Highway driving risk prediction method and system based on control unit

Abstract

The present invention provides a method and system for predicting driving risks on highways based on a control unit. First, the method and system can automatically extract key risk areas, i.e., X driving risk road condition data, from target driving road condition data, and then generate detailed risk knowledge point estimation data by performing detailed risk knowledge point estimation on these risk areas. By adopting a target driving risk estimation network with optimized parameters, the present invention significantly improves the accuracy and reliability of driving risk prediction. During the training process, the network parameters are continuously optimized by comparing the prior risk knowledge points of the sample driving road condition data with the estimated risk knowledge points until the training termination requirements are met, thereby ensuring the prediction performance of the network model. Therefore, the present invention can not only effectively identify potential driving risks on highways, but also provide scientific and reasonable control plan suggestions for traffic management departments, thereby improving driving safety on highways and reducing the occurrence of traffic accidents.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for predicting highway driving risks based on a control unit. Background Art

[0002] With the rapid development of highways and the continuous increase in the number of vehicles, the issue of highway driving safety has become increasingly prominent. Traditional driving risk management mainly relies on manual patrols, surveillance cameras, and data analysis after traffic accidents. These methods have problems such as delayed response, limited coverage, and inability to predict risks in real time. Therefore, how to achieve real-time and accurate prediction of highway driving risks has become a technical problem that needs to be solved urgently.

[0003] In the existing technology, although some methods based on data analysis and model prediction have been applied to driving risk identification, these methods often rely on simple statistical models or rule judgments, and it is difficult to fully and accurately capture the correlation between complex driving conditions and risks. In particular, when faced with changing road conditions, different traffic flows, and differences in driver behavior, the prediction accuracy of traditional methods will drop significantly.

[0004] In addition, existing driving risk prediction systems usually lack the ability to make detailed predictions for specific control units (such as a section or intersection of a highway). Since different sections and intersections of highways have unique traffic characteristics and risk patterns, it is necessary to conduct targeted risk prediction and control plan formulation for these specific areas. However, traditional methods often fail to provide sufficiently detailed prediction results, resulting in a lack of targetedness and effectiveness of control measures. 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 method for predicting highway driving risks based on a control unit, the method comprising:

[0006] Acquire target traffic condition data, which is traffic condition data to be analyzed obtained by real-time traffic condition monitoring of a target control unit of a target highway;

[0007] The target driving road condition data is loaded into a target driving risk estimation network, and a target risk driving area positioning unit in the target driving risk estimation network is used to extract X driving risk road condition data from the target driving road condition data;

[0008] Using a target risk knowledge point estimation unit in the target driving risk estimation network to estimate the X driving risk road condition data, and generate risk knowledge point estimation data of the target driving road condition data, wherein X is not less than 1, and the X driving risk road condition data are used to represent regional road condition data in the target driving road condition data whose estimated risk confidence is not less than a threshold confidence;

[0009] Making a control plan decision for the control node corresponding to the target control unit based on the risk knowledge point estimation data of the target driving road condition data;

[0010] Among them, the target driving risk estimation network is a deep learning network generated by using sample driving road condition data and sample area road condition data to optimize the parameters of the initialized driving risk estimation network until the following training termination requirements are met: the prior risk knowledge point data of the sample driving road condition data and the estimated risk knowledge point data of the target driving risk road condition data meet the first training error requirement, and the target driving risk road condition data is the driving risk road condition data with the largest estimated risk confidence in the sample driving road condition data output by the initialized risk driving area positioning unit in the initialized driving risk estimation network.

[0011] On the other hand, an embodiment of the present invention also provides a highway driving risk prediction system based on a control unit, 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.

[0012] Based on the above aspects, the embodiment of the present application obtains the driving road condition data of the target highway in real time, and uses the target driving risk estimation network of deep learning to perform accurate risk prediction and analysis. The method can first automatically extract key risk areas from complex driving road condition data, that is, X driving risk road condition data, which represent road condition information with high risk confidence. Subsequently, by performing detailed risk knowledge point estimation on these risk areas, detailed risk knowledge point estimation data is generated, which provides a data basis for subsequent control plan decision-making. By adopting the target driving risk estimation network with optimized parameters, the present invention significantly improves the accuracy and reliability of driving risk prediction. During the training process, by comparing the prior risk knowledge points of the sample driving road condition data with the estimated risk knowledge points, the network parameters are continuously optimized until the strict training termination requirements are met, thereby ensuring the prediction performance of the network model. This risk prediction method based on deep learning can not only effectively identify potential highway driving risks, but also provide scientific and reasonable control plan suggestions for traffic management departments, thereby improving highway driving safety and reducing the occurrence of traffic accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the execution flow of the highway driving risk prediction method based on the control unit provided in an embodiment of the present invention.

[0014] Figure 2 It is a schematic diagram of the hardware architecture of a highway driving risk prediction system based on a control unit provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of a highway driving risk prediction method based on a control unit provided by an embodiment of the present invention. The highway driving risk prediction method based on a control unit is introduced in detail below.

[0016] Step S110, obtaining target traffic condition data, wherein the target traffic condition data is traffic condition data to be analyzed obtained by real-time traffic condition monitoring of a target control unit of a target highway.

[0017] In this embodiment, for a highway traffic monitoring system in a certain area, the target highway is a highway with a large traffic volume and complex terrain in the area, for example, the highway passes through a mountainous area and has multiple tunnels and viaducts. The target control unit can be a specific section of the highway, such as the section from a tunnel entrance to the next service area.

[0018] Therefore, the server obtains driving road condition data through multiple sensors installed in this target control unit. These sensors include cameras, speed sensors, meteorological sensors, etc. The camera is used to capture the driving status of vehicles on the road, such as the density of vehicles, whether there are vehicles changing lanes illegally, etc. The speed sensor can detect the driving speed of the vehicle, which helps to analyze whether the vehicle is speeding or driving too slowly (perhaps due to an emergency). The meteorological sensor is responsible for collecting weather information, such as whether there is rain, snow, fog and other weather conditions, because different weather conditions will have a great impact on driving safety.

[0019] For example, at a certain moment, the camera captured a high density of vehicles on the road, some lanes had slow-moving vehicles, the speed sensor showed that the speed of some vehicles was much lower than the normal speed limit, and the weather sensor detected that there was mist on the road. The server integrates the information collected from various sensors to form target traffic data. This target traffic data contains a lot of complex information, such as the location coordinates, speed, driving direction, and current weather conditions of different vehicles. These target traffic data are all traffic data to be analyzed in preparation for subsequent risk assessment.

[0020] Step S120: loading the target driving road condition data into a target driving risk estimation network, and extracting X driving risk road condition data from the target driving road condition data using a target risk driving area positioning unit in the target driving risk estimation network.

[0021] In this embodiment, the previously acquired target driving road condition data can be loaded into the target driving risk estimation network. This target driving risk estimation network is a deep learning network that has been trained and optimized with a large amount of data. The target risk driving area positioning unit starts to process the target driving road condition data.

[0022] Assume that the target driving road condition data contains information about a road section with a length of 10 kilometers, a total of 500 vehicles, and different weather conditions and road conditions. The target risk driving area positioning unit will first split the target driving road condition data into Z regional road condition data according to certain rules. For example, it can be split into one area per 1 kilometer, so that Z is equal to 10. Then, the target risk driving area positioning unit will evaluate the risk confidence of each regional road condition data.

