Risk control analysis method and insurance service system based on Internet of Vehicles accident big data

Abstract

The invention provides a risk control analysis method and system based on Internet of Vehicles accident big data, and the accuracy and robustness of risk control knowledge prediction are effectively improved. According to the method, multiple pieces of template accident monitoring data in Internet of Vehicles accident big data are acquired, and a data set containing a target risk control description tag and a disturbance risk control description tag is constructed for each piece of template accident monitoring data, so that the diversity of training samples is enriched. A sample Internet of Vehicles accident monitoring data sequence is generated in combination with an accident feature logic chain, and high-quality training data is provided for a risk control knowledge prediction network. Furthermore, parameter optimization is performed on the initialized risk control knowledge prediction network by using the sample Internet of Vehicles accident monitoring data sequences, and the finally generated target risk control knowledge prediction network can more accurately predict and identify the risk control knowledge category in the Internet of Vehicles accident, so that risk control and safety management in the Internet of Vehicles environment are facilitated, and the risk control accuracy is improved. And the intelligent level and practicability of risk control analysis are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of vehicle networking technology, and in particular to a risk control analysis method and system based on vehicle networking accident big data. Background Art

[0002] With the rapid development and popularization of Internet of Vehicles technology, information exchange between vehicles is becoming more and more frequent, and a large amount of Internet of Vehicles accident data is also generated. These accident data contain rich risk control information, which is of great significance for improving road traffic safety, optimizing vehicle design, and improving traffic management strategies. However, how to effectively extract and utilize risk control knowledge from massive Internet of Vehicles accident data has become a problem that needs to be solved urgently.

[0003] Traditional risk control analysis methods often rely on manual experience and expert knowledge to analyze accident data one by one. This method is not only inefficient, but also limited by human subjective judgment, making it difficult to fully and accurately mine risk control information from accident data. In recent years, although some risk control analysis methods based on machine learning have been proposed, most of these methods rely on a large amount of labeled data and are difficult to effectively capture the complex relationships and features in accident data.

[0004] In addition, the diversity and complexity of accident risk control description labels in the Internet of Vehicles accident data also poses challenges to risk control analysis. Different accidents may involve different risk control knowledge categories, and the same accident may also contain multiple risk control description labels. How to effectively process these risk control description labels and combine them with accident characteristics to generate representative sample data is the key to improving the effectiveness of risk control analysis. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a risk control analysis method based on Internet of Vehicles accident big data, the method comprising: Obtaining X template Internet of Vehicles accident monitoring data from the Internet of Vehicles accident big data, where X is a positive integer; For each template IoV accident monitoring data, obtain the accident risk control description label of the current template IoV accident monitoring data as the target risk control description label, and obtain the accident risk control description labels in Y other template IoV accident monitoring data from X-1 template IoV accident monitoring data as the disturbance risk control description label, where Y is a positive integer less than Z, and Z is the number of predefined risk control knowledge categories, which is the limit number of risk control knowledge categories in the prediction result predicted by the risk control knowledge label; Based on the target risk control description label, the disturbance risk control description label and the accident feature logic chain of each template IoV accident monitoring data, sample IoV accident monitoring data is generated by aggregation, and a sample IoV accident monitoring data sequence is generated, wherein the target risk control description label is a target risk control analysis result corresponding to the sample IoV accident monitoring data, and the sample IoV accident monitoring data sequence includes multiple sample IoV accident monitoring data, and each sample IoV accident monitoring data is generated based on one template IoV accident monitoring data; Based on the sample Internet of Vehicles accident monitoring data sequence, the network parameters of the initialized risk control knowledge prediction network are updated to generate a target risk control knowledge prediction network.

[0006] On the other hand, an embodiment of the present invention also provides an insurance service system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0007] Based on the above aspects, the embodiments of the present application effectively improve the accuracy and robustness of risk control knowledge prediction. The method first obtains multiple template accident monitoring data in the Internet of Vehicles accident big data, and constructs a data set containing target risk control description labels and disturbance risk control description labels for each template accident monitoring data, enriching the diversity of training samples. By combining the accident feature logic chain to generate a sample Internet of Vehicles accident monitoring data sequence, high-quality training data is provided for the risk control knowledge prediction network. Furthermore, these sample data are used to optimize the parameters of the initialized risk control knowledge prediction network, and the target risk control knowledge prediction network finally generated can more accurately predict and identify the risk control knowledge categories in Internet of Vehicles accidents, thereby facilitating risk control and safety management in the Internet of Vehicles environment, and significantly improving the intelligence level and practicality of risk control analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a schematic diagram of the execution flow of the risk control analysis method based on Internet of Vehicles accident big data provided by an embodiment of the present invention.

[0009] Figure 2 It is a schematic diagram of the hardware architecture of the insurance service system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0010] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of a risk control analysis method based on big data of vehicle networking accidents provided by an embodiment of the present invention. The risk control analysis method based on big data of vehicle networking accidents is introduced in detail below.

[0011] Step S110, obtaining X template Internet of Vehicles accident monitoring data in the Internet of Vehicles accident big data, where X is a positive integer.

[0012] In this embodiment, in the environment of the Internet of Vehicles, there is a large amount of accident-related data. For example, a large Internet of Vehicles service provider stores various information from many vehicles in its database. These vehicles are equipped with advanced sensors and communication equipment, which can transmit data such as vehicle status, environmental information and driver behavior in real time. When an accident occurs, these data are fully recorded to form Internet of Vehicles accident monitoring data. Assume that there are 1,000 Internet of Vehicles accident monitoring data in the database of this Internet of Vehicles service provider. Now, in order to carry out the subsequent construction and training of the risk control knowledge prediction network, it is necessary to select a certain number of template Internet of Vehicles accident monitoring data from them, where X=100 is set. The selection of these 100 template Internet of Vehicles accident monitoring data can be based on a variety of methods, such as random selection in chronological order, or selection after division according to the geographical location area where the accident occurred. For example, first classify according to the city where the vehicle is located, and then select from the accident data corresponding to each city according to a certain proportion to ensure that these 100 template Internet of Vehicles accident monitoring data are representative. Each template of IoV accident monitoring data contains a wealth of information, such as the speed change curve of the vehicle at the time of the accident (the speed drop from normal driving speed to the moment of the accident), status information of various vehicle components (for example, whether the brake system had abnormal temperature changes or pressure fluctuations before the accident), environmental information (including weather conditions at the time, such as whether it was raining or snowing, road visibility, and the slope and curvature of the road, etc.) and driver behavior information (such as the driver's operations before the accident, such as whether there were sudden turns, acceleration or braking operations, and the driver's fatigue level, which can be obtained by monitoring the driver's eye closure frequency through the in-car camera).

[0013] Step S120, for each template Internet of Vehicles accident monitoring data, obtain the accident risk control description label of the current template Internet of Vehicles accident monitoring data as the target risk control description label, and obtain Y accident risk control description labels in other template Internet of Vehicles accident monitoring data from X-1 template Internet of Vehicles accident monitoring data as disturbance risk control description labels, where Y is a positive integer less than Z, and Z is the number of predefined risk control knowledge categories, which is the limit number of risk control knowledge categories in the prediction result predicted by the risk control knowledge label.

[0014] Take the 100 template IoV accident monitoring data selected previously as an example. For one of the template IoV accident monitoring data (assuming it is the 30th data), first obtain its accident risk control description label as the target risk control description label. This accident risk control description label covers multiple aspects of the assessment of accident risk. For example, it may include a detailed analysis of the cause of the accident, such as "because the driver did not reasonably adjust the braking force according to the vehicle speed and road conditions on a slippery road, and the vehicle's braking system had slight wear and tear, resulting in reduced braking performance, the vehicle lost control and collided at the bend." This involves a comprehensive consideration of driver behavior, vehicle status and environmental factors. In addition, this accident risk control description label may also include a quantitative assessment of the accident risk, such as "according to the various data at the time of the accident, the risk level of this accident is high, with a probability of about 30%, and the estimated loss is 5,000 yuan for vehicle repair costs and 3,000 yuan for medical expenses (if there are injuries), and the indirect economic loss is about 1,000 yuan due to traffic congestion caused by the accident."

[0015] Then, determine the number of predefined risk control knowledge categories Z, assuming Z=20. Determine Y accident risk control description labels of other template Internet of Vehicles accident monitoring data in the remaining 99 template Internet of Vehicles accident monitoring data (i.e., X-1) as disturbance risk control description labels. For example, randomly determine Y=5. For example, one of the five disturbance risk control description labels may be "the vehicle is driving at night, the driver's vision is obstructed, and the road construction sign ahead is not found in time. The vehicle rushes into the construction area at a high speed and causes a collision. The risk level is high, and the probability of occurrence is about 40%. The estimated loss is serious damage to the vehicle, the repair cost is 8,000 yuan, there may be casualties, and the medical expenses are estimated to be 5,000 yuan (if there are casualties). At the same time, the construction equipment is damaged, and the compensation cost is 2,000 yuan." In this way, the acquisition operation of the target risk control description label and the disturbance risk control description label is performed for each template Internet of Vehicles accident monitoring data.

