Intelligent connected vehicle technology risk assessment method based on heterogeneous dual machine learning

Abstract

The application discloses an intelligent networked vehicle technology risk assessment method based on heterogeneous double machine learning, constructs a hybrid space-time synthesis HSTS three-layer heterogeneous data environment, fuses multi-source data such as actuarial benchmarks, new energy collision tests and ecological driving behaviors and completes standardized preprocessing; a semi-parametric multiplication structure equation of a Tweedie compound Poisson-Gamma distribution is established to separate technical causal core parameters; a symmetric K-fold cross fitting strategy is adopted to estimate high-dimensional nuisance functions and propensity scores through double-branch machine learning; a Neyman orthogonal score function is constructed to eliminate regularization interpolation bias, and a pure safety bonus causal parameter of an intelligent driving algorithm after stripping hardware maintenance cost and driving behavior interference is solved; precise assessment of technology and risk decoupling is realized, the offset proportion of hardware maintenance premium to technical safety value is quantified, and the paradox that the more advanced the technology is, the higher the risk is assessed is solved.

Description

Technical Field

[0001] This invention relates to the field of big data analysis technology for intelligent connected vehicles, specifically to a technology based on the dual machine learning DML-Tweedie framework, which removes interference parameters, quantifies the relevant risks of integrated intelligent connected vehicle technology equipment, and achieves accurate risk assessment. Background Technology

[0002] With the rapid popularization of software-defined vehicles (SDVs) and intelligent connected vehicles (ICVs), advanced driver assistance systems (ADAS) have been widely applied to various types of vehicles, with sophisticated sensors such as LiDAR and cameras being integrated into vulnerable parts of the vehicle (such as bumpers). While this integration method can effectively reduce the frequency of accidents and improve vehicle driving safety, it has also given rise to a "maintenance paradox"—that is, even minor collisions can damage sophisticated sensors, leading to an exponential increase in hardware repair costs. This presents a new challenge to the risk assessment of intelligent connected vehicle technologies, as existing assessment technology systems have many significant shortcomings and cannot meet the needs of accurate assessment.

[0003] Currently, risk assessment methods for intelligent connected vehicles can be mainly divided into three categories: The first category is traditional statistical assessment methods, which have long relied on generalized linear models (GLM) and zero-inflated Poisson (ZIP) models, relying solely on static information such as vehicle displacement, vehicle price, and historical accident data. These methods cannot adapt to the high-dimensional and nonlinear characteristics of multi-source dynamic data in intelligent connected vehicles. The second category is vehicle-to-everything (V2X) big data assessment schemes. With the rise of intelligent connected technologies, these schemes mainly use simple kinematic indicators such as rapid acceleration / deceleration and speeding frequency as the basis for assessment. They can only preliminarily characterize driving-related risks and cannot focus on the risk contribution of the integrated technology equipment itself. The third category is general machine learning assessment schemes. Currently, some leading institutions in the industry are beginning to try using high-dimensional prediction models such as XGBoost and random forests, combined with interpretable tools such as SHAP, to conduct correlation analysis on vehicle collision frequency and collision severity. However, these methods have not solved the assessment bias caused by the "maintenance paradox" and lack effective causal inference capabilities.

[0004] From the perspective of the assessment methods used by various entities, technology research and development and evaluation departments mainly use methods such as real vehicle crash tests and virtual simulation tests to evaluate the theoretical accident reduction effect of ADAS (i.e., the technological safety dividend); relevant assessment agencies mainly focus on the correlation between vehicle collision frequency and collision severity to conduct risk assessments of integrated technology equipment. However, none of the above assessment methods have formed a scientific and comprehensive technical system, making it impossible to accurately quantify technical risks and support the optimization and upgrading of integrated technology equipment.