[0023] For each area, factors such as vehicle density, speed distribution, weather conditions, and road structure are comprehensively considered. For example, in one area, the vehicle density is too high, the average vehicle speed is less than 30% of the normal speed, and due to the misty weather, visibility is low, so the risk confidence of this area will be relatively high. Assuming that the threshold confidence is set to 0.6, after evaluation, it is found that the risk confidence of 3 areas is not less than this threshold confidence, then X is equal to 3. The road condition data of these 3 areas are extracted as X driving risk road condition data. These data contain detailed vehicle driving information, weather conditions, road conditions and other information, which are important basis for subsequent risk knowledge point estimation.

[0024] Step S130, using the target risk knowledge point estimation unit in the target driving risk estimation network to estimate the X driving risk road condition data, and generate risk knowledge point estimation data of the target driving road condition data, wherein X is not less than 1, and the X driving risk road condition data are used to represent the regional road condition data in which the estimated risk confidence is not less than the threshold confidence in the target driving road condition data.

[0025] In this embodiment, the target risk knowledge point estimation unit starts to process the previously extracted X driving risk road condition data. First, the target risk driving area positioning unit in the target driving risk estimation network encodes and represents the X driving risk road condition data to generate X graph encoding vectors.

[0026] For example, for the road condition data of the three risk areas extracted previously (X=3), the target risk driving area positioning unit will encode various information in the road condition data of each area. For vehicle information, it may be encoded based on factors such as the type of vehicle (such as cars, trucks, buses, etc.), the speed range of the vehicle, and the distance between vehicles. For weather information, it will be encoded based on factors such as the concentration of fog and whether there is rainfall. For road information, factors such as the slope of the road and the curvature of the curve will be considered for encoding. In this way, the road condition data of each area is converted into a graph encoding vector.

[0027] Then, the three graph encoding vectors are averaged and loaded into the fully connected mapping branch in the target risk knowledge point estimation unit. The fully connected mapping branch predicts this average vector based on the model parameters obtained from the previous training. This risk knowledge point is a complex logical description label, which may contain multiple aspects.

[0028] For example, the risk knowledge point estimation data may be described as follows: "In area 1, due to the high vehicle density (more than 100 vehicles per kilometer) and the average vehicle speed is 30% lower than the normal speed in foggy weather (visibility less than 500 meters), there is a risk of chain rear-end collisions and the risk level is high. In area 2, although the vehicle density is moderate, there is a risk of vehicle loss of control due to the large slope of the road and a small number of vehicles speeding (exceeding the speed limit by 20%). The risk level is medium. In area 3, the vehicle density is low, but there is ice in the fog, and vehicles are prone to skidding. There is a risk of skidding and the risk level is medium." This risk knowledge point estimation data comprehensively describes the complex information such as the status of each risk area, risk type, and risk level.

[0029] Step S140: making a control plan decision for the control node corresponding to the target control unit based on the risk knowledge point estimation data of the target driving road condition data.

[0030] Among them, the target driving risk estimation network is a deep learning network generated by using sample driving road condition data and sample area road condition data to optimize the parameters of the initialized driving risk estimation network until the following training termination requirements are met: the prior risk knowledge point data of the sample driving road condition data and the estimated risk knowledge point data of the target driving risk road condition data meet the first training error requirement, and the target driving risk road condition data is the driving risk road condition data with the largest estimated risk confidence in the sample driving road condition data output by the initialized risk driving area positioning unit in the initialized driving risk estimation network.

[0031] In this embodiment, the server makes a control plan decision for the control node corresponding to the target control unit based on the risk knowledge point estimation data generated for the target driving road condition data.

[0032] Continuing with the above example, for the high-risk situation of a chain rear-end collision in area 1, the server can send instructions to the traffic signs in the area through the control node to reduce the speed limit of the road section, and at the same time remind the driver to pay attention to the distance between vehicles and drive slowly on the electronic display screen. For the medium-risk situation of vehicle loss of control in area 2, the server can notify the control node to arrange road patrol personnel to focus on this area and, if necessary, warn speeding vehicles. For the medium-risk situation of skidding in area 3, the control node can contact the relevant departments to carry out salt and de-icing operations in the area (if conditions permit) based on the decision of the server, and set up warning signs at the entrance of the road to remind drivers to pay attention to the icy conditions of the road surface.

[0033] In addition, if the overall risk level is high, the server can also adjust the traffic flow of the target control unit through the control node, such as limiting the number of vehicles entering the road section, or directing vehicles to other sections. In this way, by making decisions on the control node based on the risk knowledge point estimation data, the driving safety of the target highway control unit can be effectively improved and the probability of traffic accidents can be reduced.

[0034] At the same time, in order to ensure the accuracy and effectiveness of decision-making, the server will continue to monitor changes in target traffic data and adjust the control plan based on the new data. For example, if the fog dissipates and the risk caused by fog is reduced, the server will adjust the traffic instructions accordingly through the control node and restore the normal traffic management mode.

[0035] In this way, the target driving road condition data can be effectively utilized to conduct risk assessment through the target driving risk estimation network, and reasonable control plan decisions can be made for the control nodes based on the assessment results, thereby ensuring driving safety and smooth traffic in the target control unit of the target highway.

[0036] Based on the above steps, the embodiment of the present application obtains the driving road condition data of the target highway in real time, and uses the target driving risk estimation network of deep learning to perform accurate risk prediction and analysis. The method can first automatically extract key risk areas from complex driving road condition data, that is, X driving risk road condition data, which represent road condition information with high risk confidence. Subsequently, by performing detailed risk knowledge point estimation on these risk areas, detailed risk knowledge point estimation data is generated, which provides a data basis for subsequent control plan decision-making. By adopting the target driving risk estimation network with optimized parameters, the present invention significantly improves the accuracy and reliability of driving risk prediction. During the training process, by comparing the prior risk knowledge points of the sample driving road condition data with the estimated risk knowledge points, the network parameters are continuously optimized until the strict training termination requirements are met, thereby ensuring the prediction performance of the network model. This risk prediction method based on deep learning can not only effectively identify potential highway driving risks, but also provide scientific and reasonable control plan suggestions for traffic management departments, thereby improving highway driving safety and reducing the occurrence of traffic accidents.

[0037] In a possible implementation, the training termination requirement also includes: the prior risk knowledge point data of the sample driving road condition data and the risk knowledge point estimation data of the sample driving road condition data output by the initialized risk knowledge point estimation unit in the initialized driving risk estimation network meet the second training error requirement, and the prior risk knowledge point data of the sample area road condition data and the estimated risk knowledge point data of the sample area road condition data output by the initialized risk driving area positioning unit in the initialized driving risk estimation network meet the third training error requirement.

[0038] In this embodiment, the server first obtains a large amount of sample driving traffic data and sample regional traffic data. These data come from different highway sections and include various traffic conditions. For example, the sample driving traffic data includes vehicle driving conditions on multiple highway sections under different time periods and weather conditions, such as vehicle speed, density, vehicle model distribution, etc., and also includes corresponding prior risk knowledge point data. These prior risk knowledge point data are complex logical description labels derived by professional traffic analysts based on experience and actual traffic accidents, such as "During peak hours on a specific section of road, due to the high proportion of large vehicles and large speed differences, there is a high risk of collision due to frequent lane changes". The sample regional traffic data also includes road conditions, weather influences, and prior risk knowledge point data, such as prior risk knowledge points such as the proneness of skidding in rainy and foggy weather on a mountainous section of road.