[0016] Step S130, based on the target risk control description label, the disturbance risk control description label and the accident feature logic chain of each template Internet of Vehicles accident monitoring data, aggregating and generating sample Internet of Vehicles accident monitoring data, and generating a sample Internet of Vehicles accident monitoring data sequence, wherein the target risk control description label is the target risk control analysis result corresponding to the sample Internet of Vehicles accident monitoring data, and the sample Internet of Vehicles accident monitoring data sequence includes multiple sample Internet of Vehicles accident monitoring data, and each sample Internet of Vehicles accident monitoring data is generated based on a template Internet of Vehicles accident monitoring data.

[0017] Still taking the 30th template IoV accident monitoring data mentioned above as an example. First, construct the accident feature logic chain of this data. Accident feature data is extracted from the template IoV accident monitoring data. In terms of vehicle status feature information, it is assumed that the vehicle's tire tread is severely worn (this can be obtained by detecting the tire tread depth data through the vehicle sensor), which belongs to the vehicle mechanical condition feature information; the vehicle's steering angle suddenly increased immediately before the accident, which is the vehicle control-related feature information. In terms of environmental status feature information, it was raining and the road was slippery, which is the natural environment feature information; the road is a mountain bend with a large curvature, which is the road environment feature information. In terms of driver behavior feature information, the driver frequently checked his mobile phone before the accident, which shows that he has poor driving habits. At the same time, through the analysis of the in-car camera, it was found that the driver's eyes showed signs of fatigue, which belongs to the driver's status feature information.

[0018] These accident feature data are classified. Severe tire wear and sudden increase in steering angle are trigger features that directly cause accidents; slippery roads on rainy days, large curvature of mountain roads, frequent checking of mobile phones by drivers, and fatigue driving are associated features that affect the development of accidents. The order of trigger features is analyzed. It is assumed that severe tire wear first affects the vehicle's handling, and then the driver's steering operation causes a sudden increase in the steering angle. The causal chain between trigger features is constructed. Then the influence mode of associated features on trigger features is analyzed. For example, slippery roads on rainy days make vehicles with severe tire wear more likely to lose control. Drivers' fatigue driving and frequent checking of mobile phones lead to steering errors at bends. The associated features are connected to the corresponding positions of the causal chain of trigger features according to the trigger features they affect, and the association relationship diagram between the associated features and the trigger features is constructed. Then the mutual influence between the associated features is analyzed. For example, fatigue driving may cause drivers to check their mobile phones more frequently. A two-way influence relationship chain is established. Finally, the causal relationship chain between trigger features, the association relationship diagram between associated features and trigger features, and the influence relationship chain between associated features are integrated to generate an integrated accident feature logic chain.

[0019] Next, the target risk control description label and the disturbance risk control description label are randomly combined and connected to generate risk control description label combination data. Assuming that the target risk control description label is the label mentioned above about accidents caused by braking on slippery roads and wear of the vehicle's brake system, and the disturbance risk control description label is the label of rushing into the construction area due to obstructed vision at night, a new logical description is formed after combination, such as "because the driver did not reasonably adjust the braking force according to the vehicle speed and road conditions on the slippery road, and the vehicle's brake system was slightly worn, resulting in reduced braking performance, and the vehicle's vision was obstructed at night and failed to timely discover the road construction signs ahead, the vehicle lost control at the bend and rushed into the construction area and collided". If the number of templates of the target risk control description label and the disturbance risk control description label is less than the global number of risk control knowledge labels of the target risk control knowledge prediction network, assuming that the global number of risk control knowledge labels is 15, and the number of templates after combination is 2, based on the number difference of 15-2=13, set labels are added to the risk control description label combination data, such as "it is not detected whether the vehicle's safety auxiliary system (such as the anti-lock braking system) is working properly" and other labels. Finally, the risk control description label combination data is aggregated with the accident feature logic chain of the current template IoV accident monitoring data to generate sample IoV accident monitoring data. For example, the combined risk control description label is integrated with the vehicle status, environmental status, and driver behavior information in the accident feature logic chain to form a complete sample IoV accident monitoring data, which contains detailed accident features, risk analysis, and other information. This operation is performed on each template IoV accident monitoring data to generate a sample IoV accident monitoring data sequence.

[0020] Step S140: updating the network parameters of the initialized risk control knowledge prediction network based on the sample Internet of Vehicles accident monitoring data sequence to generate a target risk control knowledge prediction network.

[0021] Assume that the initial risk control knowledge prediction network is a neural network model based on deep learning, such as a multi-layer convolutional neural network (CNN) or recurrent neural network (RNN) structure. First, after generating a sample Internet of Vehicles accident monitoring data sequence based on the target risk control description label, the disturbance risk control description label and the accident feature logic chain aggregation, the network parameters of the initial risk control knowledge prediction network are updated.

[0022] One way is to obtain two IoV accident monitoring data and the corresponding target risk control analysis results from the sample IoV accident monitoring data sequence. For example, two previously generated sample IoV accident monitoring data are selected, one is the sample data generated by the 30th template IoV accident monitoring data, and the other is the sample data generated by the 50th template IoV accident monitoring data. The corresponding target risk control analysis results are the risk assessment results contained in the risk control description tags determined previously. Based on the set fusion weights, the feature representations corresponding to the two IoV accident monitoring data are fused and calculated, and the feature representations corresponding to the corresponding two target risk control analysis results are fused and calculated. For example, the fusion weights are set to 0.6 and 0.4, and the vehicle state features, environmental features, driver behavior features, etc. are fused according to this weight. At the same time, the feature representations in the target risk control analysis results such as risk level, probability of occurrence, and expected loss are fused, and the fusion calculation results are used as the input data to be learned. Then, the input data to be learned is used to update the network parameters of the initialized risk control knowledge prediction network, such as adjusting the weight matrix and bias vector in the neural network. After multiple such operations, the network is gradually optimized and the target risk control knowledge prediction network is finally generated.

[0023] Another way is to divide the Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence into multiple training groups. For example, all the sample Internet of Vehicles accident monitoring data are evenly divided into 5 training groups. For each training group, the training group is loaded into the first candidate neural network and the second candidate neural network respectively for network parameter update, and both candidate neural networks generate network parameter configuration by initializing the risk control knowledge prediction network. In each candidate neural network, learning and parameter adjustment are performed according to the input sample Internet of Vehicles accident monitoring data and the corresponding target risk control analysis results. Then, the Internet of Vehicles accident monitoring data with different risk control knowledge prediction results in the first prediction learning data and the second prediction learning data are obtained as the Internet of Vehicles accident monitoring data to be optimized. Assuming that the risk level of a certain sample Internet of Vehicles accident monitoring data is predicted to be high in the first prediction learning data, and is predicted to be medium in the second prediction learning data, this sample data is determined as the Internet of Vehicles accident monitoring data to be optimized. Based on the error results of the IoV accident monitoring data to be optimized in the first prediction learning data, IoV accident monitoring data with a set proportion (assuming 30%) is selected from the IoV accident monitoring data to be optimized as the first batch, and based on the error results of the IoV accident monitoring data to be optimized in the second prediction learning data, IoV accident monitoring data with the same set proportion is selected from the IoV accident monitoring data to be optimized as the second batch. For example, if there are 10 IoV accident monitoring data to be optimized, 3 are selected in the first batch and 3 are also selected in the second batch. Then, the second candidate neural network is fine-tuned based on the first batch, and the first candidate neural network is fine-tuned based on the second batch. Finally, the fine-tuned first candidate neural network and the second candidate neural network are output as the target risk control knowledge prediction network.

[0024] Another way is to reorganize the sample IoV accident monitoring data and the disturbed IoV accident monitoring data in the sample IoV accident monitoring data sequence based on multiple predefined network training objectives to generate multiple target sample data corresponding to each network training objective. For example, the predefined network training objectives include improving the prediction accuracy of high-risk accidents, reducing the misjudgment rate of low-risk accidents, etc. For the goal of improving the prediction accuracy of high-risk accidents, select those sample data with high-risk target risk control analysis results from the sample IoV accident monitoring data sequence, and recombine them according to their accident feature logic chain and risk control description labels to generate corresponding target sample data. For each network training objective, select sample data from the corresponding target sample data as a customized training group. For example, select 50 sample data from the target sample data for improving the prediction accuracy of high-risk accidents as a customized training group. Then, based on this customized training group, the risk control knowledge label prediction of the corresponding network training objective is performed in the initialized risk control knowledge prediction network to generate prediction learning data for each network training objective. For example, after training this customized training group, we obtain prediction learning data for the accuracy of high-risk accident prediction. Finally, we train the initial risk control knowledge prediction network based on the prediction learning data of each network training target to generate the target risk control knowledge prediction network. In this way, the generated target risk control knowledge prediction network can have better performance in many aspects.