[0005] While existing assessment methods provide some support for risk management in intelligent connected vehicles (ICVs), the following significant technical shortcomings still exist in the context of the rapid development of ICVs, which are also technical problems that urgently need to be solved in the field: 1. The maintenance paradox obscures the safety value of technology: Most existing assessment methods are based on the results of hardware repairs after a vehicle collision, ignoring the impact of the integrated logic of precision components such as sensors on repair costs. The high hardware repair premium obscures the accident reduction value brought by ADAS algorithms in the assessment, making it impossible to accurately identify the safety value of the integrated technology equipment itself, resulting in the contradiction of "installing intelligent driving systems but presenting higher risk assessment results".

[0006] 2. Assessment bias caused by confounding bias: Existing assessment methods are mostly based on correlation analysis models, which lack causal decoupling capabilities and cannot effectively isolate various interference parameters. On the one hand, safe intelligent driving technologies are often tied to experienced drivers or high-end models, and existing models cannot separate the pure safety contribution of integrated technology devices from complex driver behavior noise (such as sudden braking habits). On the other hand, when traditional black-box machine learning models process high-dimensional telemetry data, they are prone to regularization bias due to the lack of causal inference filters, resulting in inaccurate estimation of the safety value of integrated technology devices.

[0007] 3. Attribution error in correlation models: Pure correlation assessment models (such as pure prediction XGBoost models) are prone to erroneous assessment logic, which may directly include high hardware maintenance costs in the assessment of technical variables, leading to the erroneous conclusion that "vehicles equipped with intelligent driving integrated devices are at higher risk". At the same time, traditional GLM and ZIP models cannot effectively handle high-dimensional nonlinear noise, which may significantly underestimate the real risk mitigation value of intelligent driving integrated technology devices, resulting in serious inadequacy of assessment accuracy.

[0008] 4. Poor adaptability to complex data: Collision-related data of intelligent connected vehicles have an extremely high zero-value inflation rate (approximately 79.28%) and obvious heavy-tailed distribution characteristics (consisting of Tweedie compound Poisson-gamma distribution). Existing evaluation systems cannot adapt to the characteristics of such semi-continuous data, making it difficult to balance the nonlinear relationship between collision frequency and collision severity, and the evaluation results lack stable statistical robustness.

[0009] Furthermore, existing assessment methods often focus on theoretical model construction, failing to employ a heterogeneous dual machine learning framework. They lack deep integration with actual in-vehicle data and various types of benchmark data, resulting in poor practicality and an inability to effectively identify the safety value of integrated technology equipment, thus hindering equipment optimization and upgrades. Therefore, developing an assessment method based on heterogeneous dual machine learning, capable of constructing a complete data environment, effectively removing interfering parameters, and accurately quantifying technical risks, has become a pressing technical challenge in the field of intelligent connected vehicles. Summary of the Invention

[0010] This invention proposes a risk assessment method for intelligent connected vehicle technology based on heterogeneous dual machine learning, which can solve the problems of difficulty in removing interference, insufficient accuracy, and poor implementation in the risk assessment of integrated technology equipment in existing intelligent connected vehicles. It provides a risk assessment method for intelligent connected vehicle technology based on heterogeneous dual machine learning, providing reliable technical support for the optimization of integrated technology equipment.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: A risk assessment method for intelligent connected vehicle technology based on heterogeneous dual machine learning, comprising the following steps performed using computer equipment: S1. Acquire actuarial benchmark data, new energy vehicle crash test data, and ecological driving behavior hybrid spatiotemporal synthetic HSTS data environment data with trajectory indicators, and perform standardized preprocessing on the acquired multi-source data to construct a basic dataset to support risk assessment. S2. Based on the hybrid spatiotemporal synthetic HSTS data environment, a semi-parametric multiplication structure equation is constructed. The core parameters of risk assessment are defined by a logarithmic link function as an expression containing a binary intervention variable D and a high-dimensional perturbation function g(X). The target causal parameter θ, which characterizes the safety value of intelligent driving technology, is separated from the complex confounding factors. S3. A symmetric K-fold cross-fit strategy is used to segment the basic dataset. A two-branch machine learning algorithm is used to perform unbiased estimation of the high-dimensional perturbation function g(X) and the propensity score m(X) respectively. S4. Based on the log-linked exponential dispersion model, construct a Neyman orthogonal score function that is adapted to the zero-value expansion and heavy-tailed distribution characteristics of intelligent connected vehicle collision data and conforms to the Tweedie composite Poisson-gamma distribution. S5. Using the high-dimensional disturbance function g(X) and the estimated propensity score m(X) obtained in step S3, substitute them into the Neyman orthogonal scoring function to solve the empirical scoring equation and calculate the target causal parameter θ after removing hardware maintenance costs and driver behavior interference. S6. Based on the target causal parameter θ obtained by the solution, output the risk assessment results of intelligent connected vehicle technology with decoupling of technology and risk, which are used for the optimization and safety management of integrated technology equipment in intelligent connected vehicles.

