Vehicle accident risk prediction method and device, storage medium, and equipment

Abstract

The application discloses a vehicle accident risk prediction method and device, a storage medium and equipment, relates to the technical field of artificial intelligence and financial technology, and mainly aims to solve the problem of poor prediction accuracy of vehicle accident risk. The method comprises the following steps: acquiring vehicle multi-source data, wherein the vehicle multi-source data is used for representing data collected from multiple sources of the vehicle; determining short-term features and long-term features corresponding to the vehicle multi-source data, and constructing a plurality of weight connection graphs of different time nodes based on the vehicle multi-source data; performing risk prediction on the short-term features, the long-term features and the weight connection graphs by using a risk prediction model that has been trained, to obtain a vehicle accident risk prediction result, wherein the risk prediction model is obtained by training a convolutional graph neural network based on the introduction of a self-attention mechanism.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence and financial technology, and in particular to a method, device, storage medium, and equipment for predicting vehicle accident risks. Background Technology

[0002] With the deep integration of insurance platforms and vehicle-to-everything (V2X) technologies, V2X and UBI (Usage-Based Insurance) technologies are rapidly becoming more widespread. They can provide favorable conditions for using deep learning technology to predict vehicle accident risks by utilizing high-frequency driving trajectories, vehicle conditions, and real-time road and weather information.

[0003] Currently, existing vehicle accident risk prediction typically uses time series algorithms as the data-driven basis to predict vehicle accident risk. However, when time series models are combined with vehicle business operations, it is difficult to simultaneously take into account both short-term fluctuations and long-term trends. For example, the high incidence of accidents in the short term during heavy rain may be diluted by the long-term low-risk history, resulting in an underestimation of the risk on that day. At the same time, time series models rely on historical patterns and are slow to respond to unforeseen factors (such as extreme weather, major events, etc.). Insufficient fusion of multi-source time series data can significantly ignore spatial dependence characteristics and make it difficult to capture the risk spillover between adjacent road sections or areas, thus affecting the accuracy of vehicle accident risk prediction. Summary of the Invention

[0004] In view of this, this application provides a method, device, storage medium, and equipment for predicting vehicle accident risks, with the main purpose of solving the problem of poor accuracy in existing vehicle accident risk prediction.

[0005] According to one aspect of this application, a method for predicting vehicle accident risk is provided, comprising: acquiring multi-source vehicle data, wherein the multi-source vehicle data is used to characterize data collected from multiple sources on the vehicle; The short-term and long-term features corresponding to the vehicle multi-source data are determined, and multiple weighted connection graphs at different time points are constructed based on the vehicle multi-source data. The risk prediction model, which has been trained, is used to predict the short-term features, the long-term features, and the weighted connection graph to obtain the vehicle accident risk prediction result. The risk prediction model is a convolutional graph neural network trained with a self-attention mechanism.

[0006] Furthermore, before obtaining the vehicle accident risk prediction result by performing risk prediction on the short-term features, the long-term features, and the weighted connectivity graph using the risk prediction model that has already been trained, the method further includes: A convolutional graph neural network with self-attention mechanism is constructed, the convolutional graph neural network including an input layer, a spatial layer, a fusion layer, a prediction layer and an output layer; Obtain a training sample set, which includes short-term samples, long-term samples, and weighted connection graph samples with multi-task prediction risk labels. The convolutional graph neural network is trained based on the training sample set to obtain the risk prediction model. The spatial layer performs weighted fusion on the input short-term samples, long-term samples, and weighted connection graph samples, and then inputs the fusion layer for multi-graph fusion. The prediction layer performs convolutional dimensionality reduction on the graph features after multi-graph fusion and outputs the results through the output layer.

[0007] Furthermore, the method also includes: The short-term and long-term samples are fused to perform feature fusion, and the fused feature samples are then weighted and fused with the weighted connection graph samples before being input into the fusion layer.

[0008] Furthermore, determining the short-term and long-term features corresponding to the multi-source vehicle data includes: Based on the recurrent neural network model that has completed model training, feature extraction is performed on the multi-source vehicle data to obtain real-time features. The multi-source vehicle data includes vehicle information, driving behavior, road information, environmental information, external event information, and financial business information. The historical feature sequence of the vehicle is obtained, and the real-time features are filtered based on the historical feature sequence to obtain short-term features and long-term features.