[0039] The initialized driving risk estimation network has an initialized risk driving area positioning unit and an initialized risk knowledge point estimation unit. When the parameter optimization starts, the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit in the initialized driving risk estimation network are subjected to the ath round of combined network parameter learning based on the following operations (a is a positive integer not less than 1, and the risk driving area positioning unit and the risk knowledge point estimation unit generated by the 0th round of network parameter learning are the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit in the initialized driving risk estimation network whose parameters have not been updated).

[0040] The sample area road condition data is loaded into the risk driving area positioning unit generated by the a-1 round of network parameter learning to generate the estimated risk knowledge point data of the sample area road condition data output by the a-1 round of network parameter learning. For example, for a section of mountainous sample area road condition data containing curves and uphill and downhill slopes, the risk driving area positioning unit generated by the a-1 round of network parameter learning will output the estimated risk knowledge point data for the road condition data in this area based on the existing parameters and the vehicle speed, curve curvature, slope and other information in the data, such as "due to the large curve curvature and the vehicle speed close to the speed limit, there is a risk of skidding on a slippery road surface."

[0041] Next, the sample driving road condition data is loaded into the risk driving area positioning unit generated by the a-1th round of network parameter learning and the risk knowledge point estimation unit generated by the a-1th round of network parameter learning, and the estimated risk knowledge point data of the target driving risk road condition data output by the a-1th round of network parameter learning and the risk knowledge point estimation data of the sample driving road condition data output by the a-1th round of network parameter learning are generated. Specifically, when the sample driving road condition data is loaded into the risk driving area positioning unit generated by the a-1th round of network parameter learning, the sample driving road condition data is first disassembled into Z sample area road condition data, for example, a long section of the sample driving road condition data of a highway is disassembled into multiple areas according to fixed distances or specific traffic signs. Then, the risk driving area positioning unit generated by the a-1th round of network parameter learning is used to extract the sample area road condition data with an estimated risk confidence not less than the threshold confidence from the Z sample area road condition data, and generate Y risk sample area road condition data (Y not less than 1) output by the ath round of network parameter learning. At the same time, the sample area road condition data with the largest estimated risk confidence among the Z sample area road condition data is output as the target driving risk road condition data, and the risk knowledge point corresponding to the estimated risk confidence of the target driving risk road condition data is output as the estimated risk knowledge point data of the target driving risk road condition data. Then, the Y risk sample area road condition data are loaded into the risk knowledge point estimation unit generated by the a-1th round of network parameter learning to generate the risk knowledge point estimation data of the sample driving road condition data output by the ath round of network parameter learning.

[0042] In this process, error evaluation is also required in the a-th round of combined network parameter learning. The estimated risk confidence corresponding to the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample driving road condition data are loaded into the first training cost function to generate the first training cost parameter. For example, if the estimated risk confidence of a certain risk in the estimated risk knowledge point data of the target driving risk road condition data is 0.7, and the corresponding prior risk confidence in the prior risk knowledge point data of the sample driving road condition data is 0.8, the first training cost parameter is calculated by the first training cost function. Similarly, the estimated risk confidence corresponding to the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample driving road condition data are loaded into the second training cost function to generate second training cost parameters; the estimated risk confidence corresponding to the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample area road condition data are loaded into the third training cost function to generate third training cost parameters.

[0043] Then determine whether the first training cost parameter meets the first training error requirement, whether the second training cost parameter meets the second training error requirement, and whether the third training cost parameter meets the third training error requirement. When it is extracted that the first training cost parameter meets the first training error requirement, the second training cost parameter meets the second training error requirement, and the third training cost parameter meets the third training error requirement, determine that the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning, the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning, and the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning meet the training termination requirements. At this time, terminate the network parameter optimization, and output the risk driving area positioning unit and risk knowledge point estimation unit generated by the a-1-th round of network parameter learning as the target risk driving area positioning unit and target risk knowledge point estimation unit in the target driving risk estimation network. In this way, the construction of the target driving risk estimation network is completed, so as to facilitate the subsequent risk assessment of the target driving road condition data.

[0044] In a possible implementation manner, before loading the target driving road condition data into the target driving risk estimation network, the method further includes:

[0045] Step S101, obtaining the sample driving road condition data and the sample area road condition data.

[0046] Step S102, using the sample driving road condition data and the sample area road condition data, the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit in the initialized driving risk estimation network are combined and iteratively optimized to generate the target driving risk estimation network. During the network parameter optimization process, if the initialized driving risk estimation network does not meet the training termination requirements, the neuron weight information in the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit is updated. If the initialized driving risk estimation network meets the training termination requirements, the network parameter optimization is terminated, and the initialized driving risk estimation network when the network parameter optimization is terminated is output as the target driving risk estimation network, and the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit when the network parameter optimization is terminated are respectively output as the target risk driving area positioning unit and the target risk knowledge point estimation unit in the target driving risk estimation network.

[0047] In a possible implementation, step S102 includes:

[0048] Based on the following operations, the initialization risk driving area positioning unit and the initialization risk knowledge point estimation unit in the initialization driving risk estimation network are subjected to the ath round of combined network parameter learning, where a is a positive integer not less than 1, and the risk driving area positioning unit and the risk knowledge point estimation unit generated by the 0th round of network parameter learning are the initialization risk driving area positioning unit and the initialization risk knowledge point estimation unit in the initialization driving risk estimation network whose parameters have not been updated, including:

[0049] Step S1021, loading the sample area road condition data into the risk driving area positioning unit generated by the a-1th round of network parameter learning, generating estimated risk knowledge point data of the sample area road condition data output by the ath round of network parameter learning.

[0050] Step S1022, load the sample driving road condition data into the risk driving area positioning unit generated by the a-1th round of network parameter learning and the risk knowledge point estimation unit generated by the a-1th round of network parameter learning, generate the estimated risk knowledge point data of the target driving risk road condition data output by the a-1th round of network parameter learning, and generate the risk knowledge point estimation data of the sample driving road condition data output by the a-1th round of network parameter learning.

[0051] Step S1023, when the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning, the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning, and the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning meet the training termination requirements, the network parameter optimization is terminated, and the risk driving area positioning unit and the risk knowledge point estimation unit generated by the a-1-th round of network parameter learning are respectively output as the target risk driving area positioning unit and the target risk knowledge point estimation unit in the target driving risk estimation network.

[0052] In this embodiment, the first step is to obtain sample traffic condition data and sample regional traffic condition data. The server collects these data from a huge traffic data repository, which is widely sourced and representative. The sample traffic condition data covers the driving conditions of many highway sections under different conditions. For example, highways are selected from different geographical locations, such as mountain highways, plain highways, and coastal highways. For the sample traffic condition data of mountain highways, it contains information such as the speed change of vehicles on the climbing section, the vehicle driving trajectory at the bend, and the vehicle density in the tunnel; the plain highway is more about the vehicle overtaking situation in the long straight section, the traffic flow distribution at different time periods, etc.; the coastal highway may record the stability of the vehicle in strong wind weather, etc. At the same time, each sample traffic condition data is accompanied by corresponding prior risk knowledge point data, which are compiled by professional traffic analysts based on years of experience, a large amount of historical traffic accident data and related traffic engineering principles. For example, at a certain bend on a mountain highway, the prior risk knowledge point data may indicate that there is a high risk of vehicle rollover due to the large curvature of the bend and some drivers speeding. The risk situation description here includes information on road characteristics, vehicle behavior, risk level, etc., and is a complex logical description label.