[0012] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0013] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0014] As can be seen from the above technical solution, the intelligent connected vehicle technology risk assessment method based on heterogeneous dual machine learning of the present invention achieves multiple technological innovations in the field of intelligent connected vehicle technology risk assessment, as follows: 1. Causal-driven model innovation: Unlike traditional correlation assessment models (such as GLM or pure XGBoost models), this invention is the first to apply the heterogeneous dual machine learning (DML) framework to the risk assessment of intelligent connected vehicle technology. By utilizing Neyman orthogonality, it effectively solves the complex confounding bias problem between driving behavior and technology exposure, significantly improving the accuracy and reliability of risk assessment.

[0015] 2. Innovative Data Distribution Adaptation: In response to the zero-value inflation and heavy-tailed distribution characteristics of collision-related data of intelligent connected vehicles, this invention constructs a causal score function adapted to the Tweedie composite Poisson-Gamma distribution (p=1.10), filling the technical gap in the application of high-dimensional nonlinear causal inference to semi-continuous risk assessment data, and improving the adaptability and statistical robustness of the assessment model to complex data.

[0016] 3. Innovation in multi-source data synthesis mechanism: A hybrid spatiotemporal synthesis (HSTS) environment is proposed to integrate actuarial benchmark data, new energy vehicle collision technology data and high-dimensional telemetry behavior data across domains. This effectively overcomes the technical bottleneck of unobservable counterfactual data in the field of autonomous driving and provides more comprehensive and reliable data support for risk assessment.

[0017] Specifically, the beneficial effects of this invention are as follows: At the model level, the Neyman orthogonal scoring function is used to make the estimation of the target causal parameter θ immune to the estimation error of the driving habit-related parameter g(X) in the first order, thus ensuring mathematical robustness and evaluation accuracy in the context of ultra-high dimensional variables. At the application level, risk assessment that achieves "separation of technology and risk" can provide accurate basis for the optimization of related technologies and safety management of intelligent connected vehicles, and avoid misjudging the value of technology. At the research and industry level, it provides a mathematical foundation-level "causal filter" to support researchers in conducting more granular heterogeneity effect (HTE) research on the safety of different intelligent driving functions such as AEB and NOA. At the same time, it provides scientific and technical support for the construction of a standard system for risk management throughout the entire life cycle of intelligent connected vehicles, guides OEMs to focus on the maintainability and integration cost of sensors, and encourages the industry to reduce hardware maintenance pressure through technological innovation. At the industry level, it can help domestically produced intelligent driving vehicles demonstrate the advanced nature of their intelligent driving algorithms through precise technical risk assessment evidence, providing technical support for their participation in international market competition.

