Large language model-based vehicle owner behavior data-driven vehicle insurance risk assessment method

Abstract

The invention discloses a vehicle owner behavior data driving type vehicle insurance risk assessment method based on a large language model, and relates to the technical field of risk assessment. Comprehensive vehicle owner and vehicle behavior data are collected from multiple channels, and deep analysis is performed on the data through the large language model; driving behavior features, historical accident record features and vehicle use condition features highly related to the vehicle insurance risk are extracted, and the influence of various factors on the risk is accurately reflected in combination with a nonlinear conversion function; key features are converted into numerical vectors and stored in the form of a multi-dimensional feature matrix, and high-quality input data are provided for a deep learning model; a risk assessment model constructed by adopting a PyTorch framework is combined with loss function optimization and a gradient descent algorithm, so that efficient model training and prediction are realized; and the risk assessment result is adjusted in real time through dynamic data updating, so that the accuracy of vehicle insurance risk assessment and the response speed to the behavior change of the vehicle owner are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of risk assessment, and in particular to a vehicle insurance risk assessment method driven by vehicle owner behavior data based on a large language model. Background Art

[0002] Auto insurance refers to an insurance product that provides risk protection for motor vehicles. It aims to provide financial compensation to the insured (car owner) according to the contract when the vehicle is involved in an accident, stolen, damaged or causes loss to a third party. Auto insurance is an indispensable part of the modern transportation system. It not only protects the property safety of car owners, but also maintains the safety and order of road traffic. With the rapid development of big data and artificial intelligence technology, the auto insurance industry has gradually introduced intelligent tools to improve the accuracy and efficiency of risk assessment. The behavioral data of car owners and the historical data of vehicles are important bases for auto insurance companies to assess the risks of car owners and formulate auto insurance cost strategies.

[0003] In the existing technology, most auto insurance companies mainly rely on traditional risk assessment methods. Traditional risk assessment methods are based on limited structured data such as the owner's historical violation records, vehicle usage and maintenance records. Although they can reflect some of the owner's behavioral characteristics, due to their single data dimension, they cannot cover the owner's overall driving habits and potential risk factors. In addition, traditional risk assessment methods usually use empirical rules or simple statistical models for analysis, and lack the ability to process unstructured data, making it difficult for existing risk assessment systems to conduct a comprehensive and detailed analysis of the owner's potential accident risks, thereby affecting the rationality of auto insurance costs.

[0004] To solve this problem, we propose a car owner behavior data-driven auto insurance risk assessment method based on a large language model. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model, which can effectively solve the problems in the above-mentioned background technology.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] The vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model includes the following steps:

[0008] S1. Obtaining the behavior data of multiple users and constructing multiple owner and vehicle behavior data sets for standardization, wherein the owner and vehicle behavior data sets include insurance company historical assessment report data, driving behavior data, owner violation record data, accident record data, motor vehicle usage status data, and motor vehicle maintenance record data;

[0009] S2. Based on the large language model, the standardized car owner and vehicle behavior data set of each user is deeply analyzed to extract key features related to auto insurance risk, including driving behavior features, historical accident record features and vehicle usage features. The extracted key features of each user are converted into numerical vectors through the large language model and integrated and stored in the form of a multi-dimensional feature matrix;

[0010] S3. Input the multi-dimensional feature matrix into the PyTorch deep learning framework to train the risk assessment model, use the insurance company's historical assessment report data as labels to build a supervised learning model, and use the risk assessment model after multiple rounds of iterations and model optimization to predict the risk level of the car owner and the vehicle;

[0011] S4. Input the car owner and vehicle behavior data set of the user to be evaluated into the trained risk assessment model to generate a risk value, and generate a risk assessment result based on the risk level corresponding to the risk value to support the insurance company to dynamically adjust the auto insurance premium;

[0012] S5. Generate a natural language report based on the large language model, and explain the impact of key factors in the owner and vehicle behavior data set on the risk value calculation based on the risk assessment results;

[0013] S6. Monitor the behavior data of the owner and the vehicle in real time through the vehicle-mounted equipment, dynamically adjust the risk value, the generated risk level and the risk assessment report according to the latest data, and update them according to the real-time behavior data to reflect the changes in the owner's driving habits;

[0014] S7. Monitor the performance of the risk assessment model in real time and update the model parameters regularly.

[0015] Preferably, the S1 specifically includes:

[0016] Collect the behavior data of multiple users from multiple sources, and build multiple owner and vehicle behavior data sets D by integrating multi-source heterogeneous data. total ;

[0017] The vehicle owner and vehicle behavior data set includes insurance company historical assessment report data, driving behavior data, vehicle owner violation record data, accident record data, vehicle usage status data, and vehicle maintenance record data;

[0018] Driving behavior data is used to record the driver's behavior data during historical driving, including speeding, sudden acceleration, sudden deceleration, turning, night driving, long-distance driving, fatigue driving and seat belt use;

[0019] The owner's violation record data is used to record the owner's violation behaviors during the historical driving process, and the owner's violation record data includes the violation type, violation number, violation time, violation location and violation severity;

[0020] The accident record data is used to record the owner's historical accident records and insurance claim records. The historical accident records include the accident time, accident location, accident type, accident frequency, liability determination and loss extent. The insurance claim records include compensation items, compensation amounts and compensation items.

[0021] The motor vehicle usage status data is used to record usage status information and vehicle status information, wherein the motor vehicle usage status information includes the age and mileage of the vehicle;

[0022] The motor vehicle maintenance record data is used to record the vehicle maintenance record and maintenance record, and the maintenance record includes maintenance time, maintenance items, maintenance frequency, maintenance time, maintenance items and maintenance frequency;

[0023] The insurance company's historical assessment report data is used to record the insurance company's past risk assessment data on the vehicle owner and vehicle;

[0024] For the collected car owner and vehicle behavior dataset D total Perform data cleaning, which includes removing duplicate data, erroneous data, and correcting missing values, to generate a cleaned car owner and vehicle behavior dataset D clean ;

[0025] For the cleaned car owner and vehicle behavior dataset D clean Formatting is performed to convert multi-source heterogeneous data into a unified standard format to meet the requirements of model processing, and the formatted owner and vehicle behavior dataset D is obtained. formatted ;

[0026] For the formatted owner and vehicle behavior dataset D formatted Perform deduplication processing to remove redundant records, ensure data uniqueness, and finally generate a standardized car owner and vehicle behavior dataset D standard ;

[0027] The vehicle owner and vehicle behavior dataset D total :

[0028] D total =D c ∪D v ∪D a ∪D u ∪D m ∪D h ;

[0029] Among them, D crepresents driving behavior data, D v Indicates violation record data, D a Indicates accident record data, D u Indicates the motor vehicle usage status data, D m Represents motor vehicle maintenance record data, including vehicle maintenance records and maintenance records, D h Represents the insurance company's historical assessment report data, including the insurance company's past risk assessment data on the vehicle owner and the vehicle.