[0053] The sample area road condition data focuses on the impact of specific area conditions on driving. Taking a section of a mountain highway as an example, the sample area road condition data may include information such as the slope of the section, the friction coefficient of the road surface, and the surrounding environment (such as whether it is close to a cliff). It also comes with prior risk knowledge point data. For example, on rainy days, due to the large slope and reduced friction coefficient of the road surface, the braking distance of the vehicle will increase significantly, which is easy to cause a rear-end collision. This is a prior risk knowledge point data for the road conditions in this area.

[0054] After obtaining these data, the server begins to use the sample driving road condition data and the sample regional road condition data to perform combined iterative parameter optimization on the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit in the initialized driving risk estimation network. This process is based on the a-th round of combined network parameter learning of the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit (a is a positive integer not less than 1, and the risk driving area positioning unit and the risk knowledge point estimation unit generated by the 0th round of network parameter learning are the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit in the initialized driving risk estimation network whose parameters have not been updated).

[0055] In the a-th round of combined network parameter learning, the sample area road condition data is first loaded into the risk driving area positioning unit generated by the a-1 round of network parameter learning to generate the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning. For example, for the sample area road condition data of a section of a mountain highway, the section contains continuous curves and some sections have a certain slope. The risk driving area positioning unit generated by the a-1 round of network parameter learning will analyze the sample area road condition data based on the previously learned parameters. It will comprehensively consider the slope value of the section, the curvature radius of the curve, the width of the road, and other information, as well as the experience accumulated in the previous processing of similar road conditions (reflected in the existing network parameters), so as to generate the estimated risk knowledge point data for this sample area road condition data. If the slope of the road section is large and the radius of curvature of the curve is small, according to the calculation and judgment of the network, the estimated risk knowledge point data that may be output is that due to the road slope and curve factors, the vehicle is prone to lose control when driving at high speed, and the risk level is high. The risk level here is determined based on the evaluation mechanism within the network and is interrelated with other elements in the risk knowledge point data (such as road conditions and vehicle behavior expectations).

[0056] Next, the sample driving road condition data is loaded into the risk driving area positioning unit generated by the a-1th round of network parameter learning and the risk knowledge point estimation unit generated by the a-1th round of network parameter learning, and the estimated risk knowledge point data of the target driving risk road condition data output by the a-1th round of network parameter learning and the risk knowledge point estimation data of the sample driving road condition data output by the a-1th round of network parameter learning are generated. The specific operations are as follows:

[0057] When the sample driving road condition data is loaded into the risk driving area positioning unit generated by the a-1 round of network parameter learning, the server will first decompose the sample driving road condition data into Z sample area road condition data. Taking a long section of highway sample driving road condition data mixed with mountainous areas and plains as an example, the server may decompose it into Z sample area road condition data according to different terrain features, traffic signs or fixed distance intervals. Assuming that the total length of this highway is 100 kilometers, the server decomposes it into an area every 10 kilometers, then Z is equal to 10. Then, using the risk driving area positioning unit generated by the a-1 round of network parameter learning, the sample area road condition data with an estimated risk confidence not less than the threshold confidence is extracted from the Z sample area road condition data, and the Y risk sample area road condition data (Y is not less than 1) output by the a round of network parameter learning is generated. For example, among the 10 sample area road condition data, after analysis by the risk driving area positioning unit, it is found that the risk confidence of 3 of the areas is not less than the set threshold confidence, which may be because these 3 areas have special road conditions (such as dense curves, road construction, etc.) or special traffic conditions (such as excessive vehicle density, large speed difference, etc.), then Y is equal to 3. At the same time, the sample area road condition data with the largest estimated risk confidence among the Z sample area road condition data is output as the target driving risk road condition data, and the risk knowledge point corresponding to the estimated risk confidence of the target driving risk road condition data is output as the estimated risk knowledge point data of the target driving risk road condition data. For example, among the 10 areas, there is an area with road construction and extremely high vehicle density, and the risk confidence of this area is the largest, then the road condition data of this area is determined as the target driving risk road condition data, and according to the network's analysis of this area, the output estimated risk knowledge point data may be due to the narrowing of the lanes and the high vehicle density caused by road construction, the safety distance between vehicles is difficult to ensure, and there is a high risk of rear-end collision.

[0058] Afterwards, the Y risk sample area road condition data are loaded into the risk knowledge point estimation unit generated by the a-1 round of network parameter learning to generate the risk knowledge point estimation data of the sample driving road condition data output by the a round of network parameter learning. For the three risk sample area road condition data, the risk knowledge point estimation unit will conduct an in-depth analysis of these data based on the relationship between various risk factors learned previously and the corresponding weight parameters. For example, for one of the risk sample area road condition data, there are many large vehicles in the area with slow speeds, and some small vehicles frequently overtake. The risk knowledge point estimation unit will comprehensively consider factors such as vehicle type, speed difference, overtaking behavior, etc., calculate according to the existing model parameters, and output the risk knowledge point estimation data for this sample driving road condition data. It may be that due to the slow speed of large vehicles and the frequent overtaking of small vehicles, side scratches and rear-end collisions are likely to occur, and the risk level is medium.

[0059] In this process, it is necessary to determine whether the training termination requirements are met. When the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning, the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning, and the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning meet the training termination requirements, the network parameter optimization is terminated, and the risk driving area positioning unit and risk knowledge point estimation unit generated by the a-1-th round of network parameter learning are output as the target risk driving area positioning unit and target risk knowledge point estimation unit in the target driving risk estimation network.

[0060] The process of judging whether the training termination requirements are met involves comparisons in multiple aspects. For the sample driving road condition data, the estimated risk confidence corresponding to the estimated risk knowledge point data of the target driving risk road condition data output by the ath round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample driving road condition data are loaded into the first training cost function to generate the first training cost parameter. For example, if the estimated risk confidence for a certain risk factor (such as vehicle rear-end collision risk) in the estimated risk knowledge point data of the target driving risk road condition data is 0.7, and the corresponding prior risk confidence in the prior risk knowledge point data of the sample driving road condition data is 0.8, the first training cost function will generate the first training cost parameter based on the difference between the two values ​​and the calculation rules set inside the function. Similarly, the estimated risk confidence corresponding to the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample driving road condition data are loaded into the second training cost function to generate second training cost parameters; the estimated risk confidence corresponding to the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample area road condition data are loaded into the third training cost function to generate third training cost parameters.

[0061] Then, the server will check whether the first training cost parameter meets the first training error requirement, whether the second training cost parameter meets the second training error requirement, and whether the third training cost parameter meets the third training error requirement. If all three training cost parameters meet their corresponding training error requirements, then it is determined that the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning, the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning, and the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning meet the training termination requirements. At this point, the network parameter optimization process is terminated, and the risk driving area positioning unit and risk knowledge point estimation unit generated by the a-1-th round of network parameter learning are used as the target risk driving area positioning unit and target risk knowledge point estimation unit in the target driving risk estimation network. In this way, the target driving risk estimation network is generated, and preparations are made for the subsequent risk assessment of the actual target driving road condition data.