[0018] In summary, this invention, relying on a heterogeneous dual machine learning framework and combining a hybrid spatiotemporal synthesis (HSTS) data environment, semi-parametric multiplicative structural equations, and Neyman orthogonal scoring functions, can effectively remove interfering parameters such as driving behavior and hardware maintenance, accurately separate and calculate the target causal parameter θ (pure risk reduction effect). It addresses the technical shortcomings of traditional assessment methods, such as regularization bias, poor data adaptability, and attribution errors, significantly improving the accuracy of risk assessment for intelligent connected vehicle technologies. It is adaptable to risk data with high zero-value inflation and heavy-tailed distribution, closely aligning with actual vehicle scenarios and demonstrating strong applicability. It can be widely applied in areas such as risk management, technology optimization, scientific research, and industry standard setting for integrated intelligent connected vehicle technologies and equipment, fully leveraging the technical advantages of heterogeneous dual machine learning to achieve precise quantification of technical safety value and promote the standardized development of the intelligent connected vehicle technology field. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall process of the risk assessment method based on heterogeneous dual machine learning in this embodiment of the invention. Figure 2 Here is a flowchart of the improved algorithm in this embodiment of the invention: Dual Machine Learning (DML) with K-fold Crossfit; Figure 3 This is a waterfall chart of claim losses used for quantitative risk attribution and offset ratio analysis in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] Step S1: Acquire actuarial benchmark data, new energy vehicle crash test data, and ecological driving behavior data with trajectory indicators to construct a hybrid spatiotemporal synthetic HSTS three-layer heterogeneous data environment, complete data standardization preprocessing, and provide reliable data support for risk assessment. Step S2: Based on the hybrid spatiotemporal synthetic data environment, construct a semi-parametric multiplication structure equation, define the core parameters of risk assessment as expressions containing binary intervention variables D and high-dimensional perturbation functions g(X) through a logarithmic link function, and separate the target causal parameter θ; Step S3: The samples in the data environment are segmented using a symmetric K-fold cross-fit strategy, and the high-dimensional perturbation function g(X) and the propensity score m(X) are estimated by a two-branch algorithm using a machine learning algorithm; Step S4: Based on the log-linked exponential dispersion model, construct the Neyman orthogonal score function suitable for the Tweedie composite Poisson-gamma distribution. ; Step S5: Using the estimated value obtained in step S3, solve the empirical score equation based on the Neyman orthogonal score function to calculate the target causal parameter θ, where θ is the pure risk reduction effect; Step S6: Based on the target causal parameter θ, output the risk assessment result of technology-risk separation to provide data support for the optimization and safety management of integrated technology equipment for intelligent connected vehicles; The following are detailed explanations: 1. Construct a Hybrid Spatiotemporal Synthesis (HSTS) data environment In step S1, the present invention addresses the counterfactual unobservability problem by constructing an HSTS environment, the core of which lies in joint calibration using three layers of heterogeneous data: (1) Actuarial layer data calibration: Define the probability pattern of claims occurrence. This embodiment is based on the freMTPL dataset for calibration, and the main parameters include: ①p (Tweedie power parameter): The estimated value is 1.10. This parameter determines that the distribution is between the Poisson (p=1, describing frequency) and the gamma (p=2, describing severity) distribution, indicating that the risk is dominated by the claim frequency. It is suitable for describing the low-speed collision scenarios in which intelligent driving systems intervene most frequently.

[0022] ②φ (discrete parameter): The estimated value is 19,056.29, which is used to finely characterize the nonlinear dispersion ratio relationship between the mean and variance.

[0023] (2) Technical layer data calibration: Define severity anchor points. Based on statistical data of vulnerable parts of 40 new energy vehicles, the core parameter is hardware repair premium. The data shows that for L2+ level and above vehicles, due to the high-value sensors being densely integrated in vulnerable areas such as the front / side of the vehicle, the claim expenditure is about RMB 1002.71 higher than the baseline.

[0024] (3) Behavioral layer data extraction: A high-dimensional interference space is defined. By accessing 30,000 ecological driving telemetry data collected from the vehicle network bus, high-dimensional dynamic features such as rapid acceleration, rapid deceleration, sharp turning speed, and following distance (TTC) are extracted. Real-time kinematic scores are synthesized through a weighted fusion algorithm. The real-time kinematic scores are used to quantify the driver's driving aggression and serve as the core hybrid control variable of the behavioral layer. All features are preprocessed using a standardized matrix and used as the control variable feature vector X.