[0009] Furthermore, the construction of multiple weighted connection graphs at different time points based on the vehicle multi-source data includes: Calculate the weighted adjacency matrix of the vehicle multi-source data according to different time points; The weighted adjacency matrix is ​​normalized, and the connecting edges are determined based on a sparsity strategy. Multiple weighted connection graphs are constructed based on the connection edges and the normalized weighted adjacency matrix. These multiple weighted connection graphs include correlation graphs, causal graphs, K-nearest neighbor graphs, and temporal similarity graphs.

[0010] Furthermore, the vehicle accident risk prediction results include accident classification, accident frequency, and financial business quotas; the method also includes: When the accident classification matches a preset classification object or the number of accidents matches a preset accident threshold, the connection edges in the weighted connection graph are adjusted, and the business quota in the financial business information is adjusted. The business quota is used to represent the equity ratio of vehicle accidents in financial business.

[0011] Furthermore, after acquiring the multi-source vehicle data, the method further includes: The multi-source vehicle data is cleaned and missing values ​​are filled in. The cleaned and supplemented multi-source vehicle data is transformed into a feature matrix using a time-space mapping method to determine the short-term and long-term features corresponding to the multi-source vehicle data.

[0012] According to another aspect of this application, a vehicle accident risk prediction device is provided, comprising: The acquisition module is used to acquire multi-source vehicle data, which is used to characterize the data collected from multiple sources on the vehicle. The determination module is used to determine the short-term and long-term features corresponding to the vehicle multi-source data, and to construct multiple weighted connection graphs at different time points based on the vehicle multi-source data. The prediction module is used to perform risk prediction on the short-term features, the long-term features, and the weighted connection graph using a risk prediction model that has been trained to obtain the vehicle accident risk prediction result. The risk prediction model is trained based on a convolutional graph neural network with a self-attention mechanism.

[0013] Furthermore, the device also includes: a construction module and a training module. The building module is used to construct a convolutional graph neural network that incorporates a self-attention mechanism. The convolutional graph neural network includes an input layer, a spatial layer, a fusion layer, a prediction layer, and an output layer. The acquisition module is used to acquire a training sample set, which includes short-term samples, long-term samples, and weighted connection graph samples with multi-task prediction risk labels. The training module is used to train the convolutional graph neural network based on the training sample set to obtain the risk prediction model. The spatial layer performs weighted fusion on the input short-term samples, long-term samples and weighted connection graph samples and then inputs them to the fusion layer for multi-graph fusion. The prediction layer performs convolutional dimensionality reduction on the graph features after multi-graph fusion and outputs them through the output layer.

[0014] Furthermore, the device also includes: The fusion module is used to perform feature fusion on the short-term samples and the long-term samples, so as to perform weighted fusion of the fused feature samples with the weighted connection graph samples and input them to the fusion layer.

[0015] Furthermore, The determining module is specifically used to extract features from the multi-source vehicle data based on a recurrent neural network model that has been trained, to obtain real-time features. The multi-source vehicle data includes vehicle information, driving behavior, road information, environmental information, external event information, and financial business information. The module also obtains historical feature sequences of the vehicle and filters the real-time features based on the historical feature sequences to obtain short-term and long-term features.

[0016] Furthermore, The construction module is specifically used to calculate the weighted adjacency matrix of the vehicle multi-source data according to different time nodes; normalize the weighted adjacency matrix and determine the connection edges based on the sparsity strategy; and construct multiple weighted connection graphs based on the connection edges and the normalized weighted adjacency matrix. The multiple weighted connection graphs include correlation graphs, causal graphs, K-nearest neighbor graphs and temporal similarity graphs.

[0017] Furthermore, the vehicle accident risk prediction results include accident classification, accident frequency, and financial business quotas; the device also includes: The adjustment module is used to adjust the connection edges in the weighted connection graph and the business quota in the financial business information when the accident classification matches a preset classification object or the number of accidents matches a preset accident threshold. The business quota is used to represent the equity ratio of vehicle accidents in financial business.

[0018] Furthermore, the device also includes: The transformation module is used to clean the vehicle multi-source data and fill in missing values; it performs matrix transformation on the cleaned and supplemented vehicle multi-source data through time-space mapping to obtain a feature matrix, so as to determine the short-term and long-term features corresponding to the vehicle multi-source data.

[0019] According to another aspect of this application, a storage medium is provided that stores at least one executable instruction, which causes a processor to perform an operation corresponding to the vehicle accident risk prediction method described above.

[0020] According to another aspect of this application, a terminal is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described vehicle accident risk prediction method.