[0062] In a possible implementation, step S1022 includes:

[0063] Step S1022-1, load the sample driving road condition data into the risk driving area positioning unit generated by the a-1th round of network parameter learning, generate the estimated risk knowledge point data of the target driving risk road condition data output by the ath round of network parameter learning, and Y risk sample area road condition data output by the ath round of network parameter learning, wherein Y is not less than 1, and the Y risk sample area road condition data are used to represent the area road condition data in the sample driving road condition data whose estimated risk confidence is not less than the threshold confidence, and the target driving risk road condition data are used to represent the sample area road condition data with the largest estimated risk confidence in the sample driving road condition data.

[0064] Step S1022-2, loading the Y risk sample area road condition data into the risk knowledge point estimation unit generated by the a-1th round of network parameter learning, and generating risk knowledge point estimation data of the sample driving road condition data output by the ath round of network parameter learning.

[0065] In a possible implementation, step S1022-1 includes:

[0066] Step S1022-11, load the sample driving road condition data into the risk driving area positioning unit generated by the a-1th round of network parameter learning, and use the risk driving area positioning unit generated by the a-1th round of network parameter learning to disassemble the sample driving road condition data into Z sample area road condition data.

[0067] Step S1022-12, using the risk driving area positioning unit generated by the a-1th round of network parameter learning, extract the sample area road condition data with an estimated risk confidence not less than the threshold confidence from the Z sample area road condition data, and generate the Y risk sample area road condition data output by the ath round of network parameter learning, where Z is not less than Y.

[0068] Step S1022-13, using the risk driving area positioning unit generated by the a-1th round of network parameter learning, outputs the sample area road condition data with the largest estimated risk confidence among the Z sample area road condition data as the target driving risk road condition data, and outputs the risk knowledge points corresponding to the estimated risk confidence of the target driving risk road condition data as the estimated risk knowledge point data of the target driving risk road condition data.

[0069] In this embodiment, the server starts to process the sample traffic condition data and loads the sample traffic condition data into the risk driving area positioning unit generated by the a-1 round of network parameter learning. The sample traffic condition data here is carefully selected from a large amount of historical traffic data and contains rich information, such as driving conditions from different highway sections, different time periods and different traffic conditions. Taking the highway network in a certain area as an example, the sample traffic condition data covers information such as driving conditions in mountainous sections, plain sections, bridge sections and tunnels. This information includes the speed of the vehicle, the type of vehicle (such as cars, trucks, buses, etc.), the distance between vehicles, the curvature of the road, the slope, the road surface conditions, and the impact of weather on driving.

[0070] When the sample driving road condition data is loaded into the risk driving area positioning unit generated by the a-1 round of network parameter learning, the unit will first use the existing network parameters to decompose the sample driving road condition data into Z sample area road condition data. For example, the sample driving road condition data processed by the server is a 100-kilometer-long highway data. The risk driving area positioning unit may decompose it into a unit of 10 kilometers, so Z is equal to 10. During the decomposition process, the risk driving area positioning unit will comprehensively consider multiple factors, such as determining the boundaries of each sample area road condition data based on the physical characteristics of the road (such as road signs, curves, slopes, etc.) and traffic flow characteristics (such as changes in vehicle density, etc.).

[0071] Next, the risk driving area positioning unit generated by the a-1th round of network parameter learning is used to extract the sample area road condition data with an estimated risk confidence not less than the threshold confidence from the Z sample area road condition data, thereby generating the Y risk sample area road condition data output by the ath round of network parameter learning. Assuming the threshold confidence is set to 0.6, in the 10 sample area road condition data, the risk driving area positioning unit will calculate the risk confidence for each area according to its internal risk assessment model. This calculation process will take into account many factors. For example, if a sample area road condition data contains a road construction area, and the vehicle density in this area is large, the speed is slow, and the vehicle types are mixed (large trucks and small cars frequently staggered), then the risk confidence of this area may be high. Assuming that the risk confidence of this area is 0.7, it meets the condition of not less than the threshold confidence. After evaluating the traffic data of all Z sample areas, assuming that the risk confidence of three areas is not less than the threshold confidence, then Y is equal to 3, and the traffic data of these three areas are determined as the traffic data of Y risk sample areas.

[0072] Then, using the risk driving area positioning unit generated by the a-1th round of network parameter learning, the sample area road condition data with the largest estimated risk confidence among the Z sample area road condition data is output as the target driving risk road condition data, and the risk knowledge point corresponding to the estimated risk confidence of the target driving risk road condition data is output as the estimated risk knowledge point data of the target driving risk road condition data. Among the 10 sample area road condition data mentioned above, after the calculation of the risk confidence, it is found that there is an area containing a long curve and there is local dense fog near the curve. At the same time, the vehicle speed difference in this area is large (some vehicles are speeding). The risk confidence of this area is the highest, assuming it is 0.8. Then the sample area road condition data of this area is determined as the target driving risk road condition data. For this target driving risk road condition data, the risk driving area positioning unit will output the corresponding estimated risk knowledge point data according to the risk knowledge point mapping relationship inside it. For example, this risk knowledge point data may be described as due to factors such as long curves, dense fog and large differences in vehicle speeds, there is a high risk of vehicle skidding, rear-end collisions and collisions in the area. The risk situation description here includes multiple risk factors and risk levels (high risk), which is a complex logical description label, and this risk knowledge point data corresponds to the estimated risk confidence level of 0.8 of the target driving risk road condition data.

[0073] After generating the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning and the Y risk sample area road condition data, the Y risk sample area road condition data are loaded into the risk knowledge point estimation unit generated by the a-1th round of network parameter learning to generate the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning. For these Y risk sample area road condition data, each data contains complex driving information in a specific area. Taking one of the risk sample area road condition data as an example, the area is a bridge section with moderate vehicle density but many large vehicles, and there is a crosswind in the area. The risk knowledge point estimation unit will conduct a comprehensive analysis of various factors in the road condition data of this risk sample area based on the previously learned network parameters. It will consider factors such as the driving stability of large vehicles under the influence of crosswinds and the changes in the safe distance between vehicles in this case. After performing similar analysis on each risk sample area road condition data, the risk knowledge point estimation unit will generate the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning. This risk knowledge point estimation data will provide a comprehensive description of the risk situation in the entire sample driving road condition data. For example, it may be described as certain areas in the sample driving road condition data (including the areas corresponding to the Y risk sample area road condition data analyzed previously) where there are risks of vehicle rollover, collision, etc. due to driving stability issues of large vehicles in special environments (such as crosswinds) and factors such as vehicle density. Different risk levels will be given according to the risk level of different areas, forming a comprehensive risk knowledge point estimation data for the sample driving road condition data.

[0074] During the whole process, through the collaborative work of the risk driving area positioning unit and the risk knowledge point estimation unit, various information in the sample driving road condition data is used to gradually generate the estimated risk knowledge point data of the target driving risk road condition data and the risk knowledge point estimation data of the sample driving road condition data according to the established rules and network parameters. These data are of great significance for the subsequent construction of an accurate target driving risk estimation network and risk assessment of actual driving road conditions.

[0075] In a possible implementation, in the a-th round of combined network parameter learning, the method further includes:

[0076] Step A110, load the estimated risk confidence corresponding to the estimated risk knowledge point data of the target driving risk road condition data output by the ath round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample driving road condition data into the first training cost function to generate the first training cost parameter.

