AI-based early warning method and system for driving safety on slippery roads

Abstract

The present invention provides an AI-based slippery road driving safety warning method and system, which relates to the field of artificial intelligence technology. First, the dangerous area information in the candidate slippery road conditions is accurately extracted, and the driving state description features extracted from the vehicle sensor data are combined to effectively identify the dangerous situations that the vehicle may encounter in the present and future. Through the circular reasoning mechanism, the method can accurately predict the change sequence of the dangerous area based on a variety of dynamic factors, such as the interference features to be injected, the reasoning unit length, etc., which not only improves the accuracy and timeliness of the driving safety warning, but also enables the warning system to better adapt to the complex and changeable slippery road environment, thereby providing more reliable safety protection for the driver and significantly reducing the risk of traffic accidents caused by slippery roads.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI-based slippery road driving safety warning method and system. Background Art

[0002] In the current traffic environment, slippery roads are one of the main causes of traffic accidents. Especially in rainy and snowy days or when there is water or oil on the road, the slippery condition of the road will significantly reduce the adhesion between the tire and the road surface, increasing the risk of the vehicle skidding and losing control. Therefore, how to effectively warn of driving safety on slippery roads has become a technical problem that needs to be solved urgently.

[0003] Traditional driving safety warning methods mainly rely on vehicle sensors, such as wheel speed sensors and acceleration sensors, to detect the real-time status of the vehicle, such as speed, acceleration, and side slip, and then determine whether the vehicle is in a dangerous state. However, these methods can often only respond to dangers that have already occurred, but cannot predict and warn potential dangerous areas in advance. In addition, traditional methods also lack sufficient intelligence and adaptability when dealing with complex and changeable slippery road environments. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides an AI-based slippery road driving safety warning method, the method comprising:

[0005] Acquire candidate slippery road condition data, and extract dangerous area information corresponding to the road in the candidate slippery road condition data through a road condition dangerous area prediction network;

[0006] Acquire vehicle sensor data, and extract driving state description features of the vehicle sensor data through a driving state description network;

[0007] Extracting a target description feature from the driving state description feature according to a set time interval, and acquiring a plurality of forward slippery road condition data corresponding to the target description feature;

[0008] According to the interference features to be injected, the inference unit length corresponding to the loop reasoning stage, the dangerous area information, the target description features, and the forward slippery road condition data, a driving safety inference network is used to infer the dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data according to at least two rounds of loop reasoning stages, so as to perform a slippery road driving safety warning based on the dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data.

[0009] On the other hand, an embodiment of the present invention further provides an AI-based slippery road driving safety warning system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0010] Based on the above aspects, the embodiment of the present application realizes accurate early warning of driving safety on slippery roads by integrating a road condition dangerous area prediction network, a driving state description network and a driving safety reasoning network. Specifically, the present invention first accurately extracts the dangerous area information in the candidate slippery road conditions, and combines the driving state description features extracted from the vehicle sensor data to effectively identify the dangerous situations that the vehicle may encounter in the present and future. Through the circular reasoning mechanism, the method can accurately predict the change sequence of the dangerous area based on a variety of dynamic factors, such as the interference features to be injected, the reasoning unit length, etc., which not only improves the accuracy and timeliness of driving safety warnings, but also enables the early warning system to better adapt to the complex and changeable slippery road environment, thereby providing drivers with more reliable safety protection and significantly reducing the risk of traffic accidents caused by slippery roads. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the AI-based slippery road driving safety warning method provided in an embodiment of the present invention.

[0012] Figure 2 It is a schematic diagram of the hardware architecture of the AI-based slippery road driving safety warning system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 The figure is a flow chart of an AI-based slippery road driving safety warning method provided by an embodiment of the present invention. The AI-based slippery road driving safety warning method is introduced in detail below.

[0014] Step S110, obtaining candidate slippery road condition data, and extracting dangerous area information corresponding to the road in the candidate slippery road condition data through a road condition dangerous area prediction network.

[0015] In this embodiment, there are multiple road sections in a certain area that may be slippery. For example, on some winding mountain road sections in mountainous areas, the road surface is often slippery due to early morning dew or poor drainage after rainfall; and some roads near rivers are also prone to slippery conditions due to the influence of water vapor.