[0021] By employing the above technical solutions, the technical solutions provided in the embodiments of this application have at least the following advantages: This application provides a vehicle accident risk prediction method, apparatus, storage medium, and device. Compared with the prior art, the embodiments of this application acquire multi-source vehicle data, which is used to characterize data collected from multiple sources on the vehicle; determine the short-term and long-term features corresponding to the multi-source vehicle data, and construct multiple weighted connection graphs at different time points based on the multi-source vehicle data; and use a risk prediction model that has completed model training to predict the short-term features, the long-term features, and the weighted connection graphs to obtain the vehicle accident risk prediction result. The risk prediction model is trained based on a convolutional graph neural network with a self-attention mechanism, which improves the ability to capture complex spatiotemporal patterns, effectively integrates heterogeneous data from multiple sources such as vehicle networks, road networks, weather, and events, and incorporates multi-dimensional relationships such as temporal similarity and causal precedence into the modeling scope through multi-graph construction, graph attention networks, and spatial Transformers, making up for the problem of insufficient perception of spatial spillover risks, improving adaptability to scene changes, and greatly improving the prediction accuracy of vehicle accident risks.

[0022] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a vehicle accident risk prediction method provided in an embodiment of this application is shown; Figure 2 This illustration shows a schematic diagram of a model framework provided in an embodiment of this application; Figure 3 This paper shows a block diagram of a vehicle accident risk prediction device provided in an embodiment of this application; Figure 4 A schematic diagram of the structure of a terminal provided in an embodiment of this application is shown. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] The embodiments of this invention can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0027] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0028] Based on this, in one embodiment, the present invention provides a method for predicting vehicle accident risks. Taking the application of this method to computer equipment such as servers as an example, the server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms, such as an intelligent insurance platform.

[0029] This application provides a method for predicting vehicle accident risks, such as... Figure 1As shown, the method includes: 101. Obtain multi-source vehicle data.

[0030] In this embodiment, multi-source vehicle data is used to characterize data collected from multiple sources for vehicles. In the vehicle accident risk prediction scenario, to obtain data related to vehicle accidents and ensure the effectiveness and accuracy of risk prediction, multi-source vehicle data includes vehicle information, driving behavior, road information, environmental information, external event information, and financial business information. Data sources cover financial business (such as insurance), vehicle registration information, and driver registration information, etc., and this embodiment does not impose specific limitations. Vehicle information includes vehicle model, vehicle age, cumulative mileage, traffic violation records, etc. Driving behavior includes vehicle speed, number of sudden braking and sharp turns, nighttime driving duration, and driving time distribution, etc. Road information includes road grade, real-time traffic flow, congestion index, construction and accident warnings, etc. Environmental information includes rainfall, visibility, temperature, and weather warnings, etc. External event information includes information on holidays, sporting events, large gatherings, and the start of the school season, etc. Financial business information is used to characterize information related to vehicle financial transactions, including claims records (accident time, location, type, and amount, etc.) and insurance amounts, etc., and this embodiment does not impose specific limitations.

[0031] It should be noted that the multi-source data in this application embodiment can be collected based on direct user input or based on different data source systems. This application embodiment does not impose any specific limitations.

[0032] 102. Determine the short-term and long-term features corresponding to the multi-source vehicle data, and construct multiple weighted connection graphs at different time points based on the multi-source vehicle data.

[0033] In this embodiment, short-term features are used to characterize the features extracted by the vehicle from vehicle information, driving behavior, road information, environmental information, external event information, and financial business information within a predetermined short time period. Long-term features are used to characterize the features extracted by the vehicle from vehicle information, driving behavior, road information, environmental information, external event information, and financial business information within a predetermined long time period. The division between short and long time periods can be configured based on risk prediction needs, preferably one week as the short time period and one month as the long time period, to distinguish between short-term and long-term features. This embodiment does not impose specific limitations. Simultaneously, the current execution end constructs multiple weighted connection graphs at different time nodes based on vehicle information, driving behavior, road information, environmental information, external event information, and financial business information. These weighted connection graphs can include correlation graphs, causal graphs, K-nearest neighbor graphs, and temporal similarity graphs. That is, connection graphs are constructed with different structural relationships for the multi-source vehicle data to perform multi-dimensional risk prediction.

[0034] It should be noted that the weighted connection graph is a graph structure built based on the adjacent features of the data, and different weighted connection graphs are obtained for different building methods. This application does not make specific limitations on the embodiments.

[0035] 103. Using the risk prediction model that has been trained, perform risk prediction on the short-term features, the long-term features, and the weighted connection graph to obtain the vehicle accident risk prediction result.