[0077] Step A120, load the estimated risk confidence corresponding to the risk knowledge point estimation data of the sample driving road condition data output by the ath round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample driving road condition data into the second training cost function to generate a second training cost parameter.

[0078] Step A130, load the estimated risk confidence corresponding to the estimated risk knowledge point data of the sample area road condition data output by the ath round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample area road condition data into the third training cost function to generate a third training cost parameter.

[0079] Step A140, determining whether the first training cost parameter meets the first training error requirement, whether the second training cost parameter meets the second training error requirement, and whether the third training cost parameter meets the third training error requirement.

[0080] Step A150, when it is extracted that the first training cost parameter meets the first training error requirement, the second training cost parameter meets the second training error requirement, and the third training cost parameter meets the third training error requirement, it is determined that the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning, the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning, and the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning meet the training termination requirements.

[0081] In this embodiment, first, the server will load the estimated risk confidence corresponding to the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample driving road condition data into the first training cost function to generate the first training cost parameter. Taking the previously mentioned highway sample driving road condition data as an example, assuming that the target driving risk road condition data is a specific section of a mountain highway, due to the presence of continuous bends, local dense fog weather and large differences in vehicle speeds, the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning is a high risk situation of vehicle skidding, rear-end collision and collision, and the corresponding estimated risk confidence is 0.8. The prior risk knowledge point data of the sample driving road condition data is derived by professionals based on a large amount of historical data and traffic engineering principles. The prior risk knowledge point data of the section may also be a situation where there is a high risk of vehicle driving, and the corresponding prior risk confidence is 0.9.

[0082] The first training cost function is constructed according to a specific algorithm, and it will be calculated based on the values ​​of these two confidences. This function may take into account the difference between the estimated risk confidence and the prior risk confidence, the absolute value of the difference, or other related calculation methods. For example, the first training cost function may be to calculate the square of the difference between the two confidences, that is, \((0.8-0.9)^2=0.01\), and this 0.01 is the first training cost parameter. This parameter reflects the degree of difference between the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning and the prior risk knowledge point data of the sample driving road condition data.

[0083] Next, the server loads the estimated risk confidence corresponding to the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample driving road condition data into the second training cost function to generate the second training cost parameter. Assume that in the sample driving road condition data, the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning indicates that there is a certain collision risk in multiple areas due to factors such as mixed vehicle types, speed differences of different vehicles, and road conditions (such as potholes in some sections). The corresponding estimated risk confidence is 0.7. The prior risk confidence corresponding to the risk judgment of these areas of the prior risk knowledge point data of the sample driving road condition data is 0.8. The second training cost function is calculated according to its internal calculation rules, such as calculating the absolute value of the difference between the two, that is, \(\vert0.7-0.8\vert=0.1\), 0.1 is the second training cost parameter, which reflects the degree of deviation between the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning and the prior risk knowledge point data.

[0084] Then, the server loads the estimated risk confidence corresponding to the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning and the prior risk confidence corresponding to the prior risk knowledge point data of the sample area road condition data into the third training cost function to generate the third training cost parameter. For example, for a certain sample area road condition data, it is a section of a plain highway. The estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning indicates that there is a risk of vehicle collision due to the sudden increase in the traffic volume of the section during a specific period and some vehicles have illegal lane changes. The corresponding estimated risk confidence is 0.6. The prior risk confidence of the prior risk knowledge point data of the sample area road condition data for this section in this case is 0.7. The third training cost function is calculated according to its own algorithm. It is assumed that the absolute value of the difference between the ratio of the two and 1 is calculated, that is, \(\vert\frac{0.6}{0.7}-1\vert=\vert0.857-1\vert=0.143\), 0.143 is the third training cost parameter, which represents the degree of deviation between the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning and the prior risk knowledge point data.

[0085] Afterwards, the server determines whether the first training cost parameter meets the first training error requirement, whether the second training cost parameter meets the second training error requirement, and whether the third training cost parameter meets the third training error requirement. The first training error requirement, the second training error requirement, and the third training error requirement are pre-set numerical ranges or thresholds, which are determined based on the need to build an accurate target driving risk estimation network. For example, the first training error requirement may be set to a first training cost parameter less than 0.05, the second training error requirement may be set to a second training cost parameter less than 0.15, and the third training error requirement may be set to a third training cost parameter less than 0.2.

[0086] For the first training cost parameter 0.01 calculated previously, it is less than 0.05 and meets the first training error requirement; the second training cost parameter 0.1 is less than 0.15 and meets the second training error requirement; the third training cost parameter 0.143 is less than 0.2 and meets the third training error requirement. When the server extracts that the first training cost parameter meets the first training error requirement, the second training cost parameter meets the second training error requirement, and the third training cost parameter meets the third training error requirement, it can be determined that the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning, the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning, and the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning meet the training termination requirements. This means that in the a-th round of network parameter learning, through the processing of the sample driving road condition data and the sample area road condition data, the differences between the various risk knowledge point data and the prior risk knowledge point data are already within an acceptable range, and the learning of the network parameters has reached a relatively ideal state, and no further parameter adjustment is required, so the network parameter optimization process can be terminated. This process ensures that the target driving risk estimation network can accurately perform risk assessment on driving road condition data, providing a reliable basis for subsequent processing of target driving road condition data.

[0087] In a possible implementation, step S130 includes:

[0088] Step S131: Utilize the target risk driving area positioning unit in the target driving risk estimation network to encode and represent the X driving risk road condition data to generate X graph encoding vectors.

[0089] Step S132: Load the average vector of the X graph encoding vectors into the fully connected mapping branch in the target risk knowledge point estimation unit, and use the fully connected mapping branch to predict the average vector to generate risk knowledge point estimation data of the target driving road condition data.

[0090] In a possible implementation, step S120 includes:

[0091] Step S121: Decompose the target driving road condition data into Z regional road condition data using the target risk driving area positioning unit in the target driving risk estimation network, where Z is not less than X.

[0092] Step S122, using the target risk driving area positioning unit to extract the X driving risk road condition data from the Z area road condition data, the X driving risk road condition data are used to represent the area road condition data of the Z area road condition data whose estimated risk confidence is not less than the threshold confidence.

[0093] In this embodiment, first, in the process of extracting X driving risk road condition data from the target driving road condition data, the server uses the target risk driving area positioning unit in the target driving risk estimation network to operate. The target driving road condition data contains various driving related information of the target highway target control unit, such as the vehicle's driving speed, vehicle spacing, lane occupancy, the impact of weather conditions on the road, and the physical characteristics of the road itself (such as curves, slopes, etc.). When the target risk driving area positioning unit starts working, it will disassemble the target driving road condition data into Z regional road condition data according to certain rules, where Z is not less than X.

[0094] Take a specific target driving road condition data as an example. Assuming that this is a 50-kilometer highway section data, the target risk driving area positioning unit may be broken down into a region of 5 kilometers, so that Z is equal to 10. During the breaking process, the target risk driving area positioning unit will comprehensively consider multiple factors, such as traffic signs and geographical features (whether it is close to bridges, tunnels, etc.) of different sections. For each broken down regional road condition data, the target risk driving area positioning unit will conduct a risk assessment and calculate the estimated risk confidence of each region. This calculation process involves complex algorithms that comprehensively consider multiple factors such as vehicle driving conditions, weather, and roads in the region. For example, in one of the regions, if the vehicle density is too high (more than 100 vehicles per kilometer), there is rainy weather that makes the road slippery, and there is a curve in the region, then the risk of this region will be relatively high, and the target risk driving area positioning unit will calculate the estimated risk confidence of this region.