[0030] Preferably, the step S2 extracts key features related to the auto insurance risk, specifically comprising the following steps:

[0031] S21. Use the owner’s driving behavior data as the driving behavior variable x d ;

[0032] Based on the owner’s violation record data, construct the driving behavior variable x d The nonlinear transformation function f d (x d ), f d (x d ) is used to reflect the driving behavior variable x d The degree of impact on driving risk;

[0033] Based on the violation record data, the driving behavior data of the car owner in the period T before the violation occurred is obtained. The time period T is defined as a 60-second window before the violation event occurs. Within the time period T, each second is regarded as a discrete time point t. The driving behavior variables at each time point t in the time period T are extracted, and the function φ(v,t) is constructed to represent the driving behavior performance of the car owner at a specific time t.

[0034] Based on the driving behavior variable x d The nonlinear transformation function f d (x d ), time period T and function φ(v,t), to obtain the driving behavior characteristics F of the owner d ;

[0035] S22. Based on the accident record data, obtain historical accident records;

[0036] Based on historical accident records, obtain the accident factor x of each accident a , the accident factor x a Including accident time, accident location, accident type, accident frequency, responsibility determination and loss extent;

[0037] Construct accident factor x a The nonlinear transformation function f a (xa ), to reflect the accident factor x a The degree of impact on driving risk;

[0038] Set the time decay factor β based on historical accident records a , to reflect the changes in the impact of accidents over time and to more accurately reflect the impact of recent accidents;

[0039] Construct the time function λ of the accident frequency a (t), time function λ a (t) reflect the changing trend of accident frequency over time by accumulating the number of accidents and applying moving average technology to reduce the impact of short-term fluctuations;

[0040] Based on the time function λ a (t), time decay factor β a and accident factor x a The nonlinear transformation function f a (x a ) to obtain the owner's historical accident record feature F a ;

[0041] S23. Obtaining vehicle parameters x based on vehicle status information in the motor vehicle usage status data v , the vehicle parameters include brake system status, tire status, steering system status, engine status, gearbox status and exhaust system status;

[0042] Based on the insurance claim records of motor vehicle maintenance record data and accident record data, the support, confidence and lift between the same maintenance items and compensation items are calculated using the April algorithm;

[0043] The support degree indicates the frequency of a maintenance item and a corresponding compensation item appearing in all records, the confidence degree indicates the probability that the corresponding compensation item will also appear when a maintenance item appears, and the lift degree indicates the strength of association between the maintenance item and the corresponding compensation item;

[0044] Based on the improvement, adjust the corresponding vehicle parameter x v The adjustment factor w v , adjustment factor w v It is used to reflect the potential relationship and impact of maintenance items and maintenance frequency based on the failure of the item type;

[0045] Construct vehicle parameters x v The nonlinear transformation function f v (x v ) to assess the safety performance of the vehicle in its current condition;

[0046] Set the vehicle loss coefficient γ v The vehicle loss coefficient is used to quantify the degree of loss of the vehicle based on its service life, and obtain the vehicle service life c, the vehicle loss coefficient γ v The product of the vehicle's service life c is taken as the vehicle's wear and tear degree;

[0047] Get the vehicle's wear function δ v (t), the expression of this function is: v (t) = 1-e -λ·mileage , where λ represents the wear rate parameter of the vehicle per kilometer traveled, and mileage represents the vehicle mileage;

[0048] Based on the vehicle parameters x v The nonlinear transformation function f v (x v ), adjustment coefficient w of vehicle parameters v , vehicle loss coefficient γ v , vehicle age c and vehicle wear function δ v (t), to obtain the vehicle owner’s vehicle usage characteristics F v .

[0049] Preferably, in S21-S23, based on the large language model, the standardized car owner and vehicle behavior data set of each user is deeply analyzed to extract the driving behavior features F related to the car insurance risk. d 、Historical accident record characteristics F a and vehicle usage characteristics F v , the calculation formula is as follows:

[0050]

[0051] Among them, F d represents the driving behavior characteristics, n d represents the total number of driving behavior variables, represents the weight of the i-th driving behavior variable, represents the i-th driving behavior variable, Represents driving behavior variables The nonlinear transfer function, α d is the long-term driving behavior preference adjustment coefficient, the time dimension T represents the duration of the behavior, the function φ(v,t) represents the driving behavior performance of the owner at a specific time t, and F a Represents the characteristics of historical accident records, is the weight of the jth accident factor, Indicates the factors of the accident The nonlinear transfer function, β ais the time decay factor of the accident, λ a (t) represents the time function of the accident frequency, F v Indicates the vehicle usage characteristics, represents the weight of the kth vehicle parameter, Indicates vehicle parameters The nonlinear transfer function, γ v is the vehicle loss coefficient, δ v (t) represents the degree of wear of the vehicle according to the mileage, and c represents the age of the vehicle.

[0052] Preferably, in S2, the extracted key features of each vehicle owner and vehicle are converted into a numerical vector through a large language model, specifically including:

[0053] Based on the large language model, the extracted driving behavior features F d 、Historical accident record characteristics F a and vehicle usage characteristics F v Convert to a numerical vector V total :

[0054]

[0055] in, is the vectorized function of driving behavior characteristics, and the parameters represents the conversion weight of the i-th driving behavior feature, n d represents the number of driving behavior characteristics, is a vectorized function of historical accident characteristics, with parameters represents the conversion weight of historical accident characteristics, n a represents the number of historical accident characteristics, is the vectorized function of the k-th vehicle usage feature, with parameters represents the conversion weight of the k-th vehicle usage feature, n v The number of vehicle usage characteristics.

[0056] Preferably, the S2 is stored in the form of a multi-dimensional feature matrix, specifically including:

[0057] The numerical vector V total With multidimensional feature matrix M features Store in the form of, and obtain the multi-dimensional feature matrix M features :

[0058]

[0059] Among them, V dz 、V az and Vvz are the numerical vectors of the driving behavior characteristics, historical accident characteristics, and vehicle usage characteristics of the z-th user, respectively.