[0025] 2. Constructing the ICV risk structure equation to separate the target parameters In step S2, based on the HSTS data environment constructed in step S1, a semi-parametric multiplicative structural equation adapted to the Tweedie composite Poisson-Gamma distribution is established to decouple the nonlinear production relationship between the dependent and independent variables:

[0026] The above equation can be expanded using the log-link function to be defined as a linear combination of a binary intervention variable D and a high-dimensional perturbation function g(X): The detailed definitions and actuarial implications of the relevant core parameters are as follows: (1) Y (response variable) represents the unit mileage claims expenditure or expected net premium generated by the vehicle during its life cycle. This variable has semi-continuous characteristics. In this embodiment, there is a zero claims phenomenon of about 79.28% (zero value inflation) and it is accompanied by right-skewed heavy tail characteristics. Therefore, it must be assumed that it follows a Tweedie distribution.

[0027] (2) D (Processing Variable / Technology Exposure): A binary indicator variable. D=1 represents that the vehicle is equipped with the intelligent driving technology to be evaluated (such as AEB, NOA); D=0 represents that the baseline conventional vehicle is not equipped with the technology. This parameter is determined by the vehicle's high-fidelity configuration list.

[0028] (3) θ (target causal parameter): The core estimation parameter of this invention. It represents the pure marginal reduction effect of the intelligent driving algorithm on accident risk after completely eliminating the interference of external hardware manufacturing / maintenance costs and filtering out differences in driver behavior. A negative value indicates a reduction in risk.

[0029] (4) g(X) (noise function): represents the driver's baseline risk function. This is a high-dimensional nonlinear function used to fit the driving habit noise captured in the telemetry data, thereby "filtering out" spurious associations (self-selection bias) caused by the driver being more robust or more aggressive.

[0030] (5) X (feature vector): contains a series of confounding factors, such as driver age, gender, mileage and real-time kinematic score (the average score in this embodiment is 2.61).

[0031] 3. Employ a symmetric K-fold cross-fit strategy for bi-branch unbiased estimation. To completely eliminate interpolation bias caused by regularization or overfitting in high-dimensional machine learning models, this step strictly follows... Figure 2 The flowchart of the improved algorithm shown performs symmetric K-fold cross-fitting: (1) Sample splitting: The basic dataset W is randomly divided into K mutually exclusive subsets (K=5 in this embodiment), denoted as follows: , ,..., .

[0032] (2) Symmetrical cross-training loop: Initiates a loop process for each k (from 1 to K). For the current loop step k, specifies the current fold. As a test set (reasoning set), specify that, except for The remaining dataset W\ This serves as an auxiliary training set. The system trains the model on the auxiliary training set and performs out-of-sample predictions on the test set, specifically employing a two-branch parallel machine learning architecture: ① Branch A (Propensity Score Estimation): Using a high-dimensional control variable X, an XGBoost classification model is trained on the training set and used to predict the preference score m on the test set. )=E[ | This branch is used to quantify the probability of self-selection based on vehicle configuration and driver profile, thereby controlling for self-selection bias at the vehicle purchase level.

[0033] ②Branch B (High-Dimensional Disturbance Function Estimation): The Tweedie Loss regression function is built into the machine learning model. It uses the high-dimensional control variable X to perform nonlinear fitting on the dependent variable Y, and outputs the baseline risk prediction sequence on the test set. This is used to accurately quantify baseline risk losses caused by driver behavior and physical noise.

[0034] 4. Construct a Neyman orthogonal score function suitable for the Tweedie distribution. In step S4, in order to eliminate the interpolation bias caused by machine learning regularization, combined with the attached... Figure 2 The algorithm flow of this invention involves constructing a specific orthogonal scoring function. :

[0035] The nonlinear multi-product structure of this orthogonal scoring function consists of the product of two physical residuals: (1) Handling residuals : This represents the net technology exposure purely from exogenous technology configuration after removing user profile factors such as personal background and behavioral characteristics through the propensity score model.

