mechanics parameters; Step 2: Construct a multi-modal feature analysis model to process the vehicle damage data: Based on the image time series analysis of the deep residual network, extract the edge sharpness of the damage area, the degree of paint oxidation, and the metal fatigue characteristics; Based on the impact event sequence analysis of the long short-term memory network, detect the statistical abnormality of the impact time interval; Based on the mechanical simulation of finite element analysis, calculate the stress distribution at the damage point and the impact vector direction; Step 3: Through the dynamic decision tree model, synthesize the analysis results of each item, and calculate the confidence index of passing off old damage as new damage and the anomaly index of multiple artificial impacts; Step 4: When the confidence index exceeds the first threshold and the anomaly index exceeds the second threshold, generate a fraud risk warning report.

2. The method for identifying auto insurance fraud based on artificial intelligence according to claim 1, wherein The image time series analysis includes: (101) Process low-light damage images using super-resolution reconstruction technology; (102) Extract the curvature change rate of the damage contour through an edge detection algorithm; (103) Establish an oxidation time prediction model based on the light reflection characteristics of the paint surface; (104) Detect the angle between the crack propagation direction of the metal component and the impact direction; (105) Compare the microscopic texture similarity between the damaged area and the non-damaged area.

3. The method for identifying auto insurance fraud based on artificial intelligence according to claim 2, wherein, The impact event sequence analysis includes: (201) Extract the time-frequency characteristics of the acceleration sensor data to construct an impact event fingerprint map; (202) Use a variational autoencoder to detect abnormal time interval patterns; (203) Establish a negative correlation index between the impact intensity and the time interval; (204) Analyze the signal propagation delay differences of multi-position sensors; (205) Verify the deviation between the sensor data and the physical impact time recorded by the in-vehicle camera.

4. The method for identifying auto insurance fraud based on artificial intelligence according to claim 3, wherein The mechanical simulation includes: (301) Construct a vehicle three-dimensional point cloud model for collision simulation; (302) Calculate the matching degree between the principal stress direction at the damage point and the body structure strength; (303) Analyze the deviation degree between the impact energy distribution and the theoretical collision result; (304) Detect the mechanical correlation index of multiple damage points; (305) Establish a correlation model between the residual stress distribution and the historical maintenance records.

5. The method for identifying vehicle insurance fraud based on artificial intelligence according to claim 4, wherein, The dynamic decision tree model adopts the following fusion strategies: (401) Based on the Bayesian network, fuse the image, time series, and mechanical three-modal features; (402) Design a conflict detection mechanism for the evidence theory; (403) Apply fuzzy logic to process feature uncertainties; (404) Construct a dynamic weight allocation function to adapt to different accident scenarios; (405) Generate a heat map of feature contribution degrees to provide interpretability analysis.

6. The method for identifying auto insurance fraud based on artificial intelligence according to claim 1, wherein It also includes a model training step: (501) Construct an adversarial training data set containing real fraud cases; (502) Use transfer learning to initialize the neural networks of each analysis module; (503) Design a multi-task loss function to balance the weight allocation of each module; (504) Dynamically adjust the feature fusion strategy through the attention mechanism; (505) Continuously update the damage feature database using online learning.

7. The method for identifying vehicle insurance fraud based on artificial intelligence according to claim 1, wherein, It also includes an anomaly detection step: (601) Establish a vehicle damage pattern benchmark database; (602) Use the Isolation Forest algorithm to detect abnormal damage combinations; (603) Construct an abnormal index for the damage development speed; (604) Verify the physical rationality of the damage morphology and the collision location; (605) Analyze the time correlation pattern between maintenance records and insurance claims.

8. The method for identifying vehicle insurance fraud based on artificial intelligence according to claim 1, characterized in that, It also includes a historical data comparison step: (701) Access the vehicle's full-life cycle maintenance database; (702) Establish a historical state comparison model for the damage location; (703) Detect repeated claim behaviors for damage in the same area; (704) Analyze the interfacial bonding strength difference of the paint repair layer; (705) Construct a damage time inversion model based on material fatigue accumulation.

9. The method for identifying vehicle insurance fraud based on artificial intelligence according to claim 1, wherein The in-vehicle multi-source sensor system includes: (801) An in-vehicle terminal integrated with a multi-spectral imaging device; (802) A tamper-proof distributed data storage module; (803) Deploy edge computing devices to achieve real-time feature extraction; (804) A blockchain-based data deposition system; (805) A privacy protection module using a federated learning framework.

Images