[0060] Preferably, the S3 specifically comprises the following steps:

[0061] S31. Based on the PyTorch deep learning framework, the multidimensional feature matrix M features Input into the risk assessment model for training, and M features The input data X used as the risk assessment model is:

[0062] X=M features ;

[0063] S32. Using multi-layer neural network to build risk assessment model risk (X; θ), where θ is the risk assessment model parameter set, including the weight matrix and bias vector, and the calculation process is as follows:

[0064]

[0065] Among them, H (l) represents the neuron output of the lth layer, W (l) and b (l) are the weight matrix and bias vector of the lth layer respectively, σ(·) is the activation function, The risk value predicted by the risk assessment model;

[0066] S33. Use the insurance company's historical assessment data as label Y to construct the loss function L(θ) and optimize the risk assessment model parameter θ. The loss function combines the risk value interval [0,1000] and the weight factor w i Defined as:

[0067]

[0068] Where N is the number of samples, is the predicted risk value of the risk assessment model for the i-th sample, Y i is the actual risk value label, is the weight coefficient of the sample. The higher the risk value, the higher the weight w. i The larger the w i Used to emphasize the importance of high-risk samples to risk assessment model training, λ∥θ∥ 2 is a regularization term to prevent overfitting of the risk assessment model, and λ is a regularization coefficient;

[0069] S34. Iteratively update the risk assessment model parameters θ through the gradient descent algorithm and back propagation method:

[0070]

[0071] Among them, θ (t) is the risk assessment model parameter of the tth iteration, η is the learning rate, is the gradient of the loss function with respect to the parameters. After the risk assessment model is trained, the performance of the risk assessment model is evaluated using the validation set, and the mean absolute error E is calculated. val and the coefficient of determination R 2 :

[0072]

[0073] Among them, M is the number of samples in the validation set, is the predicted risk value of the risk assessment model for the jth validation sample, Y j is the actual risk value label, is the average value of the actual risk value of the validation set;

[0074] By evaluating E val and R 2 The prediction accuracy and fitting effect of the risk assessment model can be judged by the value, so as to decide whether it is necessary to adjust the risk assessment model parameters or retrain the risk assessment model.

[0075] Preferably, the S4 specifically comprises the following steps:

[0076] S41. The owner and vehicle behavior dataset D to be evaluated standard Input to the trained risk assessment model f risk (X; θ), generate the risk value of each car owner

[0077]

[0078] in, is the predicted risk value of the car owner, ranging from 0 to 1000, X is the input standardized car owner and vehicle behavior data set, and θ is the trained risk assessment model parameter;

[0079] S42. Based on the risk value of the car owner Car owners are classified into three risk levels: low risk, medium risk and high risk according to the preset risk value range. The classification criteria are as follows:

[0080] Low risk: Medium risk: High risk: The higher the risk value, the greater the accident risk for the car owner. The car owner will be provided with a corresponding risk assessment report based on the risk level.

[0081] S43. Dynamically adjust the car insurance premium based on the risk value of the car owner.

[0082] Preferably, the S5 comprises:

[0083] Generate a natural language report based on the big language model to explain the results of the risk assessment report. The natural language report content is automatically generated and combined with the driving behavior characteristics of the car owner, the vehicle usage characteristics and the historical accident record characteristics. The big language model automatically generates a report text containing risk analysis to explain the source of the calculated risk value and describe the impact of each feature separately, providing an explanation of the final risk value;

[0084] Specifically include:

[0085] S51. Based on driving behavior characteristics, the natural language report explains the impact of driving behavior variables, violation frequency and violation type on the calculation of risk value, and specifically explains the weights of different driving behavior variables in the risk assessment model;

[0086] S52. The natural language report analyzes the accident types, accident frequency and compensation amount of the car owner in the past based on the characteristics of historical accident records, explains the impact of historical accident records on the calculation of risk value, and explains the impact of the time weight of historical accidents on the calculation of risk value;

[0087] S53. The natural language report shall describe the safety performance of the vehicle in its current state, the vehicle loss coefficient, the potential relationship and impact of the maintenance items and maintenance frequency on the occurrence of failures of the item type based on the vehicle usage characteristics, and shall specifically explain the impact of the above factors on the calculation of the risk value;

[0088] S54. Based on the impact of the above-analyzed features on the calculation of risk values, all data are integrated to generate a final natural language report for explaining the risk value. The generated natural language report is formatted and output. Supported formats include PDF, HTML and Word documents.

[0089] Preferably, the multiple sources in S1 include internal systems of insurance companies, 4S stores, maintenance service platforms, vehicle sensors, traffic violation systems and third-party data platforms.

[0090] Compared with the prior art, the present invention provides a vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model, which has the following beneficial effects:

[0091] The present invention collects comprehensive owner and vehicle behavior data from multiple channels, conducts in-depth analysis of the data through a large language model, extracts driving behavior features, historical accident record features, and vehicle usage features that are highly correlated with auto insurance risks, and combines nonlinear conversion functions to accurately reflect the impact of various factors on risks; driving behavior features, historical accident record features, and vehicle usage features are used as key features, converted into numerical vectors and stored in the form of multi-dimensional feature matrices, providing high-quality input data for deep learning models; a risk assessment model constructed using the PyTorch framework, combined with loss function optimization and gradient descent algorithms, achieves efficient model training and prediction, and ensures the accuracy of auto insurance costs and the rationality of claims; and adjusts risk assessment results in real time through dynamic data updates, significantly improving the accuracy of auto insurance risk assessment and the speed of response to changes in owner behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 The overall flow chart of the vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model of the present invention;

[0093] Figure 2 A schematic diagram of the process of training the risk assessment model provided by the present invention. DETAILED DESCRIPTION

[0094] In order to make the technical means, creative features, objectives and effects of the present invention easy to understand, the present invention will be further described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0095] In order to address the shortcomings of the existing technology, such as Figure 1 , Figure 2 As shown, the present invention provides a vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model, comprising the following steps:

[0096] S1. Obtaining the behavior data of multiple users and constructing multiple owner and vehicle behavior data sets for standardization, wherein the owner and vehicle behavior data sets include insurance company historical assessment report data, driving behavior data, owner violation record data, accident record data, motor vehicle usage status data, and motor vehicle maintenance record data;

[0097] S2. Based on the large language model, the standardized car owner and vehicle behavior data set of each user is deeply analyzed to extract key features related to auto insurance risk, including driving behavior features, historical accident record features and vehicle usage features. The extracted key features of each user are converted into numerical vectors through the large language model and integrated and stored in the form of a multi-dimensional feature matrix;