[0025] Based on the above steps, the embodiment of the present application effectively improves the accuracy and robustness of risk control knowledge prediction. The method first obtains multiple template accident monitoring data in the Internet of Vehicles accident big data, and constructs a data set containing target risk control description labels and disturbance risk control description labels for each template accident monitoring data, enriching the diversity of training samples. By combining the accident feature logic chain to generate a sample Internet of Vehicles accident monitoring data sequence, high-quality training data is provided for the risk control knowledge prediction network. Furthermore, these sample data are used to optimize the parameters of the initialized risk control knowledge prediction network, and the target risk control knowledge prediction network finally generated can more accurately predict and identify the risk control knowledge categories in Internet of Vehicles accidents, thereby facilitating risk control and safety management in the Internet of Vehicles environment, and significantly improving the intelligence level and practicality of risk control analysis.

[0026] In a possible implementation, step S120 includes: Step S121, obtaining any one of the X template Internet of Vehicles accident monitoring data as the current template Internet of Vehicles accident monitoring data.

[0027] Step S122, randomly determine a number as the disturbance label number Y in the number interval where the number of risk control knowledge labels predicted by the initialized risk control knowledge prediction network is the limit number.

[0028] Step S123: based on the disturbance tag quantity Y, obtaining the template Internet of Vehicles accident monitoring data except Y template Internet of Vehicles accident monitoring data from the X-1 template Internet of Vehicles accident monitoring data as other template Internet of Vehicles accident monitoring data.

[0029] In this embodiment, the previously mentioned Internet of Vehicles accident monitoring data is continued as an example. There are 100 template Internet of Vehicles accident monitoring data in the Internet of Vehicles service provider database, that is, X=100. In this process, first, any one of the template Internet of Vehicles accident monitoring data is obtained as the current template Internet of Vehicles accident monitoring data, for example, the 40th template Internet of Vehicles accident monitoring data is selected. Then, assuming that the limit number determined by the number of risk control knowledge labels predicted by the initialization risk control knowledge prediction network is between 10 and 20, a number is randomly determined within this interval as the number of disturbance labels Y, assuming that Y=12 is determined. Then, based on this number of disturbance labels Y, the template Internet of Vehicles accident monitoring data other than 12 template Internet of Vehicles accident monitoring data are obtained from the remaining 99 template Internet of Vehicles accident monitoring data (i.e., X-1) as other template Internet of Vehicles accident monitoring data. These 12 excluded template Internet of Vehicles accident monitoring data are randomly selected, and the remaining template Internet of Vehicles accident monitoring data will be used for the subsequent operation of obtaining disturbance risk control description labels.

[0030] In a possible implementation, step S130 includes: Step S131: randomly combine and connect the target risk control description tag and the disturbance risk control description tag to generate risk control description tag combination data.

[0031] Step S132: If the number of templates of the target risk control description label and the disturbance risk control description label is less than the global number of risk control knowledge labels of the target risk control knowledge prediction network, a setting label is added to the risk control description label combination data based on the difference between the global number of risk control knowledge labels and the number of templates.

[0032] Step S133, aggregating the risk control description tag combination data with the accident feature logic chain of the current template Internet of Vehicles accident monitoring data to generate sample Internet of Vehicles accident monitoring data.

[0033] In this embodiment, first, the 40th template Internet of Vehicles accident monitoring data has its corresponding accident risk control description label as the target risk control description label. This label describes the accident risk situation in detail, such as a sudden tire blowout when the vehicle is driving at high speed (mechanical condition characteristic information in the vehicle state characteristic information), improper operation by the driver in an emergency (driver behavior characteristic information), and icy road at the time (natural environment characteristic information in the environmental state characteristic information), which causes the vehicle to skid and collide with the roadside guardrail. The risk level is high, with a probability of occurrence of about 35%. The estimated vehicle repair cost is 8,000 yuan, there may be medical expenses for personal injuries estimated at 3,000 yuan, and traffic diversion costs of 1,000 yuan. At the same time, disturbance risk control description labels are obtained from other previously determined template Internet of Vehicles accident monitoring data. For example, one of them may be that when the vehicle was driving in a mountainous area, due to incorrect guidance of the navigation system (the road environment characteristic information associated with the environmental state characteristic information was inaccurate), the driver drove along the wrong route into an undeveloped road section, and the vehicle fell into a pothole (the control-related characteristic information in the vehicle state characteristic information was affected). The risk level is high, the probability of occurrence is about 30%, and the estimated rescue cost is 5,000 yuan.

[0034] The target risk control description label and the disturbance risk control description label are randomly combined and connected to generate risk control description label combination data. The combined label may be due to a sudden tire blowout when the vehicle was driving at high speed, the driver operated improperly in an emergency, and the road was icy at the time. At the same time, the vehicle was driving in the mountainous area due to incorrect guidance from the navigation system. The driver followed the wrong route and entered an undeveloped road section. The vehicle fell into a pothole, causing the vehicle to skid and collide with the roadside guardrail. The comprehensive risk level is assessed as high, and the probability of occurrence is approximately 38%. The estimated total cost includes vehicle repairs of 8,000 yuan, medical treatment of 3,000 yuan, traffic diversion of 1,000 yuan, and rescue of 5,000 yuan.

[0035] Assume that the global number of risk control knowledge labels of the target risk control knowledge prediction network is 18, and the number of templates after the target risk control description label and the disturbance risk control description label are combined is 2, and the difference between the two is 16. Based on this difference, set labels are added to the risk control description label combination data, such as adding vehicle safety systems (such as whether the vehicle stability control system is working normally when the accident occurs). Finally, this risk control description label combination data is aggregated with the accident feature logic chain of the 40th template Internet of Vehicles accident monitoring data to generate sample Internet of Vehicles accident monitoring data. The accident feature logic chain contains detailed information such as the tire pressure change before the tire burst, the driver's operating sequence at the moment of the tire burst, the specific range and thickness of the road ice, and the positioning information deviation when the navigation system misguides. This information is integrated with the risk control description label combination data to form a complete sample Internet of Vehicles accident monitoring data, which contains multi-dimensional information such as the risk situation analyzed from multiple aspects of the accident and detailed accident characteristics.

[0036] In a possible implementation, the step of generating the accident feature logic chain of the current template Internet of Vehicles accident monitoring data includes: Step A110, extracting accident feature data from the current template Internet of Vehicles accident monitoring data, the accident feature data including vehicle state feature information, environment state feature information and driver behavior feature information, the vehicle state feature information including vehicle mechanical condition feature information and vehicle control related feature information, the environment state feature information including natural environment feature information and road environment feature information, the driver behavior feature information including driving habit feature information and driver state feature information.

[0037] Step A120, classifying the accident feature data to generate trigger features that directly cause the accident and associated features that affect the development of the accident.

[0038] Step A130, analyze the sequence of each trigger feature in the accident process, and construct a causal chain between the trigger features based on the sequence. The construction of the causal chain is based on the logical order of event development, and the trigger feature that occurs first is taken as the cause, and the trigger feature that occurs later is taken as the effect, and they are connected in sequence to generate a directed causal chain.

[0039] Step A140, analyze the influence pattern of the associated features on the trigger features one by one, and according to the influence pattern, connect the associated features to the corresponding positions of the causal chain of the trigger features according to the trigger features they affect, and construct an association relationship diagram between the associated features and the trigger features.

[0040] Step A150, analyze the mutual influence between the associated features, establish a two-way influence relationship chain for the associated features that have a mutual influence relationship, and integrate the causal relationship chain between the trigger features, the association relationship diagram between the associated features and the trigger features, and the influence relationship chain between the associated features to generate an integrated accident feature logic chain.

[0041] In this embodiment, taking the 40th template vehicle networking accident monitoring data mentioned above as an example, first, the accident feature data is extracted from the 40th template vehicle networking accident monitoring data. In terms of vehicle status feature information, the vehicle mechanical condition feature information includes the wear degree of the vehicle tires, the wear of the brake system, and the running state of the engine. For example, the tire tread depth is only 2 mm after detection (under normal circumstances, the tread depth of new tires is about 8-10 mm), indicating that the tire is severely worn, which is a vehicle mechanical condition feature information; the brake pad thickness of the brake system is thin, and it is only 2 mm away from the limit wear thickness, which means that the braking ability of the brake system may be affected. Vehicle control related feature information includes the steering angle of the vehicle before the accident, the speed change, and the gear operation. For example, the steering angle of the vehicle before the accident suddenly increased, from about 5 degrees in normal driving to 30 degrees in an instant, and the speed dropped rapidly from 80 kilometers per hour to 20 kilometers per hour, and the gear shifting operation was frequently performed during this process.