[0036] In this embodiment of the application, the risk prediction model is trained based on a convolutional graph neural network with an introduced self-attention mechanism. In this case, the convolutional graph neural network includes an input layer, a spatial layer, a fusion layer, a prediction layer, and an output layer, so that long-term features, short-term features, and weighted connection graphs are used as inputs in sequence to perform risk prediction, and the vehicle accident risk prediction result is output through the output layer.

[0037] In a specific embodiment, such as Figure 2 As shown, the risk prediction model is trained based on a convolutional graph neural network with a self-attention mechanism. The convolutional graph neural network consists of a spatial neural network, a graph attention network, and a convolutional network for dimensionality reduction. They are connected through a fusion layer. The graph features weighted and fused in the spatial layer are input to the fusion layer for multi-graph fusion, followed by convolutional dimensionality reduction, and outputting the vehicle accident risk prediction result.

[0038] In some embodiments, such as Figure 2 As shown, normalized weights are obtained in the fusion layer based on the importance of each graph. Weighted fusion is performed to generate comprehensive spatial features. In the prediction layer, the convolutional neural network model first applies a temporal convolution (Conv2d-T) along the time dimension to the spatiotemporal feature tensor formed by stacking the fusion results at T time points, capturing dynamic change patterns across time. Then, a structural convolution (Conv2d-d) is applied along the spatial dimension to mine spatial structural correlations. Finally, the multi-task prediction results within the future time range of Δ are output, which are the vehicle accident risk prediction results, including the accident occurrence probability p (classification task), the number of accidents c (Poisson's count prediction), and the expected compensation amount m (regression or quantile regression). Furthermore, embodiments of this application can further calculate the comprehensive risk index r = cm and its confidence interval, providing a quantitative basis for risk assessment and business decision-making.

[0039] This application provides a method for predicting vehicle accident risks, which improves the ability to capture complex spatiotemporal patterns, effectively integrates heterogeneous data from multiple sources such as vehicle networks, road networks, weather, and events, and incorporates multi-dimensional relationships such as temporal similarity and causal precedence into the modeling scope through multi-graph construction, graph attention networks, and spatial Transformers. This makes up for the problem of insufficient perception of spatial spillover risks, improves the adaptability to scene changes, and greatly improves the accuracy of vehicle accident risk prediction.

[0040] In another embodiment of this application, for further definition and explanation, before the step of performing risk prediction on the short-term features, the long-term features, and the weighted connection graph using a risk prediction model that has completed model training to obtain the vehicle accident risk prediction result, the method further includes: Construct a convolutional graph neural network that incorporates a self-attention mechanism; Obtain the training sample set; The risk prediction model is obtained by training the convolutional graph neural network based on the training sample set.

[0041] To achieve vehicle accident risk prediction based on artificial intelligence, a convolutional graph neural network (CGN) with a self-attention mechanism is pre-constructed at the current execution end. This CGN includes an input layer, a spatial layer, a fusion layer, a prediction layer, and an output layer. The input layer receives short-term and long-term features from the temporal layer, and receives weighted connection graphs of various graph structures, which are then input to the spatial layer. Simultaneously, the spatial layer weights and fuses the input short-term, long-term, and weighted connection graph samples before inputting them to the fusion layer for multi-graph fusion. The prediction layer performs convolutional dimensionality reduction on the fused graph features and outputs the results through the output layer. Figure 2 As shown. Simultaneously, a training sample set is acquired to train the convolutional graph neural network model based on the training sample set, thereby obtaining the risk prediction model. In some embodiments, the training sample set includes short-term samples, long-term samples, and weighted connection graph samples with multi-task risk prediction labels. During model training, the short-term samples... Compared with long-term samples , can be combined into In the output value space layer, the space layer uses a convolutional graph neural network, including a Transformer, a graph attention network, and convolutional networks, to process each weighted connection graph sample. Information from each node is extracted and dynamically weighted to learn and output a predicted risk. .

[0042] In some embodiments, In another embodiment of this application, for further definition and explanation, the steps also include: The short-term and long-term samples are fused to perform feature fusion, and the fused feature samples are then weighted and fused with the weighted connection graph samples before being input into the fusion layer.