[0095] Next, the target risk driving area positioning unit extracts X risk driving road condition data from the Z regional road condition data, where the X risk driving road condition data are used to represent the regional road condition data with an estimated risk confidence not less than the threshold confidence in the Z regional road condition data. Assuming that the threshold confidence is set to 0.6, after calculating the risk confidence of the 10 regional road condition data, it is found that the risk confidence of 3 regions is not less than 0.6, then X is equal to 3. The road condition data of these 3 regions are extracted as X risk driving road condition data, which contain rich information, such as the specific driving speed distribution of vehicles in high-risk areas, the proportion of vehicle types, the specific curvature and slope values ​​of the road, and weather conditions and other detailed information.

[0096] Then, the target risk knowledge point estimation unit in the target driving risk estimation network is used to estimate the X driving risk road condition data to generate the risk knowledge point estimation data of the target driving road condition data. In this process, the target risk driving area positioning unit in the target driving risk estimation network is used to encode and represent the X driving risk road condition data to generate X graph encoding vectors. For each extracted driving risk road condition data, the target risk driving area positioning unit will convert various information therein into a graph encoding vector. For example, for the vehicle speed information in a driving risk road condition data, it may be encoded according to the numerical range of the speed, such as encoding 0-60 kilometers / hour as a specific numerical range, 60-120 kilometers / hour as another numerical range, etc.; for vehicle types, different types of vehicles (such as cars, trucks, buses, etc.) will also be encoded into different values; the curvature and slope of the road will also be encoded according to certain rules. In this way, each driving risk road condition data is converted into a graph encoding vector, and a total of X graph encoding vectors are generated.

[0097] Afterwards, the average vector of the X graph encoding vectors is loaded into the fully connected mapping branch in the target risk knowledge point estimation unit, and the average vector is predicted using the fully connected mapping branch to generate the risk knowledge point estimation data of the target driving road condition data. Assuming that X is equal to 3, that is, there are 3 graph encoding vectors, the target risk knowledge point estimation unit will first calculate the average vector of these 3 graph encoding vectors. This calculation process will comprehensively consider each element in each graph encoding vector and calculate according to a certain weighted average rule. After obtaining the average vector, it is loaded into the fully connected mapping branch. The fully connected mapping branch is an important component of the target risk knowledge point estimation unit. It predicts this average vector based on the network parameters obtained from the previous training.

[0098] During the prediction process, the fully connected mapping branch generates risk knowledge point estimation data for the target driving road condition data based on the values ​​in the averaged vector, combined with the internal neuron connection weights and activation functions. This risk knowledge point estimation data is a complex logical description label that contains information from multiple aspects. For example, for the three driving risk road condition data mentioned above (which may correspond to different road sections), the risk knowledge point estimation data may be described as follows: In the first high-risk area (determined by the previous driving risk road condition data), due to the high vehicle density (specific density value), the existence of curves (curve curvature value) and the rainy weather (rainfall intensity level), the vehicle has the risk of skidding and rear-end collision, and the risk level is high; in the second high-risk area, the vehicle types are mixed (the proportion of various vehicle types), some vehicles are speeding (speeding ratio) and the road has a certain slope (slope value), there is a risk of vehicle loss of control and collision, and the risk level is medium; in the third high-risk area, although the vehicle density is moderate, due to the presence of dense fog (dense fog concentration level) and low road visibility, the vehicle is prone to chain collision risk, and the risk level is medium. This risk knowledge point estimation data comprehensively reflects the risk status in the target driving road condition data, and provides an important basis for subsequent control plan decisions for the control nodes corresponding to the target control unit based on the risk knowledge point estimation data.

[0099] During the whole process, the server conducts in-depth analysis and processing of the target driving road condition data through the collaborative work of the target risk driving area positioning unit and the target risk knowledge point estimation unit in the target driving risk estimation network, accurately extracts the risk area data from the target driving road condition data, and further generates comprehensive and detailed risk knowledge point estimation data, thereby providing effective risk assessment and decision-making support for highway traffic management.

[0100] Figure 2 The hardware structure of the highway driving risk prediction system 100 based on the control unit provided in the embodiment of the present invention for implementing the above-mentioned highway driving risk prediction method based on the control unit is shown. Figure 2 As shown, the highway driving risk prediction system 100 based on the control unit may include a processor 110 , a machine-readable storage medium 120 , a bus 130 and a communication unit 140 .

[0101] 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 highway driving risk prediction system 100 based on the control unit uses to execute or use to complete the exemplary method described in the present invention.

[0102] 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 highway driving risk prediction method based on the control unit 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.

[0103] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned highway driving risk prediction system 100 based on the control unit. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0104] 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 highway driving risk prediction method based on the control unit is implemented.

[0105] 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 highway driving risk prediction method based on a control unit, characterized in that: The method comprises: Acquire target traffic condition data, wherein the target traffic condition data is traffic condition data to be analyzed obtained by real-time traffic condition monitoring of a target control unit of a target highway; The target driving road condition data is loaded into a target driving risk estimation network, and a target risk driving area positioning unit in the target driving risk estimation network is used to extract X driving risk road condition data from the target driving road condition data; Using a target risk knowledge point estimation unit in the target driving risk estimation network to estimate the X driving risk road condition data, and generate risk knowledge point estimation data of the target driving road condition data, wherein X is not less than 1, and the X driving risk road condition data are used to represent regional road condition data in the target driving road condition data whose estimated risk confidence is not less than a threshold confidence; Making a control plan decision for the control node corresponding to the target control unit based on the risk knowledge point estimation data of the target driving road condition data; Among them, the target driving risk estimation network is a deep learning network generated by using sample driving road condition data and sample area road condition data to optimize the parameters of the initialized driving risk estimation network until the following training termination requirements are met: the prior risk knowledge point data of the sample driving road condition data and the estimated risk knowledge point data of the target driving risk road condition data meet the first training error requirement, and the target driving risk road condition data is the driving risk road condition data with the largest estimated risk confidence in the sample driving road condition data output by the initialized risk driving area positioning unit in the initialized driving risk estimation network.

2. The highway driving risk prediction method based on the control unit according to claim 1 is characterized in that: The training termination requirements also include: The prior risk knowledge point data of the sample driving road condition data and the risk knowledge point estimation data of the sample driving road condition data output by the initialized risk knowledge point estimation unit in the initialized driving risk estimation network meet the second training error requirement, and the prior risk knowledge point data of the sample area road condition data and the estimated risk knowledge point data of the sample area road condition data output by the initialized risk driving area positioning unit in the initialized driving risk estimation network meet the third training error requirement.