[0016] Based on this, the server obtains candidate slippery road condition data from multiple data sources. These candidate slippery road condition data sources include meteorological sensors distributed on the road, which can monitor meteorological data such as rainfall and humidity in real time; and road surface sensors, which can detect physical quantities directly related to the degree of slipperiness, such as the friction coefficient of the road surface and the depth of water accumulation. For example, data from a meteorological sensor shows that the rainfall on a certain section of a winding mountain road reached 5 mm per hour, and the road surface sensor detected that the friction coefficient of the section was 0.3 (usually the friction coefficient of a dry road surface is between 0.6-0.8), and the depth of water accumulation was 2 cm. These data constitute part of the candidate slippery road condition data.

[0017] The server inputs the obtained candidate slippery road condition data into the road condition dangerous area prediction network. This road condition dangerous area prediction network is pre-trained with a large amount of data and can identify dangerous areas of the road under different road condition data. For example, for the data of the above-mentioned winding mountain highway section, the road condition dangerous area prediction network judges that the bends and downhill sections are dangerous areas based on information such as rainfall intensity, friction coefficient and water depth, combined with its internally stored terrain data (such as curve radius, slope, etc.). At the bend, due to wetness, the vehicle is prone to side sliding; on the downhill section, the braking distance will increase significantly, and rear-end collisions are prone to occur. The network extracts the location information of these dangerous areas (such as the latitude and longitude coordinate range indicating the specific location of the bend and downhill section), the degree of danger (for example, the comprehensive assessment of the danger level is high based on factors such as friction coefficient and water depth) and other information as the dangerous area information corresponding to the road in the candidate slippery road condition data.

[0018] Step S120, acquiring vehicle sensor data, and extracting driving state description features of the vehicle sensor data through a driving state description network.

[0019] In this scenario, the server maintains a connection with the onboard systems of many vehicles to obtain vehicle sensor data. For example, a car is driving on the previously mentioned slippery winding mountain road. The vehicle's sensors include speed sensors, acceleration sensors, steering angle sensors, etc.

[0020] The data from the speed sensor shows that the vehicle is currently traveling at 60 kilometers per hour, the acceleration sensor detects that the vehicle has a slight deceleration trend, and the steering angle sensor indicates that the vehicle is making a small left turn. After the server obtains these vehicle sensor data, it inputs them into the driving status description network.

[0021] The driving state description network will analyze and process these vehicle sensor data to extract driving state description features. For the data of 60 kilometers per hour, the driving state description network will take into account the current slippery conditions of the road and determine whether this speed is appropriate under such road conditions. On a slippery road, this speed may be relatively fast, increasing the possibility of danger. For the vehicle's deceleration trend, the driving state description network will analyze whether it is normal braking deceleration or unexpected deceleration due to slippery road conditions. For left-turn operations, the stability and safety of the turn are judged based on the speed and slippery conditions of the road. Finally, the driving state description network will output a series of driving state description features, such as "driving at a relatively high speed on a slippery winding mountain road and turning left with a tendency to decelerate" and other feature descriptions that contain key information about the vehicle's driving status.

[0022] Step S130 , extracting a target description feature from the driving state description feature according to a set time interval, and acquiring a plurality of forward slippery road condition data corresponding to the target description feature.

[0023] In this embodiment, a time interval is preset, for example, the past 10 minutes. Within the 10 minutes, the target description feature is extracted from the previously acquired driving state description feature.

[0024] Assume that within these 10 minutes, the driving state description features include a series of information such as the vehicle's speed change on different road sections, the frequency of turning operations, and the braking conditions. The server decomposes these driving state description features into multiple state description sub-features associated with the time and space domains. For example, according to the division of each 2-minute sub-interval, 5 state description sub-features are obtained. Each state description sub-feature contains relevant information such as the average speed, maximum steering angle, and number of braking times of the vehicle within these 2 minutes.