[0043] To improve the effectiveness of long-term and short-term features in predicting vehicle risk, the current execution end pre-fuses feature fusion on the short-term and long-term samples. The fused feature samples are then weighted and fused with the weighted connectivity graph samples before being input into the fusion layer. For example, Figure 2 As shown, in the time layer, short-term samples can be extracted and fused based on the GRU (Gated Recurrent Unit) algorithm. Compared with long-term samples Output At the spatial layer, GRU is a variant of recurrent neural network (RNN) that controls the flow of information through gating mechanisms, effectively handling long-term dependencies in sequential data.

[0044] It should be noted that the GRU algorithm contains two key gating mechanisms: the update gate, which determines how many short-term samples to retain and how many long-term samples to add; and the reset gate, which controls how many long-term samples to ignore in order to focus on the important features in the current multi-source data.

[0045] In another embodiment of this application, for further definition and explanation, the step of determining the short-term and long-term features corresponding to the multi-source vehicle data includes: Based on the recurrent neural network model that has completed model training, feature extraction is performed on the multi-source data of the vehicle to obtain real-time features; The historical feature sequence of the vehicle is obtained, and the real-time features are filtered based on the historical feature sequence to obtain short-term features and long-term features.

[0046] To improve the accuracy of vehicle accident risk prediction by deeply mining features effective for risk prediction from multidimensional data from both long-term and short-term perspectives, the current execution end extracts features from vehicle information, driving behavior, road information, environmental information, external event information, and financial business information based on a pre-trained recurrent neural network model when determining short-term and long-term features. This yields real-time features. Specifically, a pre-trained recurrent neural network model is first used to extract features from vehicle information, driving behavior, road information, environmental information, external event information, and financial business information, resulting in a dataset of real-time features that can be used to distinguish between long-term and short-term features. Next, the historical feature sequence of the vehicle is obtained. This historical feature sequence consists of historical data stored in the current execution end, such as historical claim vehicles, historical driving violations, and historical claim information. The historical feature sequence is then fused with the real-time features to obtain short-term and long-term features. In some implementation scenarios, the fused short-term features include the number of times vehicle A engages in normal driving behavior within a day, road congestion, frequency of rain and snow, the "back-to-school season" label 1, and the number of claims. Correspondingly, the long-term features include the number of times vehicle A engages in normal driving behavior within a week, road congestion, frequency of rain and snow, the "back-to-school season" label 0, and the number of claims. In this case, during the filtering process, the time period with a high accident frequency in the historical feature sequence can be selected as the filtering duration for short-term features, and the time period with a low accident frequency can be selected as the filtering duration for long-term features. The accident frequency can be set to a daily or weekly duration, preferably 10 times as the judgment threshold, to filter both short-term and long-term features.

[0047] It should be noted that recurrent neural networks (RNNs) are neural networks specifically designed for processing sequential data. By passing information through hidden states, they can capture temporal dependencies in the context and then extract features. In this embodiment, the recurrent neural network is trained with a large number of pre-trained samples and then called upon. The training samples may include feature samples corresponding to vehicle information, driving behavior, road information, environmental information, external event information, and financial business information, making the recurrent neural network more suitable for risk prediction business and improving the effectiveness of vehicle accident risk prediction. This embodiment does not impose specific limitations.

[0048] In another embodiment of this application, for further definition and explanation, the step of constructing multiple weighted connection graphs at different time points based on the vehicle multi-source data includes: Calculate the weighted adjacency matrix of the vehicle multi-source data according to different time points; The weighted adjacency matrix is ​​normalized, and the connecting edges are determined based on a sparsity strategy. Multiple weighted connection graphs are constructed based on the connecting edges and the normalized weighted adjacency matrix.

[0049] To achieve feature fusion based on spatial relationships and compensate for the insufficient perception of spatial spillover risks in traditional time-series models, the current execution end constructs multiple weighted connection graphs by first calculating the weighted adjacency matrix of vehicle multi-source data according to different time nodes, normalizing the weighted adjacency matrix, and determining the connection edges based on a sparsity strategy. Finally, multiple weighted connection graphs are constructed based on the connection edges and the normalized weighted adjacency matrix. In some embodiments, K weighted graphs are generated at each time slice t based on the spatial node set V of the multi-source data. ,in, The weighted adjacency matrix represents the multi-source data. After generating the adjacency matrix, it is normalized, and a sparsity strategy is used to retain strong connections with significant weights to reduce noise interference. Specifically, multiple weighted connection graphs include a correlation graph, a causal graph, a K-nearest neighbor graph, and a temporal similarity graph. The correlation graph represents the statistical correlation of historical indicators between nodes; the causal graph represents the causal dependencies obtained based on the Granger causality test; the K-nearest neighbor graph is constructed based on Euclidean distance or geographical location; and the temporal similarity graph is the temporal pattern similarity calculated based on Dynamic Time Warping (DTW). These are not specifically limited in the embodiments of this application.