[0042] The natural environment characteristic information in the environmental state characteristic information involves the weather conditions and light conditions at the time of the accident. In this accident, the weather condition was snowy, there was snow and ice on the road surface, and the visibility was low, about only 50 meters, which had a significant impact on the driving safety of the vehicle; the road environment characteristic information includes the type of road (such as expressways, mountain roads, etc.), the curvature and slope of the road, and the road surface condition (whether it is flat, whether there are potholes, etc.). This accident occurred at a bend in the mountainous area, with a large road curvature and a downhill section with a slope of 5%, and the road became very slippery due to snow and ice.

[0043] In terms of driver behavior characteristic information, driving habit characteristic information includes whether the driver has the habit of frequently changing lanes, speeding or sudden braking during driving. In this case, the driver has the habit of occasionally speeding in daily driving; driver status characteristic information includes the driver's fatigue level and concentration level at the time of the accident. Through in-car camera monitoring, it was found that the driver had frequent blinking and yawning before the accident, and his attention was not focused and he did not notice the changes in the road ahead in time.

[0044] Next, these accident feature data are classified to generate trigger features that directly cause accidents and associated features that affect the development of accidents. The trigger features that directly cause accidents include vehicle status feature information such as severe tire wear, reduced braking ability of the brake system, sudden increase in steering angle, and rapid decrease in vehicle speed. These features directly lead to the loss of control of the vehicle during driving and are the direct cause of the accident. The associated features that affect the development of accidents include snow, icy roads, and low visibility in the natural environment feature information, mountain bends and downhill sections in the road environment feature information, and fatigue driving and speeding habits in the driver behavior feature information.

[0045] Then, the sequence of each trigger feature in the accident process is analyzed, and the causal chain between the trigger features is constructed according to the sequence. First, due to severe tire wear, the friction between the vehicle and the ground is reduced. This is the first trigger feature and the initiating factor of the accident. It is regarded as the cause; because the friction is reduced, the braking distance of the vehicle becomes longer when braking, and the reduction in the braking ability of the brake system further aggravates this situation. Therefore, the reduction in the braking ability of the brake system is the second trigger feature, which is the result caused by the cause of severe tire wear. At the same time, it is the cause of the rapid decrease in vehicle speed; the rapid decrease in vehicle speed causes the driver to perform emergency steering operations to avoid collisions, resulting in a sudden increase in the steering angle. The sudden increase in the steering angle is the result of the rapid decrease in vehicle speed, which constructs a directed causal chain.

[0046] Afterwards, the influence patterns of associated features on trigger features are analyzed one by one, and according to the influence patterns, the associated features are connected to the corresponding positions of the causal chain of the trigger features according to the trigger features they affect, and a correlation relationship diagram between the associated features and the trigger features is constructed. The two associated features of snowfall and road icing have an impact on the trigger feature of severe tire wear, because snow accumulation and ice will accelerate the wear rate of tires, so snowfall and road icing are connected to the causal chain of the trigger feature of severe tire wear; the two associated features of mountain bends and downhill sections have an impact on the two trigger features of rapid drop in speed and sudden increase in steering angle. When driving on mountain bends and downhill sections, the vehicle itself needs to be more cautiously controlled. In this case, the rapid drop in speed and the sudden increase in steering angle are more likely to cause accidents, so mountain bends and downhill sections are connected to the corresponding positions of the causal chain of these two trigger features; the two associated features of driver fatigue driving and the habit of speeding also have an impact on the rapid drop in speed and the sudden increase in steering angle. Fatigue driving will make the driver slow to react, and the habit of speeding will make the vehicle more difficult to control when encountering emergencies. Therefore, these two associated features are connected to the corresponding trigger feature causal chain, thereby constructing a correlation relationship diagram between associated features and trigger features.

[0047] Finally, the mutual influence between the associated features is analyzed. For the associated features with mutual influence, a two-way influence relationship chain is established, and the causal relationship chain between the trigger features, the association relationship diagram between the associated features and the trigger features, and the influence relationship chain between the associated features are integrated to generate the integrated accident feature logic chain. For example, there is a mutual influence relationship between the two associated features of snow and road icing. Snow will cause road icing, and road icing will make the road conditions more complicated after snowing, so a two-way influence relationship chain is established; there may also be mutual influence between the two associated features of driver fatigue driving and speeding habits. Fatigue driving may make the driver more likely to speed, and speeding habits may also increase the driver's fatigue level. A two-way influence relationship chain is also established. These two-way influence relationship chains are integrated with the causal relationship chain between the trigger features and the association relationship diagram between the associated features and the trigger features previously constructed to form a complete accident feature logic chain. This accident feature logic chain comprehensively reflects the complex relationship between various factors causing accidents in the 40th template Internet of Vehicles accident monitoring data, including the cause-effect relationship and mutual influence relationship between vehicle status feature information, environmental status feature information and driver behavior feature information, providing an important basis for subsequent risk control analysis and the generation of sample Internet of Vehicles accident monitoring data.

[0048] In a possible implementation, before step S140, this embodiment can also aggregate based on the target risk control description label, the disturbance risk control description label and the accident feature logic chain of the current template Internet of Vehicles accident monitoring data to generate disturbance Internet of Vehicles accident monitoring data and add it to the sample Internet of Vehicles accident monitoring data sequence, wherein the disturbance risk control description label is the target risk control analysis result corresponding to the disturbance Internet of Vehicles accident monitoring data.

[0049] Step S140 includes: using the sample Internet of Vehicles accident monitoring data and the disturbed Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence to update the network parameters of the initialized risk control knowledge prediction network to generate the target risk control knowledge prediction network.

[0050] In this embodiment, the 40th template vehicle network accident monitoring data is taken as an example. Its target risk control description label includes the following details: the vehicle suddenly bursts when driving at high speed, the driver operates improperly, the road is icy, causing the vehicle to skid and collide with the guardrail, the risk level is high, and various expenses are expected; the disturbance risk control description label is such as the vehicle enters an undeveloped road section in the mountainous area due to navigation errors and falls into a pothole, the risk level is high and related expenses; the accident feature logic chain includes the tire pressure change before the vehicle bursts, the steering angle change, the range and thickness of the road ice, the positioning deviation during navigation errors, and other details. Aggregate these three to generate disturbance vehicle network accident monitoring data. For example, the risk factors in the target risk control description label and the special circumstances in the disturbance risk control description label are integrated according to the actual situation in the accident feature logic chain. The integrated disturbed IoV accident monitoring data records in detail the complete accident situation in which the vehicle itself has a risk of tire blowout and the road is icy, and the vehicle enters a special road section due to navigation misleading and falls into a pothole, causing the vehicle to slide sideways and collide with the guardrail, including the various risk factors involved, the state changes during the accident, and the estimated risk level and cost. Then the generated disturbed IoV accident monitoring data is added to the sample IoV accident monitoring data sequence.

[0051] In a possible implementation, the method of updating the network parameters of the initialized risk control knowledge prediction network using the sample Internet of Vehicles accident monitoring data and the disturbed Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence to generate the target risk control knowledge prediction network includes: Two vehicle networking accident monitoring data and corresponding target risk control analysis results are obtained from the sample vehicle networking accident monitoring data sequence.

[0052] Based on the set fusion weight, the feature representations corresponding to the two Internet of Vehicles accident monitoring data are fused and calculated, and the feature representations corresponding to the two corresponding target risk control analysis results are fused and calculated, and the fusion calculation results are used as input data to be learned.

[0053] Based on the generated input data to be learned, the network parameters of the initialized risk control knowledge prediction network are updated to generate the target risk control knowledge prediction network.

[0054] In this embodiment, one of the methods is as follows: two vehicle network accident monitoring data and corresponding target risk control analysis results are obtained from the sample vehicle network accident monitoring data sequence. For example, the sample vehicle network accident monitoring data generated by the 40th template vehicle network accident monitoring data and the sample vehicle network accident monitoring data generated by the 50th template vehicle network accident monitoring data are selected from the sample vehicle network accident monitoring data sequence. The sample vehicle network accident monitoring data corresponding to the 40th template accident monitoring data includes the previously integrated accident situation and risk assessment, and its target risk control analysis results are high risk level, specific probability of occurrence, and detailed cost estimate, etc.; the sample vehicle network accident monitoring data corresponding to the 50th template accident monitoring data may be another accident situation, such as a vehicle in foggy weather (natural environment characteristic information in environmental state characteristic information), due to fog lamp failure (vehicle mechanical condition characteristic information in vehicle state characteristic information), the driver failed to timely discover the slow-moving vehicle in front (driver state characteristic information in driver behavior characteristic information) and rear-ended, the risk level is high, and there are corresponding target risk control analysis results such as probability of occurrence and cost estimate.