[0025] The target description feature is determined from these state description sub-features. Assuming that in the middle 4-6 minutes of the 10 minutes, the vehicle's speed fluctuates greatly and there are multiple emergency braking operations, then the state description sub-feature corresponding to this period is determined as the target description feature.

[0026] Based on the state node order between the multiple state description sub-features, the forward state description sub-feature that is located before the target description feature is determined. Since the target description feature is the state description sub-feature of 4-6 minutes, the state description sub-feature of the previous 2-4 minutes is the forward state description sub-feature.

[0027] The server obtains multiple forward slippery road condition data from the forward dangerous area change sequence corresponding to these forward state description sub-features. For example, the forward state description sub-feature contains the driving conditions of the vehicle on a specific road section. Based on this road section information, the server obtains forward slippery road condition data such as rainfall, road friction coefficient, and water depth of the road section 2-4 minutes ago from the previously stored road condition data. These data will help in the subsequent reasoning of the dangerous area change sequence.

[0028] Step S140, based on the interference features to be injected, the reasoning unit length corresponding to the loop reasoning stage, the dangerous area information, the target description features, and the forward slippery road condition data, use the driving safety reasoning network to infer the dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data in at least two rounds of loop reasoning stages, so as to perform a slippery road driving safety warning based on the dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data.

[0029] In this embodiment, firstly, the interference feature to be injected is obtained, the inference unit length is set, and the number of cycles is set. For example, the interference feature to be injected may be to simulate the interference of sudden strong wind on vehicle driving, the inference unit length is set to 1 minute, and the number of cycles is set to 3.

[0030] According to the interference features to be injected, the set inference unit length, the dangerous area information, the target description features, and the forward slippery road condition data, the driving safety inference network is used to infer the initial dangerous area change sequence corresponding to the first loop inference stage.

[0031] The driving safety reasoning network combines the previously acquired dangerous area information (such as dangerous areas on bends and downhill sections of winding mountain roads), target description features (large speed fluctuations and multiple emergency brakes of the vehicle in 4-6 minutes), and forward slippery road condition data (rainfall and friction coefficient 2-4 minutes ago), taking into account the interference features to be injected (strong wind interference), and uses 1 minute as the reasoning unit length to reason in the first loop reasoning stage. For example, in the first minute, due to the unstable vehicle speed and slippery road surface, coupled with strong wind interference, the reasoning network predicts that the degree of danger will further increase at the bend, the vehicle will have a higher risk of skidding, and the braking distance will be further extended on the downhill section, which may lead to an increased risk of rear-end collision, thereby generating an initial dangerous area change sequence and recording the changes in the degree of danger of each dangerous area within this 1 minute.

[0032] The initial danger zone change sequence is subjected to interference enhancement processing to generate interference enhancement data, and the set inference unit length is adjusted to the inference unit length corresponding to the current loop inference stage. For example, the danger degree data in the initial danger zone change sequence is enhanced according to certain rules. Assuming that the original side slip risk at the curve is 30%, it becomes 40% after interference enhancement. At the same time, according to the requirements of the current loop inference stage, the inference unit length is adjusted to a possible 2 minutes (assuming that the current loop inference stage requires a change in the inference unit length).

[0033] According to the interference enhancement data, the reasoning unit length corresponding to the current loop reasoning stage, the dangerous area information, the target description features, and the forward slippery road condition data, the driving safety reasoning network is used to infer the dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data according to the loop reasoning stage with a set number of cycles.

[0034] The driving safety reasoning network is used to initialize the reasoning state, which contains all the intermediate variables and parameters required in the reasoning process, such as the basic danger level of the current dangerous area, the initial state parameters of the vehicle, etc.

[0035] According to the inference unit length (2 minutes) corresponding to the current loop inference stage, the inference data of the initial inference unit is extracted from the interference enhancement data. The inference data contains the change information of the dangerous area of ​​the road within the inference unit length, and the target description features and forward slippery road condition data in the corresponding time period are extracted. For example, during these 2 minutes, the interference enhancement data shows that the risk of sideslip at the curve continues to increase, while the target description features indicate that the vehicle speed is still unstable, and the forward slippery road condition data shows that the rainfall has increased slightly.