3. The highway driving risk prediction method based on the control unit according to claim 2 is characterized in that: Before loading the target driving road condition data into the target driving risk estimation network, the method further includes: Acquire the sample driving road condition data and the sample area road condition data; The sample driving road condition data and the sample area road condition data are used to perform combined iterative parameter optimization on the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit in the initialized driving risk estimation network to generate the target driving risk estimation network. During the network parameter optimization process, if the initialized driving risk estimation network does not meet the training termination requirements, the neuron weight information in the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit is updated. If the initialized driving risk estimation network meets the training termination requirements, the network parameter optimization is terminated, and the initialized driving risk estimation network when the network parameter optimization is terminated is output as the target driving risk estimation network, and the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit when the network parameter optimization is terminated are respectively output as the target risk driving area positioning unit and the target risk knowledge point estimation unit in the target driving risk estimation network.

4. The highway driving risk prediction method based on the control unit according to claim 3 is characterized in that: The method of using the sample driving road condition data and the sample area road condition data to perform combined iterative parameter optimization on the initialized risk driving area positioning unit and the initialized risk knowledge point estimation unit in the initialized driving risk estimation network includes: Based on the following operations, the initialization risk driving area positioning unit and the initialization risk knowledge point estimation unit in the initialization driving risk estimation network are subjected to the ath round of combined network parameter learning, where a is a positive integer not less than 1, and the risk driving area positioning unit and the risk knowledge point estimation unit generated by the 0th round of network parameter learning are the initialization risk driving area positioning unit and the initialization risk knowledge point estimation unit in the initialization driving risk estimation network whose parameters have not been updated, including: Loading the sample area road condition data into the risk driving area positioning unit generated by the a-1th round of network parameter learning to generate estimated risk knowledge point data of the sample area road condition data output by the ath round of network parameter learning; The sample driving road condition data is loaded into the risk driving area positioning unit generated by the a-1th round of network parameter learning and the risk knowledge point estimation unit generated by the a-1th round of network parameter learning, to generate the estimated risk knowledge point data of the target driving risk road condition data output by the a-1th round of network parameter learning, and the risk knowledge point estimation data of the sample driving road condition data output by the a-1th round of network parameter learning; When the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning, the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning, and the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning meet the training termination requirements, the network parameter optimization is terminated, and the risk driving area positioning unit and the risk knowledge point estimation unit generated by the a-1-th round of network parameter learning are respectively output as the target risk driving area positioning unit and the target risk knowledge point estimation unit in the target driving risk estimation network.

5. The highway driving risk prediction method based on the control unit according to claim 4 is characterized in that: The step of loading the sample driving road condition data into the risk driving area positioning unit generated by the a-1th round of network parameter learning and the risk knowledge point estimation unit generated by the a-1th round of network parameter learning, generating the estimated risk knowledge point data of the target driving risk road condition data output by the a-1th round of network parameter learning, and the risk knowledge point estimation data of the sample driving road condition data output by the a-1th round of network parameter learning, comprises: The sample driving road condition data is loaded into the risk driving area positioning unit generated by the a-1th round of network parameter learning, and the estimated risk knowledge point data of the target driving risk road condition data output by the ath round of network parameter learning and Y risk sample area road condition data output by the ath round of network parameter learning are generated, wherein Y is not less than 1, and the Y risk sample area road condition data are used to represent the area road condition data in which the estimated risk confidence is not less than the threshold confidence in the sample driving road condition data, and the target driving risk road condition data are used to represent the sample area road condition data in which the estimated risk confidence is the largest in the sample driving road condition data; The Y risk sample area road condition data are loaded into the risk knowledge point estimation unit generated by the a-1th round of network parameter learning to generate risk knowledge point estimation data of the sample driving road condition data output by the ath round of network parameter learning.

6. The highway driving risk prediction method based on the control unit according to claim 5 is characterized in that: The method of loading the sample driving road condition data into the risk driving area positioning unit generated by the a-1th round of network parameter learning, generating the estimated risk knowledge point data of the target driving risk road condition data output by the ath round of network parameter learning, and Y risk sample area road condition data output by the ath round of network parameter learning, includes: The sample driving road condition data is loaded into the risk driving area positioning unit generated by the a-1th round of network parameter learning, and the sample driving road condition data is disassembled into Z sample area road condition data using the risk driving area positioning unit generated by the a-1th round of network parameter learning; Using the risk driving area positioning unit generated by the a-1th round of network parameter learning, sample area road condition data with an estimated risk confidence not less than the threshold confidence are extracted from the Z sample area road condition data, and the Y risk sample area road condition data output by the ath round of network parameter learning are generated, where Z is not less than Y; Utilizing the risk driving area positioning unit generated by the a-1th round of network parameter learning, the sample area road condition data with the largest estimated risk confidence among the Z sample area road condition data is output as the target driving risk road condition data, and the risk knowledge points corresponding to the estimated risk confidence of the target driving risk road condition data are output as the estimated risk knowledge point data of the target driving risk road condition data.

7. The highway driving risk prediction method based on the control unit according to any one of claims 4 to 6, characterized in that: In the a-th round of combined network parameter learning, the method further includes: Loading the estimated risk confidence corresponding to the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning and the priori risk confidence corresponding to the priori risk knowledge point data of the sample driving road condition data into a first training cost function to generate a first training cost parameter; Loading the estimated risk confidence corresponding to the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning and the priori risk confidence corresponding to the priori risk knowledge point data of the sample driving road condition data into the second training cost function to generate a second training cost parameter; Loading the estimated risk confidence corresponding to the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning and the priori risk confidence corresponding to the priori risk knowledge point data of the sample area road condition data into a third training cost function to generate a third training cost parameter; Determine whether the first training cost parameter meets the first training error requirement, whether the second training cost parameter meets the second training error requirement, and whether the third training cost parameter meets the third training error requirement; When it is extracted that the first training cost parameter meets the first training error requirement, the second training cost parameter meets the second training error requirement, and the third training cost parameter meets the third training error requirement, it is determined that the estimated risk knowledge point data of the sample area road condition data output by the a-th round of network parameter learning, the estimated risk knowledge point data of the target driving risk road condition data output by the a-th round of network parameter learning, and the risk knowledge point estimation data of the sample driving road condition data output by the a-th round of network parameter learning meet the training termination requirements.

8. The highway driving risk prediction method based on the control unit according to claim 1 is characterized in that: The step of estimating the X driving risk road condition data using a target risk knowledge point estimation unit in the target driving risk estimation network to generate risk knowledge point estimation data of the target driving road condition data includes: Using the target risk driving area positioning unit in the target driving risk estimation network to encode and represent the X driving risk road condition data, and generate X graph encoding vectors; The average vector of the X graph encoding vectors is loaded into the fully connected mapping branch in the target risk knowledge point estimation unit, and the average vector is predicted by using the fully connected mapping branch to generate risk knowledge point estimation data of the target driving road condition data.

9. The highway driving risk prediction method based on the control unit according to claim 1 is characterized in that: The step of extracting X driving risk road condition data from the target driving road condition data using the target risk driving area positioning unit in the target driving risk estimation network includes: Using the target risk driving area positioning unit in the target driving risk estimation network, the target driving road condition data is disassembled into Z area road condition data, where Z is not less than X; The target risk driving area positioning unit is used to extract the X driving risk road condition data from the Z area road condition data, and the X driving risk road condition data are used to represent the area road condition data of the Z area road condition data whose estimated risk confidence is not less than the threshold confidence.

10. A highway driving risk prediction system based on a control unit, characterized in that: The highway driving risk prediction system based on the control unit 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 highway driving risk prediction method based on the control unit as described in any one of claims 1 to 9 above.

Images