[0036] (2) Relative residuals It measures the actual observed claims payouts. The degree of relative deviation between the model's predicted risk and the risk after intelligent driving intervention and baseline risk noise correction. This specific orthogonal structure mathematically guarantees that the first-order partial derivatives of the estimation errors of the disturbance parameters g and m with respect to the target causal parameter θ are always equal to 0 in the expected sense (i.e., satisfying Neyman orthogonality and having first-order immunity to regularization bias caused by high-dimensional features).

[0037] 5. Solve the empirical score equation and calculate the pure safety bonus of the algorithm. After the system exits the K-fold loop in step S3 ( Figure 2 (If "end loop k?" is "yes"), the unbiased extrapolated predicted values ​​obtained from the full set of observation samples in their respective test folds are summarized. and An empirical scoring equation is constructed on the full dataset, where the sample mean of the full sample Neyman orthogonal scores is equal to...

[0038] The empirical moment equations described above are solved by inverse operation with respect to θ using a numerical zero-point solution algorithm (such as the Newton-Raphson iterative method). Due to the bias correction filtering mechanism of the Neyman orthogonal score function in step S4, the regularization interpolation bias of the bi-branch model is canceled out here, and the optimal estimated solution θ is obtained. In this embodiment, the precise calculation yields θ≈-1080.07 yuan (see...). Figure 3 ) 6. Output the risk assessment results and attribution gap calculation for "technology-risk separation". correspond Figure 1The "output application layer" is used in this system. The system uses the target causal parameter θ obtained in step S5, combined with the technical layer parameters calibrated in step S1, to perform the final attribution gap and value loss offsetting calculation, thereby achieving precise "technology-risk separation" risk management.

[0039] (1) Specific implementation of the pure safety bonus of the intelligent driving algorithm: The system uses the target causal parameter θ obtained in step S5 to output the pure safety bonus value in monetary units (yuan / vehicle / year). In this embodiment, θ≈-1080.07 yuan, which means that after removing hardware maintenance costs and driver behavior interference, the intelligent driving algorithm can reduce collision compensation losses by about 1080.07 yuan per vehicle per year. This value is obtained by solving the empirical moment equation of the Neyman orthogonal score function. Specifically, after setting the sample mean of the full sample Neyman orthogonal score to 0, the optimal estimate of θ is obtained by inverse solving the algorithm using the Newton-Raphson numerical iteration algorithm.

[0040] (2) Specific implementation of calculating the proportion of hardware repair premium offsetting technical safety value: The system uses the hardware claim premium calibrated by the technical layer in step S1 (1002.71 yuan in this embodiment) as the numerator and the absolute value of the pure risk reduction effect |θ| (1080.07 yuan in this embodiment) solved in step S5 as the denominator to calculate the value loss ratio = hardware claim premium / |θ| = 1002.71 / 1080.07 ≈ 92.84%. This ratio is output in percentage form, quantitatively revealing the degree to which the safety value of intelligent driving technology is offset by hardware repair costs.

[0041] (3) Specific implementation of identifying the real causes of increased collision losses in L2+ vehicles: The system outputs an attribution waterfall plot in step S6 ( Figure 3 The system provides a structured interpretation. Starting from the overall baseline risk loss of the vehicle as determined by the actuarial layer (1741.72 yuan in this example), subtracting the pure safety bonus of the algorithm (-1080.07 yuan) and adding the hardware repair premium (+1002.71 yuan), the final actual observed loss is 1664.36 yuan. By comparing the claims differences between L2+ vehicles and the baseline vehicle, the system identifies that approximately 57.57% of the claim increase is attributable to the sensor repair cost premium rather than the failure of the safety algorithm, thus correcting the assessment paradox that "the more advanced the technology, the higher the assessed risk."