[0055] Based on the set fusion weights, the feature representations corresponding to the two IoV accident monitoring data are fused and calculated, and the feature representations corresponding to the two corresponding target risk control analysis results are fused and calculated. Assume that the set fusion weights are 0.4 and 0.6. For the feature representation fusion calculation of IoV accident monitoring data, in terms of vehicle status feature information, factors such as tire wear, brake system, engine status (the 40th template-related situation) and fog lamp failure (the 50th template-related situation) should be comprehensively considered and calculated according to the fusion weights. For example, the severity of tire wear may become a new wear quantification value that comprehensively considers the wear conditions and weights of both after fusion; in terms of environmental status feature information, the quantitative representation of the impact of snow and ice (the 40th template-related situation) and foggy weather (the 50th template-related situation) on accidents should be fused; in terms of driver behavior feature information, the quantitative representation of improper driver fatigue operation (the 40th template-related situation) and failure to detect the vehicle ahead in time (the 50th template-related situation) should be fused. For the feature representation fusion calculation corresponding to the target risk control analysis result, the fusion of risk level should calculate a new risk level according to the respective risk level and fusion weight, the probability of occurrence should calculate a comprehensive probability of occurrence according to the probability of occurrence and weight of the two, and the cost estimate should also comprehensively consider the vehicle maintenance, personnel medical care, rescue and other costs in the two data and the weight to calculate a new comprehensive cost estimate, etc. Then the fusion calculation result is used as the input data to be learned.

[0056] Based on the generated input data to be learned, the network parameters of the initialized risk control knowledge prediction network are updated to generate the target risk control knowledge prediction network. The initialized risk control knowledge prediction network may be a multi-layer neural network structure, including an input layer, a hidden layer, and an output layer. The input data to be learned is input into the network, and the network is adjusted according to the difference between the input data and the pre-set target (such as accurately predicting the accident risk level, probability of occurrence, and cost estimate, etc.). The input layer receives the fused feature representations such as the vehicle state, environmental state, and driver behavior, as well as the fused risk level, probability of occurrence, and cost estimate. This information is calculated in the hidden layer through the connection weights and activation functions between neurons, for example, by adjusting the connection weights between neurons, so that the network output is closer to the actual target result. After multiple such input and adjustment processes, the parameters in the network are gradually optimized, and finally the target risk control knowledge prediction network is generated, which can more accurately predict risk-related results based on the input vehicle network accident monitoring data.

[0057] In a possible implementation, the method of updating the network parameters of the initialized risk control knowledge prediction network using the sample Internet of Vehicles accident monitoring data and the disturbed Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence to generate the target risk control knowledge prediction network includes: The Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence is divided into multiple training groups.

[0058] For each training group, the training group is loaded into the first candidate neural network and the second candidate neural network respectively to update the network parameters to generate the first prediction learning data and the second prediction learning data. The first candidate neural network and the second candidate neural network are both generated by initializing the risk control knowledge prediction network to configure the network parameters.

[0059] The Internet of Vehicles accident monitoring data having different risk control knowledge prediction results in the first predictive learning data and the second predictive learning data is obtained as the Internet of Vehicles accident monitoring data to be optimized.

[0060] Based on the error results of the to-be-optimized Internet of Vehicles accident monitoring data in the first predictive learning data, a set proportion of the to-be-optimized Internet of Vehicles accident monitoring data is selected as the first batch, and based on the error results of the to-be-optimized Internet of Vehicles accident monitoring data in the second predictive learning data, a set proportion of the to-be-optimized Internet of Vehicles accident monitoring data is selected as the second batch.

[0061] The second candidate neural network is fine-tuned based on the first batch, and the first candidate neural network is fine-tuned based on the second batch.

[0062] The first candidate neural network and the second candidate neural network are output as the target risk control knowledge prediction network.

[0063] In this embodiment, taking the example related to the vehicle network accident monitoring data mentioned above, it is assumed that the sample vehicle network accident monitoring data sequence contains sample data generated by many previous template vehicle network accident monitoring data, for example, there are 100 sample vehicle network accident monitoring data. These 100 data are divided into 5 training groups, each training group contains 20 sample vehicle network accident monitoring data. The basis for division can be based on the range of accident risk levels, for example, the first training group contains 20 sample data with a risk level of low to low, the second training group contains 20 sample data with a risk level of medium to medium-high, the third training group contains 20 sample data with a risk level of high to extremely high and the accident involves vehicle mechanical failure, the fourth training group contains 20 sample data with a risk level of high to extremely high and the accident involves environmental factors with a greater impact, and the fifth training group contains 20 sample data with a risk level of high to extremely high and the accident involves driver behavior factors that dominate.

[0064] Next, taking the first training group as an example, the 20 sample vehicle network accident monitoring data are loaded into the first candidate neural network and the second candidate neural network respectively. The two candidate neural networks are generated by initializing the risk control knowledge prediction network to configure the network parameters. For example, they have the same network structure, such as including an input layer, several hidden layers and an output layer, and have the same connection weights and bias parameters when initialized. When the sample data in the first training group is loaded into the first candidate neural network, the network learns according to the vehicle state feature information in the sample data (such as the wear of the brake system and the degree of tire wear in the vehicle mechanical condition feature information, the speed change and steering angle in the vehicle control related feature information), the environmental state feature information (such as the weather conditions in the natural environment feature information, the road type and curvature in the road environment feature information), and the driver behavior feature information (such as whether the driving habit feature information changes lanes frequently, the fatigue degree in the driver state feature information, etc.), combined with the corresponding target risk control analysis results (such as risk level, probability of occurrence, estimated cost, etc.), adjust the connection weights and bias parameters in the network, and generate the first prediction learning data. Similarly, the sample data in the first training group is loaded into the second candidate neural network for similar operations to generate second prediction learning data.

[0065] Then, in the first prediction learning data, the risk level predicted for a certain sample Internet of Vehicles accident monitoring data is medium, while in the second prediction learning data, the risk level predicted for the sample data is high. Such sample data is determined as the Internet of Vehicles accident monitoring data to be optimized. By comparing each sample data in the two prediction learning data one by one, all sample Internet of Vehicles accident monitoring data with different risk control knowledge prediction results are found.

[0066] Afterwards, assuming that the percentage is set to 30%, there are 20 Internet of Vehicles accident monitoring data to be optimized. Calculate the error result of each Internet of Vehicles accident monitoring data to be optimized in the first prediction learning data. For example, the error can be measured by the difference between the predicted risk level and the actual risk level (based on the target risk control analysis result). The larger the difference, the greater the error. Sort the 20 Internet of Vehicles accident monitoring data to be optimized according to the error size, and select the top 6 (20*30%=6) sample data as the first batch. Similarly, calculate the error result of each Internet of Vehicles accident monitoring data to be optimized in the second prediction learning data, and select 6 sample data as the second batch in the same way.

[0067] After that, the first batch of 6 sample data are reloaded into the second candidate neural network. According to the actual characteristics of these sample data (feature information such as vehicle status, environmental status, driver behavior, etc.) and the target risk control analysis results (accurate risk level, probability of occurrence, estimated cost, etc.), the connection weights and bias parameters in the second candidate neural network are fine-tuned so that the network can get closer to accurate prediction results when processing these sample data. Similarly, the second batch of 6 sample data are loaded into the first candidate neural network for similar fine-tuning operations.

[0068] Finally, after the above-mentioned fine-tuning operations, the first candidate neural network and the second candidate neural network are optimized to a certain extent. These two optimized neural networks are used as the final target risk control knowledge prediction network. This target risk control knowledge prediction network can combine the advantages of the two networks and more accurately predict the risk control knowledge labels for the accident monitoring data of the Internet of Vehicles.

[0069] In a possible implementation, the method of updating the network parameters of the initialized risk control knowledge prediction network using the sample Internet of Vehicles accident monitoring data and the disturbed Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence to generate the target risk control knowledge prediction network includes: Based on a plurality of predefined network training objectives, the sample Internet of Vehicles accident monitoring data and the disturbed Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence are reorganized to generate a plurality of target sample data corresponding to each network training objective.

[0070] For each network training target, sample data is selected from the corresponding target sample data as a customized training group.

[0071] Based on the customized training grouping, risk control knowledge label prediction of the corresponding network training target is performed in the initialized risk control knowledge prediction network to generate prediction learning data for each network training target.

[0072] The initialized risk control knowledge prediction network is trained based on the prediction learning data of each network training target to generate the target risk control knowledge prediction network.