[0098] S3. Input the multi-dimensional feature matrix into the PyTorch deep learning framework to train the risk assessment model, use the insurance company's historical assessment report data as labels to build a supervised learning model, and use the risk assessment model after multiple rounds of iterations and model optimization to predict the risk level of the car owner and the vehicle;

[0099] S4. Input the car owner and vehicle behavior data set of the user to be evaluated into the trained risk assessment model to generate a risk value, and generate a risk assessment result based on the risk level corresponding to the risk value to support the insurance company to dynamically adjust the auto insurance premium;

[0100] S5. Generate a natural language report based on the large language model, and explain the impact of key factors in the owner and vehicle behavior data set on the risk value calculation based on the risk assessment results;

[0101] S6. Monitor the behavior data of the owner and the vehicle in real time through the vehicle-mounted equipment, dynamically adjust the risk value, the generated risk level and the risk assessment report according to the latest data, and update them according to the real-time behavior data to reflect the changes in the owner's driving habits;

[0102] S7. Monitor the performance of the risk assessment model in real time and update the model parameters regularly.

[0103] Specifically, the S1 specifically includes:

[0104] Collect the behavior data of multiple users from multiple sources, and build multiple owner and vehicle behavior data sets D by integrating multi-source heterogeneous data. total , where multiple sources include insurance company internal systems, 4S stores, maintenance service platforms, vehicle sensors, traffic violation systems and third-party data platforms;

[0105] It should be noted that driving behavior data is used to record the driver's behavior data during historical driving, including speeding, sudden acceleration, sudden deceleration, turning, night driving, long-distance driving, fatigue driving and seat belt use;

[0106] The owner's violation record data is used to record the owner's violation behaviors in the historical driving process. The owner's violation record data includes the violation type, violation number, violation time, violation location and violation severity. The violation types include traffic signal violations, speeding violations, parking violations, driving behavior violations, certificate violations and license plate violations.

[0107] The accident record data is used to record the owner's historical accident records and insurance claim records. The historical accident records include the accident time, accident location, accident type, accident frequency, liability determination and loss extent. The insurance claim records include compensation items, compensation amounts and compensation items. Among them, the compensation items include brake system compensation, tire compensation, steering system compensation, engine compensation, gearbox compensation and exhaust system compensation;

[0108] The motor vehicle usage status data is used to record usage status information and vehicle status information, wherein the motor vehicle usage status information includes the age and mileage of the vehicle, and the vehicle status information includes the brake system status, tire status, steering system status, engine status, gearbox status and exhaust system status;

[0109] The motor vehicle maintenance record data is used to record the vehicle's maintenance records and maintenance records. The maintenance records include maintenance time, maintenance items, maintenance frequency, maintenance time, maintenance items and maintenance frequency. Among them, the maintenance items include brake system maintenance, tire maintenance, steering system maintenance, engine maintenance, gearbox maintenance and exhaust system maintenance.

[0110] The insurance company's historical assessment report data is used to record the insurance company's past risk assessment data on the vehicle owner and vehicle;

[0111] For the collected car owner and vehicle behavior dataset D total Perform data cleaning, which includes removing duplicate data, erroneous data, and correcting missing values, to generate a cleaned car owner and vehicle behavior dataset D clean ;

[0112] For the cleaned car owner and vehicle behavior dataset D clean Formatting is performed to convert multi-source heterogeneous data into a unified standard format to meet the requirements of model processing, and the formatted owner and vehicle behavior dataset D is obtained. formatted ;

[0113] For the formatted owner and vehicle behavior dataset D formatted Perform deduplication processing to remove redundant records, ensure data uniqueness, and finally generate a standardized car owner and vehicle behavior dataset D standard ;

[0114] The vehicle owner and vehicle behavior dataset D total :

[0115] D total =D c ∪D v ∪D a ∪D u ∪D m ∪D h ;

[0116] Among them, D c represents driving behavior data, D v Indicates violation record data, D a Indicates accident record data, D u Indicates the motor vehicle usage status data, D m Represents motor vehicle maintenance record data, including vehicle maintenance records and maintenance records, D h Represents the insurance company's historical assessment report data, including the insurance company's past risk assessment data on the vehicle owner and the vehicle.

[0117] Specifically, extracting key features related to auto insurance risks in S2 specifically includes the following steps:

[0118] S21. Use the owner’s driving behavior data as the driving behavior variable x d ,Driving behavior variables include: speeding behavior, rapid acceleration behavior, rapid deceleration behavior, turning behavior, night driving behavior, long-distance driving behavior, fatigue driving behavior and seat belt use behavior;

[0119] Based on the owner's violation record data, obtain the violation type and violation frequency, and construct the driving behavior variable x d The nonlinear transformation function f d (x d ), f d (x d ) is used to reflect the driving behavior variable x d The degree of impact on driving risk;

[0120] Based on the violation record data, the driving behavior data of the car owner in the period T before the violation occurred is obtained. The time period T is defined as a 60-second window before the violation event occurs. Within the time period T, each second is regarded as a discrete time point t. The driving behavior variables at each time point t in the time period T are extracted, and the function φ(v,t) is constructed to represent the driving behavior performance of the car owner at a specific time t.

[0121] Based on the driving behavior variable x d The nonlinear transformation function f d (x d), time period T and function φ(v,t), to obtain the driving behavior characteristics F of the owner d ;

[0122] S22. Based on the accident record data, obtain historical accident records;

[0123] Based on historical accident records, obtain the accident factor x of each accident a , the accident factor x a Including accident time, accident location, accident type, accident frequency, responsibility determination and loss extent;

[0124] Construct accident factor x a The nonlinear transformation function f a (x a ), to reflect the accident factor x a The degree of impact on driving risk;

[0125] Set the time decay factor β based on historical accident records a , to reflect the changes in the impact of accidents over time and to more accurately reflect the impact of recent accidents;

[0126] Construct the time function λ of the accident frequency a (t), time function λ a (t) reflect the changing trend of accident frequency over time by accumulating the number of accidents and applying moving average technology to reduce the impact of short-term fluctuations;

[0127] Based on the time function λ a (t), time decay factor β a and accident factor x a The nonlinear transformation function f a (x a ), obtain the owner's historical accident record feature F a ;

[0128] S23. Obtaining vehicle parameters x based on vehicle status information in the motor vehicle usage status data v , the vehicle parameters include brake system status, tire status, steering system status, engine status, gearbox status and exhaust system status;

[0129] Based on the insurance claim records of motor vehicle maintenance record data and accident record data, the support, confidence and lift between the same maintenance items and compensation items are calculated using the April algorithm;