[0036] The driving safety reasoning network is used to perform the first round of reasoning, combining the reasoning data in the initial reasoning unit, the initialized reasoning state, and the dangerous area information. The reasoning state is updated during the reasoning process, and the first reasoning result is generated. This first reasoning result describes the changing trend of the dangerous area of ​​the road in the current reasoning unit. For example, the reasoning results show that within these 2 minutes, the risk of skidding at the curve increased from 40% to 50% due to unstable vehicle speed, increased rainfall, and strong wind interference, and the risk of rear-end collision on the downhill section also increased accordingly.

[0037] According to the set reasoning unit length, move to the next reasoning unit, extract the interference enhancement data, target description features, and forward slippery road condition data in the next reasoning unit, update the reasoning state, and output the updated reasoning unit data and reasoning state. For example, in the next 2-minute reasoning unit, the interference enhancement data shows that the risk of skidding at the curve may have a slight downward trend due to some emergency measures taken by the vehicle (analyzed from the target description features), but due to the increase in the depth of water accumulation (obtained from the forward slippery road condition data), the overall risk is still at a high level.

[0038] Repeat the above steps and continuously update the reasoning unit data and reasoning status. Each reasoning result will be generated based on the previous reasoning result and the data of the current reasoning unit to gradually build a complete dangerous area change sequence corresponding to the current cyclic reasoning stage. The dangerous area change sequence contains all the reasoning results from the initial reasoning unit to the last reasoning unit, describing the changes in the dangerous area of ​​the road during the entire set time interval. For example, after 3 rounds of cyclic reasoning, the complete dangerous area change sequence shows that within the entire 10 minutes, the side slip risk at the curve first increased and then decreased slightly but remained at a high level, and the rear-end collision risk on the downhill section continued to fluctuate and increase.

[0039] The sequence of changes in the dangerous area is evaluated to determine whether there is a potential dangerous situation. If the evaluation result shows that there is a potential dangerous situation, the safety warning mechanism is triggered, and the safety warning mechanism warns the driver in various ways. For example, if the risk of sideslipping at a curve exceeds 60%, the server will determine that there is a potential dangerous situation, and then send a warning message to the driver through the vehicle's on-board system, such as displaying "The risk of sideslipping at the curve is extremely high, please slow down" on the dashboard, and the driver can also be informed through voice prompts.

[0040] In this way, effective analysis and warning of driving safety on slippery roads can be carried out to ensure the safety of drivers.

[0041] Based on the above steps, the embodiment of the present application realizes accurate early warning of driving safety on slippery roads by integrating the road condition dangerous area prediction network, the driving state description network and the driving safety reasoning network. Specifically, the present invention first accurately extracts the dangerous area information in the candidate slippery road conditions, and combines the driving state description features extracted from the vehicle sensor data to effectively identify the dangerous situations that the vehicle may encounter in the present and future. Through the circular reasoning mechanism, the method can accurately predict the change sequence of the dangerous area based on a variety of dynamic factors, such as the interference features to be injected, the reasoning unit length, etc., which not only improves the accuracy and timeliness of driving safety warnings, but also enables the early warning system to better adapt to the complex and changeable slippery road environment, thereby providing drivers with more reliable safety protection and significantly reducing the risk of traffic accidents caused by slippery roads.

[0042] In a possible implementation, step S130 includes:

[0043] Step S131 : decomposing the driving state description feature into a plurality of state description sub-features associated in time and space domains according to a set time interval.

[0044] Step S132, determining a target description feature from the plurality of state description sub-features.

[0045] Step S133: determining a forward state description sub-feature located before the target description feature based on the state node order among the plurality of state description sub-features.

[0046] Step S134, obtaining a plurality of forward slippery road condition data from the forward dangerous area change sequence corresponding to the forward state description sub-feature.