[0073] In this embodiment, the predefined network training objectives may include improving the accuracy of risk level prediction for high-risk accidents, improving the accuracy of prediction of accident probability under different environmental factors, improving the accuracy of prediction of estimated costs of accidents caused by driver behavior factors, etc. Taking improving the accuracy of risk level prediction for high-risk accidents as an example, all data with high to extremely high risk levels are screened out from the sample data in the sample vehicle network accident monitoring data sequence and the disturbed vehicle network accident monitoring data. Then, the data are recombined according to the vehicle state feature information, environmental state feature information and driver behavior feature information in these data. For example, all sample data related to vehicle brake system failure and with high to extremely high risk levels and relevant parts in the disturbed vehicle network accident monitoring data are combined together to form a new target sample data subset, which focuses more on the situation of high-risk accidents caused by vehicle brake system failure; all sample data with high to extremely high risk levels in foggy weather environment and relevant parts in the disturbed vehicle network accident monitoring data are combined together to form another target sample data subset, which focuses more on high-risk accident situations in foggy weather environment. A similar operation is performed on each network training target to generate multiple target sample data corresponding to each network training target.

[0074] Next, for the network training goal of improving the accuracy of risk level prediction for high-risk accidents, a certain number of sample data are selected from each previously generated target sample data subset as customized training groups. For example, 30 sample data are selected from the target sample data subset related to vehicle brake system failure, 20 sample data are selected from the target sample data subset under foggy weather environment, etc., and these selected sample data are combined together to form a customized training group.

[0075] Then, the customized training group is loaded into the initialized risk control knowledge prediction network, and the risk control knowledge label prediction is performed according to the network training goal. For example, for the goal of improving the accuracy of risk level prediction for high-risk accidents, the network predicts the risk level of each sample data based on the vehicle state feature information, environmental state feature information, and driver behavior feature information of the sample data in the customized training group, and compares it with the actual risk level (the risk level in the target risk control analysis result), adjusts the network parameters according to the comparison results, and generates prediction learning data for this network training goal. This prediction learning data contains the learning status of the network during this training process, such as prediction accuracy, error, and other information.

[0076] Finally, the prediction learning data of each network training target are combined to conduct comprehensive training on the initialized risk control knowledge prediction network. For example, the error information in the prediction learning data that improves the accuracy of risk level prediction is used to adjust the connection weights and bias parameters related to risk level prediction in the network; the error information in the prediction learning data that improves the accuracy of occurrence probability prediction is used to adjust the connection weights and bias parameters related to occurrence probability prediction in the network; the error information in the prediction learning data that improves the accuracy of expected cost prediction is used to adjust the connection weights and bias parameters related to expected cost prediction in the network. Through such comprehensive adjustment and training, the target risk control knowledge prediction network is finally generated, which can more accurately predict risk control knowledge labels in many aspects.

[0077] In a possible implementation, the method further includes: Step S150, obtaining any input target Internet of Vehicles accident monitoring data and multiple risk control knowledge tags, wherein the target Internet of Vehicles accident monitoring data includes an accident risk control description tag and an accident feature logic chain.

[0078] In this embodiment, it is assumed that there is a vehicle networking service platform that receives various types of data uploaded from different vehicle terminals. At this time, a new target vehicle networking accident monitoring data is acquired to the platform. For example, this target vehicle networking accident monitoring data comes from a car that had an accident on a specific road section. The accident monitoring data contains rich information, in which the accident risk control description tag records in detail some preliminary risk assessment information when the accident occurred, such as the general cause of the accident is speculated to be that when the vehicle passed through a road section with accumulated water (this belongs to the natural environment characteristic information in the environmental state characteristic information), the vehicle's electronic stability system (this is the vehicle mechanical condition characteristic information in the vehicle state characteristic information) seemed to have a short failure, and the driver (driver state characteristic information in the driver behavior characteristic information) may have made an operational error due to being frightened by the sudden situation. These factors combined caused the vehicle to skid and collide with roadside facilities. The risk level was initially estimated to be high, and the possible loss range was estimated, including the vehicle repair cost of about 5,000 yuan, and if there were injuries, the medical expenses were estimated to be 2,000 yuan.

[0079] The accident characteristic logic chain constructs various logical relationships in the process of the accident in more detail. In terms of vehicle state characteristic information, it can be seen from the car's sensor data that the vehicle's speed before the accident was 60 kilometers per hour. When approaching the flooded section, the vehicle's brake system pressure fluctuated slightly, and the working parameters of the electronic stability system were abnormal, which may indicate that the electronic stability system has a hidden fault. Although the tire tread depth of the vehicle is within the normal range, the friction between the tire and the ground is reduced due to the influence of the flooded road surface. In terms of environmental state characteristic information, the depth of water in the flooded section was measured to be about 5 cm, the slope of the section was 2% downhill, and the surrounding visibility was good. In terms of driver behavior characteristic information, the driver has a bad driving habit of occasionally checking the mobile phone during normal driving (this is the driving habit characteristic information). When encountering a flooded section and the vehicle is abnormal, the driver is overly nervous at first, and then mistakenly slams on the brakes, causing the vehicle to lose control. These vehicle state characteristic information, environmental state characteristic information, and driver behavior characteristic information are interrelated, forming an accident characteristic logic chain.

[0080] At the same time, multiple risk control knowledge labels were also obtained. These risk control knowledge labels are determined based on the risk assessment system pre-set by the Internet of Vehicles service platform. For example, one of the risk control knowledge labels may be "Risks related to vehicle mechanical failure", which aims to assess the impact of the failure of the vehicle's own mechanical components on the occurrence of the accident; another risk control knowledge label may be "Driver emergency response risk", which focuses on the impact of the driver's reaction ability and correctness of operation when encountering emergencies on the accident; and another risk control knowledge label is "Risks caused by environmental factors", which focuses on analyzing the role of environmental factors such as water accumulation, slope, visibility, etc. in accidents.

[0081] Step S160, aggregating the multiple risk control knowledge tags and the accident feature logic chain of the target Internet of Vehicles accident monitoring data to generate candidate Internet of Vehicles accident monitoring data.

[0082] Step S170: Use the target risk control knowledge prediction network to predict the risk control knowledge labels of the candidate Internet of Vehicles accident monitoring data to generate risk control knowledge label prediction results.

[0083] In this embodiment, the target risk control knowledge prediction network is obtained through a large amount of data training, and can analyze the input candidate vehicle network accident monitoring data and predict the relevant risk control knowledge labels. The generated candidate vehicle network accident monitoring data is input into the target risk control knowledge prediction network, and the input layer of the network receives various information in the candidate vehicle network accident monitoring data, including vehicle state feature information (such as quantified information such as electronic stability system failure hidden dangers, brake system pressure fluctuations, etc.), environmental state feature information (such as quantified information such as water depth, road slope, etc.), driver behavior feature information (such as quantified information such as driver's wrong operation, bad habits, etc.), and the quantified risk value converted from each risk control knowledge label and the added setting label and other information. This information is calculated in the hidden layer of the network, and the neurons in the hidden layer process the input information according to the pre-learned weights and activation functions. For example, for information on risks related to vehicle mechanical failure, the neurons in the hidden layer may calculate the quantitative risk values ​​of the input electronic stability system and braking system based on the previously learned relationship weights between different vehicle failures and accident results; for information on driver emergency response risks, the neurons calculate the relationship weights based on the relationship between different driver operations and habits and accident results; for information on risks caused by environmental factors, the neurons calculate the relationship weights based on different environmental factors and accident results. Finally, the output layer generates the risk control knowledge label prediction results. This prediction result may be a re-evaluation of the risk level of each risk control knowledge label, such as predicting that the quantitative value of "risk related to vehicle mechanical failure" becomes 0.7 (increased from 0.6 after the previous conversion), the quantitative value of "driver emergency response risk" becomes 0.9 (increased from 0.8 before), and the quantitative value of "risk caused by environmental factors" becomes 0.6 (increased from 0.5 before), and some new risk labels and their quantitative values ​​may also be predicted, such as the quantitative value of "secondary damage risk of vehicle after accident" is 0.3, etc.

[0084] Step S180: If the result dimension of the risk control knowledge label prediction result is greater than the number of labels of the multiple risk control knowledge labels, the risk control knowledge label prediction result is optimized based on the number of labels to generate a label index.