[0050] In another embodiment of this application, for further definition and explanation, the steps also include: When the accident classification matches a preset classification object or the number of accidents matches a preset accident threshold, the connection edges in the weighted connection graph are adjusted, and the business quota in the financial business information is adjusted.

[0051] To provide more forward-looking and granular support for accurate pricing and claims resource allocation at the business level, the vehicle accident risk prediction results are analyzed after the vehicle failure risk prediction is completed. These accident risk prediction results may include accident classification, accident frequency, and financial business quota. Accident classification characterizes the type of predicted accident, such as traffic violation accidents or accidents. The business quota characterizes the proportion of vehicle accidents in financial business, such as the amount of accident compensation paid. This application's embodiment does not impose specific limitations on these aspects.

[0052] It should be noted that, in order to effectively update the weighted connection graph, the current execution end compares the accident type with the preset classification object and the accident count with the preset accident threshold. When the accident classification matches the preset classification object or the accident count matches the preset accident threshold, it indicates a sudden increase or decrease in accidents, or a targeted accident prediction purpose. Therefore, the current execution end adjusts the connection edges in the weighted connection graph to improve the accuracy of the next accident risk prediction and adjusts the business amount in the financial business information to achieve adaptive adjustment of insurance claims business. The preset classification object and preset accident threshold can be configured based on accident prediction needs, and this embodiment does not impose specific limitations.

[0053] In another embodiment of this application, for further definition and explanation, after acquiring the multi-source vehicle data, the method further includes: The multi-source vehicle data is cleaned and missing values ​​are filled in. The cleaned and supplemented multi-source vehicle data is transformed into a feature matrix using a time-space mapping method to determine the short-term and long-term features corresponding to the multi-source vehicle data.

[0054] To ensure data validity and improve the accuracy of vehicle accident prediction, the current execution terminal performs data cleaning and missing value imputation on the multi-source vehicle data after acquiring it. Then, it performs matrix transformation on the cleaned and imputed multi-source vehicle data using a time-space mapping method to obtain a feature matrix, thereby determining the short-term and long-term features corresponding to the multi-source vehicle data. During data cleaning and missing value processing, a data blacklist method can be used for cleaning, and fixed character imputation can be used; this application embodiment does not specify specific limitations. In some embodiments, the cleaned and missing value-processed data can be uniformly mapped to a fixed time granularity and spatial unit, ultimately forming a feature matrix at time t. This provides a consistent and high-quality input foundation for subsequent multi-graph modeling.

[0055] This application provides a method for predicting vehicle accident risks. Compared with existing technologies, this application acquires multi-source vehicle data, including vehicle information, driving behavior, road information, environmental information, external event information, and financial business information. It determines the short-term and long-term features corresponding to the multi-source vehicle data and constructs multiple weighted connection graphs at different time points based on the multi-source vehicle data. A risk prediction model that has been trained is used to predict the short-term features, the long-term features, and the weighted connection graphs to obtain the vehicle accident risk prediction result. The risk prediction model is trained based on a convolutional graph neural network with a self-attention mechanism, which improves the ability to capture complex spatiotemporal patterns and effectively integrates heterogeneous data from multiple sources such as vehicle networks, road networks, weather, and events. By using multi-graph construction, graph attention networks, and spatial Transformers, it simultaneously incorporates multi-dimensional relationships such as temporal similarity and causal precedence into the modeling scope, compensating for insufficient perception of spatial spillover risks, improving adaptability to scene changes, and greatly improving the accuracy of vehicle accident risk prediction.

[0056] Furthermore, as a response to the above Figure 1 The implementation of the method shown in this application provides a vehicle accident risk prediction device, such as... Figure 3 As shown, the device includes: The acquisition module 21 is used to acquire vehicle multi-source data, which is used to characterize the data collected from multiple sources for the vehicle. The determination module 22 is used to determine the short-term and long-term features corresponding to the vehicle multi-source data, and to construct multiple weighted connection graphs at different time points based on the vehicle multi-source data. The prediction module 23 is used to perform risk prediction on the short-term features, the long-term features and the weighted connection graph through the risk prediction model that has been trained, and to obtain the vehicle accident risk prediction result. The risk prediction model is trained based on a convolutional graph neural network with a self-attention mechanism.