[0130] The support degree indicates the frequency of a maintenance item and a corresponding compensation item appearing in all records, the confidence degree indicates the probability that the corresponding compensation item will also appear when a maintenance item appears, and the lift degree indicates the strength of association between the maintenance item and the corresponding compensation item;

[0131] Based on the improvement, adjust the corresponding vehicle parameter x v The adjustment factor w v , adjustment factor w v It is used to reflect the potential relationship and impact of maintenance items and maintenance frequency based on the failure of the item type;

[0132] Among them, if the lift is greater than 1, it means that the maintenance item and maintenance frequency are positively correlated with the failure of the item type; if the lift is equal to 1, it means that the maintenance item and maintenance frequency are not correlated with the failure of the item type; if the lift is less than 1, it means that the maintenance item and maintenance frequency are negatively correlated with the failure of the item type;

[0133] Construct vehicle parameters x v The nonlinear transformation function f v (x v ) to assess the safety performance of the vehicle in its current condition;

[0134] Set the vehicle loss coefficient γ v The vehicle loss coefficient is used to quantify the degree of loss of the vehicle based on its service life, and obtain the vehicle service life c, the vehicle loss coefficient γ v The product of the vehicle's service life c is taken as the vehicle's wear and tear degree;

[0135] Get the vehicle's wear function δ v (t), the expression of this function is: v (t) = 1-e -λ·mileage , where λ represents the wear rate parameter of the vehicle per kilometer traveled, and mileage represents the vehicle mileage;

[0136] Based on the vehicle parameters x v The nonlinear transformation function f v (x v ), adjustment coefficient w of vehicle parameters v , vehicle loss coefficient γ v , vehicle age c and vehicle wear function δ v (t), obtain the vehicle owner’s vehicle usage characteristics F v ;

[0137] Based on the above, the standardized car owner and vehicle behavior dataset of each user is deeply analyzed based on the large language model to extract the driving behavior features F related to auto insurance risk.d 、Historical accident record characteristics F a and vehicle usage characteristics F v , the calculation formula is as follows:

[0138]

[0139] Among them, F d represents the driving behavior characteristics, n d represents the total number of driving behavior variables, represents the weight of the i-th driving behavior variable, represents the i-th driving behavior variable, Represents driving behavior variables The nonlinear transfer function, α d is the long-term driving behavior preference adjustment coefficient, the time dimension T represents the duration of the behavior, the function φ(v,t) represents the driving behavior performance of the owner at a specific time t, and F a Represents the characteristics of historical accident records, is the weight of the jth accident factor, Indicates the factors of the accident The nonlinear transfer function, β a is the time decay factor of the accident, λ a (t) represents the time function of the accident frequency, F v Indicates the vehicle usage characteristics, represents the weight of the kth vehicle parameter, Indicates vehicle parameters The nonlinear transfer function, γ v is the vehicle loss coefficient, δ v (t) represents the degree of wear and tear of the vehicle according to the mileage, and c represents the age of the vehicle;

[0140] Specifically, in S2, the extracted key features of each car owner and vehicle are converted into numerical vectors through a large language model, including:

[0141] Based on the large language model, the extracted driving behavior features F d 、Historical accident record characteristics F a and vehicle usage characteristics F v Convert to a numerical vector V total :

[0142]

[0143] in, is the vectorized function of driving behavior characteristics, parameter θ di represents the conversion weight of the i-th driving behavior feature, nd represents the number of driving behavior characteristics, is a vectorized function of historical accident characteristics, with parameters represents the conversion weight of historical accident characteristics, n a represents the number of historical accident characteristics, is the vectorized function of the k-th vehicle usage feature, with parameters represents the conversion weight of the k-th vehicle usage feature, n v The number of features representing vehicle usage;

[0144] It should be noted that, in this embodiment, the vectorization functions of the driving behavior characteristics, historical accident characteristics, and vehicle usage characteristics are standard vectorization functions in the NumPy library;

[0145] According to the change amplitude information of the driving behavior characteristics, the historical accident characteristics and the vehicle usage characteristics, the conversion weights of the driving behavior characteristics, the historical accident characteristics and the vehicle usage characteristics are obtained. For example, when the change amplitude of the driving behavior characteristics is a, the conversion weight of the driving behavior characteristics is set to 1 / 3a;

[0146] Specifically, S2 is stored in the form of a multi-dimensional feature matrix, including:

[0147] The numerical vector V total With multidimensional feature matrix M features Store in the form of, and obtain the multi-dimensional feature matrix M features :

[0148]

[0149] Among them, V dz 、V az and V vz are the numerical vectors of the driving behavior characteristics, historical accident characteristics, and vehicle usage characteristics of the z-th user, respectively;

[0150] It should be noted that by converting the key features extracted in the above steps into numerical vectors through a large language model and storing them in the form of a multi-dimensional feature matrix, the data processing process is simplified and high-quality input data is provided for the deep learning model;

[0151] Specifically, S3 includes the following steps:

[0152] S31. Based on the PyTorch deep learning framework, the multidimensional feature matrix M features Input into the risk assessment model for training, and M features The input data X used as the risk assessment model is:

[0153] X=Mfeatures ;

[0154] S32. Using multi-layer neural network to build risk assessment model risk (X; θ), where θ is the risk assessment model parameter set, including the weight matrix and bias vector, and the calculation process is as follows:

[0155]

[0156] Among them, H (l) represents the neuron output of the lth layer, W (l) and b (l) are the weight matrix and bias vector of the lth layer respectively, σ(·) is the activation function, The risk value predicted by the risk assessment model;

[0157] S33. Use the insurance company's historical assessment data as label Y to construct the loss function L(θ) and optimize the risk assessment model parameter θ. The loss function combines the risk value interval [0,1000] and the weight factor w i Defined as:

[0158]

[0159] Where N is the number of samples, is the predicted risk value of the risk assessment model for the i-th sample, Y i is the actual risk value label, is the weight coefficient of the sample. The higher the risk value, the higher the weight w. i The larger the w i Used to emphasize the importance of high-risk samples to risk assessment model training, λ∥θ∥ 2 is a regularization term to prevent overfitting of the risk assessment model, and λ is a regularization coefficient;

[0160] S34. Iteratively update the risk assessment model parameters θ through the gradient descent algorithm and back propagation method:

[0161]