[0042] (4) Specific implementation of causal decoupling: The system synchronously runs the traditional unbiased ordinary least squares correlation model (OLS) to obtain the correlation estimate (approximately -209.89 yuan in this embodiment), which is compared with θ≈-1080.07 yuan obtained by the DML orthogonal debiased model. The effectiveness of causal decoupling is verified by improving the recognition capability by more than 5 times. Finally, a comparison report is output to provide quantitative data support for sensor integration structure optimization, maintainability design, actuarial pricing of vehicle insurance, and the formulation of standards for risk management throughout the entire life cycle of intelligent connected vehicles.

[0043] The specific technical benefits of this invention are as follows: (1) Accurately restore the technological security dividend: Against the backdrop of the surge in hardware maintenance costs caused by the “maintenance paradox”, this model successfully restored the real pure risk reduction effect that was hidden (θ≈-1080.07 yuan). Compared with the traditional evaluation model, the technological risk avoidance value it identified was increased by more than 5 times, and the technological security value was accurately quantified.

[0044] (2) Quantitative offset ratio of technical value: Through model accounting and hedging analysis, it was accurately identified that about 92.84% of the technical value was offset by the high hardware maintenance premium, which clarified the boundary between technical safety value and hardware maintenance cost, and provided clear data reference for the structural optimization and maintainability design of integrated technology equipment for intelligent connected vehicles.

[0045] (3) Solving the misjudgment of technical risk assessment: The model reveals that the real reason for the 57.57% increase in collision-related losses of Tier 3 (L2+) vehicles is the sensor maintenance cost premium, rather than the failure of safety algorithms. It solves the assessment paradox of "the more advanced the technology, the higher the assessment risk" from the perspective of decoupling the causal relationship between technology and risk, and corrects the erroneous judgment and correlation attribution error of traditional models.

[0046] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0047] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0048] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the intelligent connected vehicle technology risk assessment methods based on heterogeneous dual machine learning in the above embodiments.

[0049] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0050] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0052] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0053] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A risk assessment method for intelligent connected vehicle technology based on heterogeneous dual machine learning, characterized in that, Includes the following steps: S1. Acquire actuarial benchmark data, new energy vehicle crash test data, and ecological driving behavior data with trajectory indicators to construct a hybrid spatiotemporal synthetic HSTS three-layer heterogeneous data environment. The actuarial benchmark data includes at least vehicle historical claim frequency data, claim amount data, policy risk exposure data, and vehicle static attribute data. Standardize and preprocess the acquired multi-source data to construct a basic dataset that supports risk assessment. S2. Based on the hybrid spatiotemporal synthetic HSTS data environment, a semi-parametric multiplication structure equation is constructed. The core parameters of risk assessment are defined by a logarithmic link function as an expression containing a binary intervention variable D and a high-dimensional perturbation function g(X). The target causal parameter θ, which characterizes the safety value of intelligent driving technology, is separated from the complex confounding factors. S3. A symmetric K-fold cross-fit strategy is used to segment the basic dataset. A two-branch machine learning algorithm is used to perform unbiased estimation of the high-dimensional perturbation function g(X) and the propensity score m(X) respectively. S4. Based on the log-linked exponential dispersion model, construct a Neyman orthogonal score function that is adapted to the zero-value expansion and heavy-tailed distribution characteristics of intelligent connected vehicle collision data and conforms to the Tweedie composite Poisson-gamma distribution. S5. Using the high-dimensional disturbance function g(X) and the estimated value of the tendency score m(X) obtained in step S3, substitute them into the Neyman orthogonal scoring function to solve the empirical scoring equation, and calculate the unbiased target causal parameter θ after removing hardware maintenance costs and driver behavior interference. The target causal parameter θ is the pure safety bonus corresponding to the intelligent driving algorithm. S6. Based on the target causal parameter θ obtained by the solution, output the risk assessment results of intelligent connected vehicle technology with decoupling of technology and risk, which are used for the optimization and safety management of integrated technology equipment in intelligent connected vehicles.