[0057] Furthermore, the device also includes: a construction module and a training module. The building module is used to construct a convolutional graph neural network that incorporates a self-attention mechanism. The convolutional graph neural network includes an input layer, a spatial layer, a fusion layer, a prediction layer, and an output layer. The acquisition module is used to acquire a training sample set, which includes short-term samples, long-term samples, and weighted connection graph samples with multi-task prediction risk labels. The training module is used to train the convolutional graph neural network based on the training sample set to obtain the risk prediction model. The spatial layer performs weighted fusion on the input short-term samples, long-term samples and weighted connection graph samples and then inputs them to the fusion layer for multi-graph fusion. The prediction layer performs convolutional dimensionality reduction on the graph features after multi-graph fusion and outputs them through the output layer.

[0058] Furthermore, the device also includes: The fusion module is used to perform feature fusion on the short-term samples and the long-term samples, so as to perform weighted fusion of the fused feature samples with the weighted connection graph samples and input them to the fusion layer.

[0059] Furthermore, The determining module is specifically used to extract features from the multi-source vehicle data based on a recurrent neural network model that has been trained, to obtain real-time features. The multi-source vehicle data includes vehicle information, driving behavior, road information, environmental information, external event information, and financial business information. The module also obtains historical feature sequences of the vehicle and filters the real-time features based on the historical feature sequences to obtain short-term and long-term features.

[0060] Furthermore, The construction module is specifically used to calculate the weighted adjacency matrix of the vehicle multi-source data according to different time nodes; normalize the weighted adjacency matrix and determine the connection edges based on the sparsity strategy; and construct multiple weighted connection graphs based on the connection edges and the normalized weighted adjacency matrix. The multiple weighted connection graphs include correlation graphs, causal graphs, K-nearest neighbor graphs and temporal similarity graphs.

[0061] Furthermore, the vehicle accident risk prediction results include accident classification, accident frequency, and financial business quotas; the device also includes: The adjustment module is used to adjust the connection edges in the weighted connection graph and the business quota in the financial business information when the accident classification matches a preset classification object or the number of accidents matches a preset accident threshold. The business quota is used to represent the equity ratio of vehicle accidents in financial business.

[0062] Furthermore, the device also includes: The transformation module is used to clean the vehicle multi-source data and fill in missing values; it performs matrix transformation on the cleaned and supplemented vehicle multi-source data through time-space mapping to obtain a feature matrix, so as to determine the short-term and long-term features corresponding to the vehicle multi-source data.

[0063] This application provides a vehicle accident risk prediction device. Compared with the prior art, this application acquires multi-source vehicle data, which represents data collected from multiple sources on the vehicle. It determines the short-term and long-term features corresponding to the multi-source vehicle data and constructs multiple weighted connection graphs at different time points based on the multi-source vehicle data. A risk prediction model that has been trained is used to predict the short-term features, the long-term features, and the weighted connection graphs to obtain the vehicle accident risk prediction result. The risk prediction model is trained based on a convolutional graph neural network with a self-attention mechanism, which improves the ability to capture complex spatiotemporal patterns and effectively integrates heterogeneous data from multiple sources such as vehicle networks, road networks, weather, and events. Through multi-graph construction, graph attention networks, and spatial Transformers, it simultaneously incorporates multi-dimensional relationships such as temporal similarity and causal precedence into the modeling scope, compensating for insufficient perception of spatial spillover risks, improving adaptability to scene changes, and greatly improving the accuracy of vehicle accident risk prediction.

[0064] According to one embodiment of this application, a storage medium is provided, the storage medium storing at least one executable instruction that can execute the vehicle accident risk prediction method in any of the above method embodiments.

[0065] Figure 4 The diagram shows a structural schematic of a terminal according to one embodiment of the present application. The specific embodiments of the present application do not limit the specific implementation of the terminal.

[0066] like Figure 4 As shown, the terminal may include: a processor 302, a communications interface 304, a memory 306, and a communications bus 308.

[0067] The processor 302, communication interface 304, and memory 306 communicate with each other via communication bus 308.

[0068] Communication interface 304 is used to communicate with other network elements such as clients or other servers.

[0069] The processor 302 is used to execute program 310, which can specifically execute the relevant steps in the above-described vehicle accident risk prediction method embodiment.

[0070] Specifically, program 310 may include program code that includes computer operation instructions.