[0162] Among them, θ (t) is the risk assessment model parameter of the tth iteration, η is the learning rate, is the gradient of the loss function with respect to the parameters. After the risk assessment model is trained, the performance of the risk assessment model is evaluated using the validation set, and the mean absolute error E is calculated. val and the coefficient of determination R 2 :

[0163]

[0164] Among them, M is the number of samples in the validation set, is the predicted risk value of the risk assessment model for the jth validation sample, Y j is the actual risk value label, is the average value of the actual risk value of the validation set;

[0165] By evaluating E val and R 2 Values ​​are used to judge the prediction accuracy and fitting effect of the risk assessment model, so as to decide whether it is necessary to adjust the risk assessment model parameters or retrain the risk assessment model;

[0166] In this embodiment, a threshold Q can be set, and the threshold Q and E val For comparison, if E val If it is less than the threshold Q, it means that the prediction accuracy of the model is high; otherwise, it means that the prediction error of the model is large, and the risk assessment model parameters need to be adjusted or the risk assessment model needs to be retrained;

[0167] If R 2 If it is between 0.9 and 1.1, it means that the model has a good fitting effect; otherwise, it means that the fitting effect is not good and the risk assessment model parameters need to be adjusted or the risk assessment model needs to be retrained;

[0168] Specifically, S4 includes the following steps:

[0169] S41. The owner and vehicle behavior dataset D to be evaluated standard Input to the trained risk assessment model f risk (X; θ), generate the risk value of each car owner

[0170]

[0171] in, is the predicted risk value of the car owner, ranging from 0 to 1000, X is the input standardized car owner and vehicle behavior data set, and θ is the trained risk assessment model parameter;

[0172] S42. Based on the risk value of the car owner Car owners are classified into three risk levels: low risk, medium risk and high risk according to the preset risk value range. The classification criteria are as follows:

[0173] Low risk: Medium risk: High risk: The higher the risk value, the greater the accident risk for the car owner. The car owner will be provided with a corresponding risk assessment report based on the risk level.

[0174] S43. Dynamically adjust the car insurance premium based on the risk value of the car owner.

[0175] Specifically, S5 includes:

[0176] Generate a natural language report based on the big language model to explain the results of the risk assessment report. The natural language report content is automatically generated and combined with the driving behavior characteristics of the car owner, the vehicle usage characteristics and the historical accident record characteristics. The big language model automatically generates a report text containing risk analysis to explain the source of the calculated risk value and describe the impact of each feature separately, providing an explanation of the final risk value;

[0177] Specifically include:

[0178] S51. The natural language report explains the impact of driving behavior characteristics on risk assessment, including explaining the impact of driving behavior variables on driving risk, the impact on the calculation of risk value, and how to improve the risk level of the car owner based on the car owner's driving behavior data; based on the car owner's historical violation record data, explaining the impact of violation type and violation frequency on driving risk, the impact on the calculation of risk value, and how to improve the risk level of the car owner; when explaining the impact of driving behavior variables, violation frequency and violation type on the calculation of risk value, the natural language report also specifically explains the weights of different driving behavior variables in the risk assessment model;

[0179] S52. Based on the characteristics of historical accident records, the natural language report analyzes the types of accidents that the car owner has experienced in the past, the frequency of accidents, and the amount of compensation paid, and explains the impact of historical accident records on the calculation of risk values. If the car owner has multiple major accident records, the natural language report emphasizes that they have a greater impact on the risk value, thus explaining why the car owner is rated as a higher risk level; the natural language report explains the time weight of historical accidents, indicating that the more recent accidents have a more significant impact on the risk value, while the earlier accidents have a smaller impact on the risk value;

[0180] S53. The natural language report describes the safety performance of the vehicle in its current state based on the vehicle usage characteristics, describes the vehicle loss coefficient based on the vehicle's age and mileage, and describes the potential relationship and impact of maintenance items and maintenance frequency on the occurrence of failures of the item type based on the vehicle's maintenance records. The natural language report specifically explains the impact of the above factors on the calculation of the risk value;

[0181] S54. Based on the impact of the above-analyzed features on the calculation of risk values, all data are integrated to generate a final natural language report for explaining the risk value. The generated natural language report is formatted and output. Supported formats include PDF, HTML and Word documents.

[0182] In summary, the advantages of the present invention are as follows: the present invention conducts in-depth analysis of the owner and vehicle behavior data based on a large language model, extracts driving behavior characteristics, historical accident record characteristics and vehicle usage characteristics related to auto insurance risks, and by constructing a nonlinear conversion function, can more accurately reflect the impact of driving behavior, accident factors and vehicle status on risks; the driving behavior characteristics, historical accident record characteristics and vehicle usage characteristics are used as key features, converted into numerical vectors and stored in the form of a multi-dimensional feature matrix, which provides input data for the deep learning model, and the PyTorch deep learning framework is used to build a risk assessment model. The model parameters are iteratively updated through the gradient descent algorithm and the back propagation method, and the risk assessment results are adjusted in real time through dynamic data updates, which significantly improves the accuracy of auto insurance risk assessment and the response speed to changes in owner behavior; the user's risk value is calculated based on the risk assessment model, which provides more accurate support for insurance companies to formulate differentiated auto insurance cost strategies; and the natural language report improves the transparency of the assessment results and enhances the insurance company's understanding of risk levels and decision support.

[0183] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. A car owner behavior data-driven auto insurance risk assessment method based on a large language model, characterized by: The steps include: S1. Obtaining the behavior data of multiple users and constructing multiple owner and vehicle behavior data sets for standardization, wherein the owner and vehicle behavior data sets include insurance company historical assessment report data, driving behavior data, owner violation record data, accident record data, motor vehicle usage status data, and motor vehicle maintenance record data; S2. Based on the large language model, the standardized car owner and vehicle behavior data set of each user is deeply analyzed to extract key features related to auto insurance risk, including driving behavior features, historical accident record features and vehicle usage features. The extracted key features of each user are converted into numerical vectors through the large language model and integrated and stored in the form of a multi-dimensional feature matrix; S3. Input the multi-dimensional feature matrix into the PyTorch deep learning framework to train the risk assessment model, use the insurance company's historical assessment report data as labels to build a supervised learning model, and use the risk assessment model after multiple rounds of iterations and model optimization to predict the risk level of the car owner and the vehicle; S4. Input the car owner and vehicle behavior data set of the user to be evaluated into the trained risk assessment model to generate a risk value, and generate a risk assessment result based on the risk level corresponding to the risk value to support the insurance company to dynamically adjust the auto insurance premium; S5. Generate a natural language report based on the large language model, and explain the impact of key factors in the owner and vehicle behavior data set on the risk value calculation based on the risk assessment results; S6. Monitor the behavior data of the owner and the vehicle in real time through the vehicle-mounted equipment, dynamically adjust the risk value, the generated risk level and the risk assessment report according to the latest data, and update them according to the real-time behavior data to reflect the changes in the owner's driving habits; S7. Monitor the performance of the risk assessment model in real time and update the model parameters regularly.