[0085] If the result dimension of the risk control knowledge label prediction result is greater than the number of labels of multiple risk control knowledge labels, for example, there are 3 risk control knowledge labels initially input, and 5 risk control knowledge labels appear in the prediction result, it is necessary to optimize the risk control knowledge label prediction result based on the number of labels and generate a label index. Since there are only 3 risk control knowledge labels initially, only focus on the situation of these 3 labels in the prediction result. According to a preset optimization rule, for example, the quantitative values ​​of these 3 labels in the prediction result are sorted from high to low according to the risk level. Assume that after sorting, the quantitative value of "driver emergency response risk" is 0.9, the highest, the quantitative value of "vehicle mechanical failure related risk" is 0.7, the second, and the quantitative value of "risk caused by environmental factors" is 0.6, the lowest. Then generate a label index based on this sorting, for example, index "driver emergency response risk" as 1, "vehicle mechanical failure related risk" as 2, and "risk caused by environmental factors" as 3.

[0086] Step S190, based on the arrangement results of the multiple risk control knowledge labels in the candidate Internet of Vehicles accident monitoring data, obtain the risk control knowledge label corresponding to the label index as the prediction result of the target Internet of Vehicles accident monitoring data, wherein the prediction result is one of the multiple risk control knowledge labels.

[0087] Since the risk control knowledge labels were previously arranged in a certain order in the candidate IoV accident monitoring data, for example, "risks related to vehicle mechanical failure" ranked first, "driver emergency response risk" ranked second, and "risks caused by environmental factors" ranked third. According to the label index generated earlier, index 1 corresponds to "driver emergency response risk", so "driver emergency response risk" is used as the prediction result of the target IoV accident monitoring data.

[0088] Step S200: updating the insurance service product of the corresponding target user based on the prediction result of the target Internet of Vehicles accident monitoring data.

[0089] Assume that the target user is the owner of the car that had the accident, and he purchased the auto insurance product of the insurance company that cooperates with the Internet of Vehicles service platform. Since the prediction result is "driver emergency response risk", the insurance company updates the owner's insurance service products based on this result. If the insurance policy of the insurance company stipulates that in the case of a high risk of driver emergency response, it is necessary to increase the insurance premium or adjust some liability scopes in the insurance terms. For example, the insurance premium rate is increased by 10%, and it is clearly stipulated in the insurance terms that if a similar accident caused by the driver's emergency response occurs in the future, the car owner needs to bear a higher proportion of the deductible. In this way, insurance companies can more accurately assess risks based on the analysis results of the Internet of Vehicles accident monitoring data, so as to reasonably adjust insurance service products, which not only guarantees their own risk control, but also provides car owners with insurance services that are more in line with the actual risk situation.

[0090] In a possible implementation, step S160 includes: Step S161: Based on the knowledge representation template corresponding to each risk control knowledge tag, knowledge conversion is performed on the multiple risk control knowledge tags to generate multiple tag input representations.

[0091] Step S162, aggregating the multiple label input representations and the accident feature logic chain of the target Internet of Vehicles accident monitoring data to generate a label aggregation result.

[0092] Step S163: If the number of the risk control knowledge labels is less than the global number of risk control knowledge labels of the target risk control knowledge prediction network, then based on the difference between the global number of risk control knowledge labels and the number of risk control knowledge labels, a set label is added to the risk control description label combination data to generate candidate Internet of Vehicles accident monitoring data.

[0093] In this embodiment, taking the risk control knowledge label "vehicle mechanical failure related risk" as an example, its corresponding knowledge representation template may stipulate that specific data should be extracted from the vehicle status feature information for quantitative representation. For example, for the hidden danger of electronic stability system failure, according to the template, if the abnormal working parameters of the electronic stability system last for more than 3 seconds (this is a threshold set in the template), the risk under this label is quantified as 0.6 (assuming that the value between 0-1 represents the risk level, 1 is the highest risk); for the pressure fluctuation of the brake system, if the fluctuation amplitude exceeds 20% of the normal range, it is quantified as 0.3. For the label "driver emergency response risk", according to its knowledge representation template, the risk of the driver's wrong operation (such as slamming on the brakes) in an emergency situation is quantified as 0.8, because this is a relatively serious wrong operation; and the driver has a bad habit of checking the mobile phone, which is quantified as 0.2 according to the template, indicating that this habit has a certain potential impact on the occurrence of accidents. For the label "risk caused by environmental factors", the depth of water accumulation is 5 cm, which is quantified as 0.5 according to the template, and the 2% downhill section is quantified as 0.1 (indicating that the impact on the accident is relatively small). In this way, multiple label input representations are generated.

[0094] Then, multiple label input representations are aggregated with the accident feature logic chain of the target vehicle network accident monitoring data to generate label aggregation results. During the aggregation process, the quantitative data in each label input representation is associated with the specific data in the accident feature logic chain. For example, the quantitative risk data of the electronic stability system and the braking system under the label "risk related to vehicle mechanical failure" are combined with the abnormal working parameters of the electronic stability system and the pressure fluctuation of the braking system of the vehicle before the accident in the accident feature logic chain to form a comprehensive description of the logical relationship between vehicle mechanical failure and accident; the quantitative risk data under the label "driver emergency response risk" is combined with the driver's operation and habits at the time of the accident in the accident feature logic chain to form a description of the logical relationship between driver-related risks and accidents; the quantitative risk data under the label "risk caused by environmental factors" is combined with environmental information such as water depth and road slope in the accident feature logic chain to form a description of the logical relationship between environmental factors and accidents. These descriptions are integrated together to form the label aggregation results.

[0095] If the number of risk control knowledge labels is less than the global number of risk control knowledge labels of the target risk control knowledge prediction network, assuming that the global number of risk control knowledge labels of the target risk control knowledge prediction network is 10, and the currently acquired risk control knowledge labels are only 3, then the difference between the two numbers is 7. Based on this number difference, set labels are added to the risk control description label combination data. These set labels may be determined according to the default risk assessment rules of the Internet of Vehicles service platform, such as adding a label such as "the potential impact of other vehicle components (such as engines, gearboxes, etc.) on accidents is not assessed", and a default quantitative risk value of 0.1 is set (indicating that the impact is temporarily considered to be small), adding a label "the impact of external interference factors (such as sudden interference from other vehicles, etc.) on accidents is not considered", and the quantitative risk value is 0.05, etc. After adding set labels in this way, candidate Internet of Vehicles accident monitoring data is generated. This candidate Internet of Vehicles accident monitoring data is a complete data that integrates the original target Internet of Vehicles accident monitoring data, the label input representation after the conversion of multiple risk control knowledge labels, and the added set labels. It can more comprehensively reflect the risk situation of the accident and meet the input requirements of the target risk control knowledge prediction network.

[0096] Figure 2 The hardware structure of the insurance service system 100 provided in the embodiment of the present invention for implementing the risk control analysis method based on the big data of vehicle network accident is shown as follows: Figure 2 As shown, the insurance service system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0097] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the insurance service system 100 uses to execute or use to complete the exemplary methods described in the present invention.

[0098] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the risk control analysis method based on the big data of vehicle network accidents in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0099] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned insurance service system 100. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0100] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned risk control analysis method based on the big data of vehicle network accidents is implemented.

[0101] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A risk control analysis method based on Internet of Vehicles accident big data, characterized in that: The method comprises: Obtaining X template Internet of Vehicles accident monitoring data from the Internet of Vehicles accident big data, where X is a positive integer; For each template IoV accident monitoring data, obtain the accident risk control description label of the current template IoV accident monitoring data as the target risk control description label, and obtain the accident risk control description labels in Y other template IoV accident monitoring data from X-1 template IoV accident monitoring data as the disturbance risk control description label, where Y is a positive integer less than Z, and Z is the number of predefined risk control knowledge categories, which is the limit number of risk control knowledge categories in the prediction result predicted by the risk control knowledge label; Based on the target risk control description label, the disturbance risk control description label and the accident feature logic chain of each template IoV accident monitoring data, sample IoV accident monitoring data is generated by aggregation, and a sample IoV accident monitoring data sequence is generated, wherein the target risk control description label is a target risk control analysis result corresponding to the sample IoV accident monitoring data, and the sample IoV accident monitoring data sequence includes multiple sample IoV accident monitoring data, and each sample IoV accident monitoring data is generated based on one template IoV accident monitoring data; Based on the sample Internet of Vehicles accident monitoring data sequence, the network parameters of the initialized risk control knowledge prediction network are updated to generate a target risk control knowledge prediction network.

2. The risk control analysis method based on Internet of Vehicles accident big data according to claim 1 is characterized in that: The method of obtaining, for each template IoV accident monitoring data, the accident risk control description label of the current template IoV accident monitoring data as the target risk control description label and obtaining, from X-1 template IoV accident monitoring data, the accident risk control description labels of Y other template IoV accident monitoring data as the disturbance risk control description label includes: Obtaining any one of the X template Internet of Vehicles accident monitoring data as the current template Internet of Vehicles accident monitoring data; In the quantity interval where the quantity of risk control knowledge labels predicted by the initialized risk control knowledge prediction network is the limit quantity, a quantity is randomly determined as the disturbance label quantity Y; Based on the disturbance label quantity Y, the template Internet of Vehicles accident monitoring data except Y template Internet of Vehicles accident monitoring data are obtained from the X-1 template Internet of Vehicles accident monitoring data as other template Internet of Vehicles accident monitoring data.