2. The method for risk assessment of intelligent connected vehicle technology based on heterogeneous dual machine learning according to claim 1, characterized in that: The parameters of the hybrid spatiotemporal synthetic HSTS data environment include an actuarial layer, a technical layer, and a behavioral layer; The actuarial layer calibrates the Tweedie power parameter p=1.10 and discrete parameter φ=19056.29 based on the freMTPL dataset, and defines the probability pattern of claims occurrence. The technology layer is based on statistical data of vulnerable parts of 40 new energy vehicles to calibrate the hardware claim premium brought about by sensor integration. The behavioral layer relies on 30,000 pieces of ecological driving telemetry data to extract high-dimensional dynamic features of rapid acceleration, rapid deceleration, and turning speed as hybrid control variables.

3. The method for risk assessment of intelligent connected vehicle technology based on heterogeneous dual machine learning according to claim 1, characterized in that: The expression for the semi-parametric multiplication structure equation in step S2 is: The detailed parameter definitions are as follows: Y is the response variable, representing the claim expense per unit mileage generated by the vehicle during its life cycle; D is a binary intervention variable; D=1 represents that the vehicle is equipped with the intelligent driving technology to be evaluated; D=0 represents the baseline configuration. θ is the target causal parameter characterizing the safety value of intelligent driving technology. It represents the pure risk reduction effect after eliminating all hardware cost interference and differences in driver behavior. g(X) is a high-dimensional disturbance function, representing the driver's baseline risk function; X is a feature vector containing the driver's age, gender, mileage, and a comprehensive driving behavior score. The comprehensive driving behavior score is calculated by weighted fusion of the number of rapid accelerations, the number of rapid decelerations, the speed of sharp turns, and the following distance (TTC) features in the vehicle telemetry data.

4. The method for risk assessment of intelligent connected vehicle technology based on heterogeneous dual machine learning according to claim 3, characterized in that: The parameters of the Neyman orthogonal scoring function in S4 include: m(X) is the propensity score: representing the conditional probability E[D|X] of predicting vehicle configuration intelligent driving technology based on feature vector X, used to eliminate exposure endogeneity bias caused by the correlation between driver behavior characteristics and technology selection; (D m(X)) represents the residual after processing; This refers to the relative residual.

5. The method for risk assessment of intelligent connected vehicle technology based on heterogeneous dual machine learning according to claim 4, characterized in that: In step S3, the symmetric K-fold crossfit is to randomly divide the dataset into K folds, with each fold serving as the test set and the rest as the training set. Two-branch machine learning includes: (1) Branch A uses the XGBoost model to train on the training set and predict the propensity score m(X) of the test set; (2) Branch B uses the Tweedie Loss loss function to train the model to fit the high-dimensional disturbance function g(X).

6. The method for risk assessment of intelligent connected vehicle technology based on heterogeneous dual machine learning according to claim 1, characterized in that: Step S6 outputs the risk assessment results of intelligent connected vehicle technology with decoupling of technology and risk, specifically including the following steps: (1) Output the pure risk reduction effect represented by the target causal parameter θ in monetary units. The value is obtained by solving the empirical moment equation using the Newton-Raphson numerical iteration algorithm after setting the sample mean of the full sample Neyman orthogonal score function to zero. (2) Using the pre-calibrated hardware compensation premium as the numerator and the absolute value of the target causal parameter θ as the denominator, calculate the ratio of the two numbers to obtain the percentage by which the hardware repair premium offsets the technical safety value. (3) The actual claim loss is calculated by subtracting the target causal parameter θ from the overall vehicle baseline risk loss and adding the hardware repair premium. The attribution of the claim increment of L2+ vehicle and baseline vehicle on sensor repair cost premium is compared and analyzed.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, the processor performs the steps of the method as described in any one of claims 1 to 6.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, the processor performs the steps of the method as described in any one of claims 1 to 6.

Images