2. According to the vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model in claim 1, the step S1 specifically comprises: Collect the behavior data of multiple users from multiple sources, and build multiple owner and vehicle behavior data sets D by integrating multi-source heterogeneous data. total ; The vehicle owner and vehicle behavior data set includes insurance company historical assessment report data, driving behavior data, vehicle owner violation record data, accident record data, vehicle usage status data, and vehicle maintenance record data; Driving behavior data is used to record the driver's behavior data during historical driving, including speeding, sudden acceleration, sudden deceleration, turning, night driving, long-distance driving, fatigue driving and seat belt use; The owner's violation record data is used to record the owner's violation behaviors during the historical driving process, and the owner's violation record data includes the violation type, violation number, violation time, violation location and violation severity; The accident record data is used to record the owner's historical accident records and insurance claim records. The historical accident records include the accident time, accident location, accident type, accident frequency, liability determination and loss extent. The insurance claim records include compensation items, compensation amounts and compensation items. The motor vehicle usage status data is used to record usage status information and vehicle status information, wherein the motor vehicle usage status information includes the age and mileage of the vehicle; The motor vehicle maintenance record data is used to record the vehicle maintenance record and maintenance record, and the maintenance record includes maintenance time, maintenance items, maintenance frequency, maintenance time, maintenance items and maintenance frequency; The insurance company's historical assessment report data is used to record the insurance company's past risk assessment data on the vehicle owner and vehicle; For the collected car owner and vehicle behavior dataset D total Perform data cleaning, which includes removing duplicate data, erroneous data, and correcting missing values, to generate a cleaned car owner and vehicle behavior dataset D clean ; For the cleaned car owner and vehicle behavior dataset D clean Formatting is performed to convert multi-source heterogeneous data into a unified standard format to meet the requirements of model processing, and the formatted owner and vehicle behavior dataset D is obtained. formatted ; For the formatted owner and vehicle behavior dataset D formatted Perform deduplication processing to remove redundant records, ensure data uniqueness, and finally generate a standardized car owner and vehicle behavior dataset D standard ; The vehicle owner and vehicle behavior dataset D total : D total =D c ∪D v ∪D a ∪D u ∪D m ∪D h ; Among them, D c represents driving behavior data, D v Indicates violation record data, D a Indicates accident record data, D u Indicates the motor vehicle usage status data, D m Represents motor vehicle maintenance record data, including vehicle maintenance records and maintenance records, D h Represents the insurance company's historical assessment report data, including the insurance company's past risk assessment data on the vehicle owner and the vehicle.

3. The vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model according to claim 2 is characterized in that: The key features related to the auto insurance risk are extracted in S2, specifically including the following steps: S21. Use the owner’s driving behavior data as the driving behavior variable x d ; Based on the owner’s violation record data, construct the driving behavior variable x d The nonlinear transformation function f d (x d ), f d (x d ) is used to reflect the driving behavior variable x d The degree of impact on driving risk; Based on the violation record data, the driving behavior data of the car owner in the period T before the violation occurred is obtained. The time period T is defined as a 60-second window before the violation event occurs. Within the time period T, each second is regarded as a discrete time point t. The driving behavior variables at each time point t in the time period T are extracted, and the function φ(v,t) is constructed to represent the driving behavior performance of the car owner at a specific time t. Based on the driving behavior variable x d The nonlinear transformation function f d (x d ), time period T and function φ(v,t), to obtain the driving behavior characteristics F of the owner d ; S22. Based on the accident record data, obtain historical accident records; Based on historical accident records, obtain the accident factor x of each accident a , the accident factor x a Including accident time, accident location, accident type, accident frequency, responsibility determination and loss extent; Construct accident factor x a The nonlinear transformation function f a (x a ), to reflect the accident factor x a The degree of impact on driving risk; Set the time decay factor β based on historical accident records a ; Construct the time function λ of the accident frequency a (t), time function λ a (t) reflect the changing trend of accident frequency over time by accumulating the number of accidents and applying moving average technology to reduce the impact of short-term fluctuations; Based on the time function λ a (t), time decay factor β a and accident factor x a The nonlinear transformation function f a (x a ) to obtain the owner's historical accident record feature F a ; S23. Obtaining vehicle parameters x based on vehicle status information in the motor vehicle usage status data v , the vehicle parameters include brake system status, tire status, steering system status, engine status, gearbox status and exhaust system status; Based on the insurance claim records of motor vehicle maintenance record data and accident record data, the support, confidence and improvement between the same maintenance items and compensation items are calculated through the Apriori algorithm; The support degree indicates the frequency of a maintenance item and a corresponding compensation item appearing in all records, the confidence degree indicates the probability that the corresponding compensation item will also appear when a maintenance item appears, and the lift degree indicates the strength of association between the maintenance item and the corresponding compensation item; Based on the improvement, adjust the corresponding vehicle parameter x v The adjustment factor w v , adjustment factor w v It is used to reflect the potential relationship and impact of maintenance items and maintenance frequency based on the failure of the item type; Construct vehicle parameters x v The nonlinear transformation function f v (x v ) to assess the safety performance of the vehicle in its current condition; Set the vehicle loss coefficient γ v The vehicle loss coefficient is used to quantify the degree of loss of the vehicle based on its service life, and obtain the vehicle service life c, the vehicle loss coefficient γ v The product of the vehicle's service life c is taken as the vehicle's wear and tear degree; Get the vehicle's wear function δ v (t), the expression of this function is: v (t) = 1-e -λ·mileage , where λ represents the wear rate parameter of the vehicle per kilometer traveled, and mileage represents the vehicle mileage; Based on the vehicle parameters x v The nonlinear transformation function f v (x v ), adjustment coefficient w of vehicle parameters v , vehicle loss coefficient γ v , vehicle age c and vehicle wear function δ v (t), to obtain the vehicle owner’s vehicle usage characteristics F v .