[0071] Processor 302 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The terminal includes one or more processors, which may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0072] Memory 306 is used to store program 310. Memory 306 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0073] Specifically, program 310 can be used to cause processor 302 to perform the following operations: Acquire multi-source vehicle data, which is used to characterize data collected from multiple sources for vehicles; The short-term and long-term features corresponding to the vehicle multi-source data are determined, and multiple weighted connection graphs at different time points are constructed based on the vehicle multi-source data. The risk prediction model, which has been trained, is used to predict the short-term features, the long-term features, and the weighted connection graph to obtain the vehicle accident risk prediction result. The risk prediction model is a convolutional graph neural network trained with a self-attention mechanism.

[0074] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0075] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for predicting vehicle accident risks, characterized in that, include: Acquire multi-source vehicle data, which is used to characterize data collected from multiple sources for vehicles; The short-term and long-term features corresponding to the vehicle multi-source data are determined, and multiple weighted connection graphs at different time points are constructed based on the vehicle multi-source data. The risk prediction model, which has been trained, is used to predict the short-term features, the long-term features, and the weighted connection graph to obtain the vehicle accident risk prediction result. The risk prediction model is a convolutional graph neural network trained with a self-attention mechanism.

2. The method according to claim 1, characterized in that, Before obtaining the vehicle accident risk prediction result by performing risk prediction on the short-term features, the long-term features, and the weighted connectivity graph using the risk prediction model that has been trained, the method further includes: A convolutional graph neural network with self-attention mechanism is constructed, the convolutional graph neural network including an input layer, a spatial layer, a fusion layer, a prediction layer and an output layer; Obtain a training sample set, which includes short-term samples, long-term samples, and weighted connection graph samples with multi-task prediction risk labels. The convolutional graph neural network is trained based on the training sample set to obtain the risk prediction model. The spatial layer performs weighted fusion on the input short-term samples, long-term samples, and weighted connection graph samples, and then inputs the fusion layer for multi-graph fusion. The prediction layer performs convolutional dimensionality reduction on the graph features after multi-graph fusion and outputs the results through the output layer.

3. The method according to claim 2, characterized in that, The method further includes: The short-term and long-term samples are fused to perform feature fusion, and the fused feature samples are then weighted and fused with the weighted connection graph samples before being input into the fusion layer.

4. The method according to claim 3, characterized in that, The determination of the short-term and long-term features corresponding to the multi-source vehicle data includes: Based on the recurrent neural network model that has completed model training, feature extraction is performed on the multi-source vehicle data to obtain real-time features. The multi-source vehicle data includes vehicle information, driving behavior, road information, environmental information, external event information, and financial business information. The historical feature sequence of the vehicle is obtained, and the real-time features are filtered based on the historical feature sequence to obtain short-term features and long-term features.

5. The method according to claim 1, characterized in that, The construction of multiple weighted connection graphs at different time points based on the vehicle multi-source data includes: Calculate the weighted adjacency matrix of the vehicle multi-source data according to different time points; The weighted adjacency matrix is ​​normalized, and the connecting edges are determined based on a sparsity strategy. Multiple weighted connection graphs are constructed based on the connection edges and the normalized weighted adjacency matrix. These multiple weighted connection graphs include correlation graphs, causal graphs, K-nearest neighbor graphs, and temporal similarity graphs.

6. The method according to any one of claims 1-5, characterized in that, The vehicle accident risk prediction results include accident classification, accident frequency, and financial business quotas; the method further includes: When the accident classification matches a preset classification object or the number of accidents matches a preset accident threshold, the connection edges in the weighted connection graph are adjusted, and the business quota in the financial business information is adjusted. The business quota is used to represent the equity ratio of vehicle accidents in financial business.

7. The method according to claim 1, characterized in that, After acquiring the multi-source vehicle data, the method further includes: The multi-source vehicle data is cleaned and missing values ​​are filled in. The cleaned and supplemented multi-source vehicle data is transformed into a feature matrix using a time-space mapping method to determine the short-term and long-term features corresponding to the multi-source vehicle data.

8. A vehicle accident risk prediction device, characterized in that, include: The acquisition module is used to acquire multi-source vehicle data, which is used to characterize the data collected from multiple sources on the vehicle. The determination module is used to determine the short-term and long-term features corresponding to the vehicle multi-source data, and to construct multiple weighted connection graphs at different time points based on the vehicle multi-source data. The prediction module is used to perform risk prediction on the short-term features, the long-term features, and the weighted connection graph using a risk prediction model that has been trained to obtain the vehicle accident risk prediction result. The risk prediction model is trained based on a convolutional graph neural network with a self-attention mechanism.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images