3. The risk control analysis method based on Internet of Vehicles accident big data according to claim 2 is characterized in that: The target risk control description label, the disturbance risk control description label and the accident feature logic chain based on each template IoV accident monitoring data are aggregated to generate sample IoV accident monitoring data, including: Randomly combine and connect the target risk control description tag and the disturbance risk control description tag to generate risk control description tag combination data; If the number of templates of the target risk control description label and the disturbance risk control description label is less than the global number of risk control knowledge labels of the target risk control knowledge prediction network, then based on the difference between the global number of risk control knowledge labels and the number of templates, adding a setting label to the risk control description label combination data; Aggregating the risk control description tag combination data with the accident feature logic chain of the current template Internet of Vehicles accident monitoring data to generate sample Internet of Vehicles accident monitoring data; The steps of generating the accident feature logic chain of the current template vehicle networking accident monitoring data include: Extracting accident feature data from the current template Internet of Vehicles accident monitoring data, the accident feature data including vehicle state feature information, environmental state feature information and driver behavior feature information, the vehicle state feature information including vehicle mechanical condition feature information and vehicle control-related feature information, the environmental state feature information including natural environment feature information and road environment feature information, and the driver behavior feature information including driving habit feature information and driver state feature information; Classifying the accident feature data to generate trigger features that directly cause the accident and associated features that affect the development of the accident; Analyze the sequence of each trigger feature in the accident process, and build a causal chain between the trigger features according to the sequence. The construction of the causal chain is based on the logical order of the event development. The trigger feature that occurs first is taken as the cause, and the trigger feature that occurs later is taken as the effect. They are connected in sequence to generate a directed causal chain. Analyze the influence patterns of the associated features on the trigger features one by one, and connect the associated features to the corresponding positions of the causal relationship chain of the trigger features according to the influence patterns, so as to construct an association relationship diagram between the associated features and the trigger features; The mutual influence between the associated features is analyzed, and for the associated features with mutual influence, a two-way influence relationship chain is established. The causal relationship chain between the trigger features, the association relationship diagram between the associated features and the trigger features, and the influence relationship chain between the associated features are integrated to generate an integrated accident feature logic chain.

4. The risk control analysis method based on Internet of Vehicles accident big data according to claim 1 is characterized in that: Before updating the network parameters of the initialized risk control knowledge prediction network based on the sample IoV accident monitoring data sequence to generate the target risk control knowledge prediction network, the method further includes: Aggregating the target risk control description label, the disturbance risk control description label and the accident feature logic chain of the current template Internet of Vehicles accident monitoring data, generating disturbance Internet of Vehicles accident monitoring data and adding it to the sample Internet of Vehicles accident monitoring data sequence, wherein the disturbance risk control description label is the target risk control analysis result corresponding to the disturbance Internet of Vehicles accident monitoring data; The updating of network parameters of the initialized risk control knowledge prediction network based on the sample Internet of Vehicles accident monitoring data sequence to generate a target risk control knowledge prediction network includes: The sample Internet of Vehicles accident monitoring data and the disturbed Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence are used to update the network parameters of the initialized risk control knowledge prediction network to generate the target risk control knowledge prediction network.

5. The risk control analysis method based on Internet of Vehicles accident big data according to claim 4 is characterized in that: The method of updating the network parameters of the initialized risk control knowledge prediction network using the sample Internet of Vehicles accident monitoring data and the disturbed Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence to generate the target risk control knowledge prediction network includes: Obtaining two vehicle networking accident monitoring data and corresponding target risk control analysis results from the sample vehicle networking accident monitoring data sequence; Based on the set fusion weight, the feature representations corresponding to the two vehicle network accident monitoring data are fused and calculated, and the feature representations corresponding to the two corresponding target risk control analysis results are fused and calculated, and the fusion calculation results are used as input data to be learned; Based on the generated input data to be learned, the network parameters of the initialized risk control knowledge prediction network are updated to generate the target risk control knowledge prediction network.

6. The risk control analysis method based on Internet of Vehicles accident big data according to claim 5 is characterized in that: The method of updating the network parameters of the initialized risk control knowledge prediction network using the sample Internet of Vehicles accident monitoring data and the disturbed Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence to generate the target risk control knowledge prediction network includes: Dividing the Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence into a plurality of training groups; For each training group, the training group is loaded into the first candidate neural network and the second candidate neural network respectively to update the network parameters, and the first prediction learning data and the second prediction learning data are generated. The first candidate neural network and the second candidate neural network are both generated by initializing the risk control knowledge prediction network to configure the network parameters; Acquire the Internet of Vehicles accident monitoring data having different risk control knowledge prediction results in the first prediction learning data and the second prediction learning data as the Internet of Vehicles accident monitoring data to be optimized; Based on the error result of the to-be-optimized Internet of Vehicles accident monitoring data in the first prediction learning data, selecting the Internet of Vehicles accident monitoring data with a set proportion from the to-be-optimized Internet of Vehicles accident monitoring data as the first batch, and based on the error result of the to-be-optimized Internet of Vehicles accident monitoring data in the second prediction learning data, selecting the to-be-optimized Internet of Vehicles accident monitoring data with the set proportion from the to-be-optimized Internet of Vehicles accident monitoring data as the second batch; Fine-tuning the second candidate neural network based on the first batch, and fine-tuning the first candidate neural network based on the second batch; The first candidate neural network and the second candidate neural network are output as the target risk control knowledge prediction network.

7. The risk control analysis method based on Internet of Vehicles accident big data according to claim 4 is characterized in that: The method of updating the network parameters of the initialized risk control knowledge prediction network using the sample Internet of Vehicles accident monitoring data and the disturbed Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence to generate the target risk control knowledge prediction network includes: Based on a plurality of predefined network training objectives, the sample Internet of Vehicles accident monitoring data and the disturbed Internet of Vehicles accident monitoring data in the sample Internet of Vehicles accident monitoring data sequence are reorganized to generate a plurality of target sample data corresponding to each network training objective; For each network training target, sample data is selected from the corresponding target sample data as a customized training group; Based on the customized training group, the risk control knowledge label prediction of the corresponding network training target is performed in the initialized risk control knowledge prediction network to generate prediction learning data for each network training target; The initialized risk control knowledge prediction network is trained based on the prediction learning data of each network training target to generate the target risk control knowledge prediction network.

8. The risk control analysis method based on Internet of Vehicles accident big data according to claim 1 is characterized in that: The method further comprises: Obtaining any input target Internet of Vehicles accident monitoring data and multiple risk control knowledge tags, wherein the target Internet of Vehicles accident monitoring data includes an accident risk control description tag and an accident feature logic chain; Aggregating the multiple risk control knowledge tags and the accident feature logic chain of the target Internet of Vehicles accident monitoring data to generate candidate Internet of Vehicles accident monitoring data; Using the target risk control knowledge prediction network to predict the risk control knowledge labels of the candidate Internet of Vehicles accident monitoring data, and generating risk control knowledge label prediction results; If the result dimension of the risk control knowledge label prediction result is greater than the number of labels of the multiple risk control knowledge labels, optimizing the risk control knowledge label prediction result based on the number of labels to generate a label index; Based on the arrangement results of the multiple risk control knowledge labels in the candidate Internet of Vehicles accident monitoring data, obtaining the risk control knowledge label corresponding to the label index as the prediction result of the target Internet of Vehicles accident monitoring data, wherein the prediction result is one of the multiple risk control knowledge labels; Based on the prediction results of the target Internet of Vehicles accident monitoring data, the insurance service products for the corresponding target users are updated.

9. The risk control analysis method based on Internet of Vehicles accident big data according to claim 8 is characterized in that: The step of aggregating the accident feature logic chain based on the multiple risk control knowledge tags and the target IoV accident monitoring data to generate candidate IoV accident monitoring data includes: Based on the knowledge representation template corresponding to each risk control knowledge label, the multiple risk control knowledge labels are converted to generate multiple label input representations; Aggregating the multiple label input representations with the accident feature logic chain of the target Internet of Vehicles accident monitoring data to generate a label aggregation result; If the number of the risk control knowledge labels is less than the global number of risk control knowledge labels of the target risk control knowledge prediction network, then based on the difference between the global number of risk control knowledge labels and the number of risk control knowledge labels, a set label is added to the risk control description label combination data to generate candidate Internet of Vehicles accident monitoring data.

10. An insurance service system, characterized in that: The insurance service system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the risk control analysis method based on vehicle network accident big data as described in any one of claims 1 to 9 above.

Images