4. The vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model according to claim 3 is characterized in that: In S21-S23, based on the large language model, the standardized car owner and vehicle behavior data set of each user is deeply analyzed to extract the driving behavior features F related to the car insurance risk. d 、Historical accident record characteristics F a and vehicle usage characteristics F v , the calculation formula is as follows: Among them, F d represents the driving behavior characteristics, n d represents the total number of driving behavior variables, represents the weight of the i-th driving behavior variable, represents the i-th driving behavior variable, Represents driving behavior variables The nonlinear transfer function, α d is the long-term driving behavior preference adjustment coefficient, the time dimension T represents the duration of the behavior, the function φ(v,t) represents the driving behavior performance of the owner at a specific time t, and F a Represents the characteristics of historical accident records, is the weight of the jth accident factor, Indicates the factors of the accident The nonlinear transfer function, β a is the time decay factor of the accident, λ a (t) represents the time function of the accident frequency, F v Indicates the vehicle usage characteristics, represents the weight of the kth vehicle parameter, Indicates vehicle parameters The nonlinear transfer function, γ v is the vehicle loss coefficient, δ v (t) represents the degree of wear of the vehicle according to the mileage, and c represents the age of the vehicle.

5. The vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model according to claim 4 is characterized in that: In S2, the extracted key features of each car owner and vehicle are converted into numerical vectors through a large language model, specifically including: Based on the large language model, the extracted driving behavior features F d 、Historical accident record characteristics F a and vehicle usage characteristics F v Convert to a numerical vector V total : in, is the vectorized function of driving behavior characteristics, parameter θ di represents the conversion weight of the i-th driving behavior feature, n d represents the number of driving behavior characteristics, is a vectorized function of historical accident characteristics, with parameters represents the conversion weight of historical accident characteristics, n a represents the number of historical accident characteristics, is the vectorized function of the kth vehicle usage feature, parameter θ vk represents the conversion weight of the k-th vehicle usage feature, n v The number of vehicle usage characteristics.

6. The vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model according to claim 5 is characterized in that: The S2 is stored in the form of a multi-dimensional feature matrix, specifically including: The numerical vector V total With multidimensional feature matrix M features Store in the form of, and obtain the multidimensional feature matrix M features : Among them, V dz 、V az and V vz are the numerical vectors of the driving behavior characteristics, historical accident characteristics, and vehicle usage characteristics of the z-th user, respectively.

7. The vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model according to claim 6, characterized in that: The S3 specifically includes the following steps: S31. Based on the PyTorch deep learning framework, the multidimensional feature matrix M features Input into the risk assessment model for training, and M features The input data X used as the risk assessment model is: X=M features ; S32. Using multi-layer neural network to build risk assessment model risk (X; θ), where θ is the risk assessment model parameter set, including the weight matrix and bias vector, and the calculation process is as follows: Among them, H (l) represents the neuron output of the lth layer, W (l) and b (l) are the weight matrix and bias vector of the lth layer respectively, σ(·) is the activation function, The risk value predicted by the risk assessment model; S33. Use the insurance company's historical assessment data as label Y to construct the loss function L(θ) and optimize the risk assessment model parameter θ. The loss function combines the risk value interval [0,1000] and the weight factor w i Defined as: Where N is the number of samples, is the predicted risk value of the risk assessment model for the i-th sample, Y i is the actual risk value label, is the weight coefficient of the sample. The higher the risk value, the higher the weight w. i The larger the w i Used to emphasize the importance of high-risk samples to risk assessment model training, λ∥θ∥ 2 is a regularization term to prevent overfitting of the risk assessment model, and λ is a regularization coefficient; S34. Iteratively update the risk assessment model parameters θ through the gradient descent algorithm and back propagation method: Among them, θ (t) is the risk assessment model parameter of the tth iteration, η is the learning rate, is the gradient of the loss function with respect to the parameters. After the risk assessment model is trained, the performance of the risk assessment model is evaluated using the validation set, and the mean absolute error E is calculated. val and the coefficient of determination R 2 : Among them, M is the number of samples in the validation set, is the predicted risk value of the risk assessment model for the jth validation sample, Y j is the actual risk value label, is the average value of the actual risk value of the validation set; By evaluating E val and R 2 The prediction accuracy and fitting effect of the risk assessment model can be judged by the value, so as to decide whether it is necessary to adjust the risk assessment model parameters or retrain the risk assessment model.

8. The vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model according to claim 7, characterized in that: The S4 specifically comprises the following steps: S41. The owner and vehicle behavior dataset D to be evaluated standard Input to the trained risk assessment model f risk (X; θ), generating the risk value of each car owner in, is the predicted risk value of the car owner, ranging from 0 to 1000, X is the input standardized car owner and vehicle behavior data set, and θ is the trained risk assessment model parameter; S42. Based on the risk value of the car owner Car owners are classified into three risk levels: low risk, medium risk and high risk according to the preset risk value range. The classification criteria are as follows: Low risk: Medium risk: High risk: The higher the risk value, the greater the accident risk for the car owner. The car owner will be provided with a corresponding risk assessment report based on the risk level. S43. Dynamically adjust the car insurance premium based on the risk value of the car owner.

9. The vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model according to claim 8, characterized in that: The S5 includes: Generate a natural language report based on the big language model to explain the results of the risk assessment report. The natural language report content is automatically generated and combined with the driving behavior characteristics of the car owner, the vehicle usage characteristics and the historical accident record characteristics. The big language model automatically generates a report text containing risk analysis to explain the source of the calculated risk value and describe the impact of each feature separately, providing an explanation of the final risk value; Specifically include: S51. Based on driving behavior characteristics, the natural language report explains the impact of driving behavior variables, violation frequency and violation type on the calculation of risk value, and specifically explains the weights of different driving behavior variables in the risk assessment model; S52. The natural language report analyzes the accident types, accident frequency and compensation amount of the car owner in the past based on the characteristics of historical accident records, explains the impact of historical accident records on the calculation of risk value, and explains the impact of the time weight of historical accidents on the calculation of risk value; S53. The natural language report shall describe the safety performance of the vehicle in its current state, the vehicle loss coefficient, the potential relationship and impact of the maintenance items and maintenance frequency on the occurrence of failures of the item type based on the vehicle usage characteristics, and shall specifically explain the impact of the above factors on the calculation of the risk value; S54. Based on the impact of the above-analyzed features on the calculation of risk values, all data are integrated to generate a final natural language report for explaining the risk value. The generated natural language report is formatted and output. Supported formats include PDF, HTML and Word documents.

10. The vehicle owner behavior data-driven auto insurance risk assessment method based on a large language model according to claim 1, characterized in that: The multiple sources in S1 include the insurance company's internal system, 4S stores, maintenance service platforms, vehicle sensors, traffic violation systems and third-party data platforms.

Images