[0047] In this embodiment, a specific time interval is set, such as the past 15 minutes, during which the vehicle has been driving on a potentially slippery winding mountain road. The server receives the driving state description features output by the driving state description network, which cover various driving state information of the vehicle in these 15 minutes. In order to analyze the vehicle state more carefully, the server decomposes the driving state description features into multiple state description sub-features associated with the time and space domains according to the set 15-minute time interval. Specifically, the server divides 15 minutes into 5 state description sub-features at intervals of 3 minutes. Each state description sub-feature contains the comprehensive driving information of the vehicle within these 3 minutes, such as the average speed, the fluctuation range of the speed, the change of the steering angle, the frequency of the braking operation, and the deviation of the vehicle's driving trajectory relative to the center line of the road. These state description sub-features not only reflect the state changes of the vehicle in the time dimension, but are also associated with the spatial position of the road where the vehicle is located.

[0048] After decomposing these multiple spatiotemporal-domain related state descriptor features, the server began to determine the target description feature from them. During the 15-minute driving process, the server found that there were special circumstances in the state descriptor features corresponding to the 6th to 9th minutes. During this time period, the vehicle's speed fluctuation range increased significantly, from the original relatively stable speed of 50-55 kilometers per hour to frequent fluctuations between 40-60 kilometers per hour, and the steering angle had multiple sudden small adjustments, and the frequency of braking operations was much higher than other time periods. Combining these factors, the server determined that the state descriptor feature corresponding to this time period was the target description feature. This target description feature can highlight the unstable driving state of the vehicle at a certain time during the driving process, and this unstable state may indicate a higher risk of danger in slippery road conditions, providing a key analysis basis for subsequent dangerous area analysis and driving safety warnings.

[0049] After determining the target description feature (i.e., the state description sub-feature corresponding to the 6th to 9th minute), the forward state description sub-feature located before the target description feature is determined based on the state node order between the multiple state description sub-features. Since the target description feature is the state description sub-feature of the 6th to 9th minute, the state description sub-feature corresponding to the 3rd to 6th minute before is the forward state description sub-feature.

[0050] Next, multiple forward slippery road condition data are obtained from the forward dangerous area change sequence corresponding to these forward state description sub-features. In the previous step, the server has constructed a dangerous area change sequence, which contains the slippery road related information and the changes in the dangerous area in different time periods. For the forward dangerous area change sequence corresponding to the forward state description sub-feature (3rd-6th minute), the server obtains multiple forward slippery road condition data from it. For example, the rainfall data of a specific section of the Panshan Highway in the 3rd-6th minute was obtained from the dangerous area change sequence, and it was found that the rainfall in this time period increased from 3 mm per hour to 4 mm per hour; at the same time, the road friction coefficient data was obtained, showing that the friction coefficient decreased from 0.4 to 0.35; the water depth data was also obtained, and the water depth increased from 1.5 cm to 2 cm. These forward slippery road condition data reflect the trend of the slippery road condition before the target description feature, which is of great significance for analyzing the relationship between the vehicle driving state corresponding to the target description feature and the road dangerous area. By combining these forward slippery road condition data, target description features and other relevant information, the server can more comprehensively infer the dangerous condition of the vehicle on slippery roads, thereby providing support for accurate driving safety warnings.

[0051] In a possible implementation, step S140 includes:

[0052] Step S141, obtaining interference features to be injected, setting the inference unit length, and setting the number of loop rounds.

[0053] Step S142, based on the interference feature to be injected, the set inference unit length, the dangerous area information, the target description feature, and the forward slippery road condition data, the driving safety inference network is used to infer the initial dangerous area change sequence corresponding to the first loop inference stage.

[0054] Step S143, performing interference enhancement processing on the initial dangerous area change sequence to generate interference enhancement data, and adjusting the set reasoning unit length to the reasoning unit length corresponding to the current loop reasoning stage.

[0055] Step S144, based on the interference enhancement data, the reasoning unit length corresponding to the current loop reasoning stage, the dangerous area information, the target description features, and the forward slippery road condition data, the driving safety reasoning network is used to infer the dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data according to the loop reasoning stage with the set number of cycles.

[0056] In a possible implementation, step S144 includes:

[0057] Step S1441, using the driving safety reasoning network to initialize the reasoning state, the reasoning state includes all the intermediate variables and parameters required in the reasoning process.

[0058] Step S1442, according to the reasoning unit length corresponding to the current loop reasoning stage, the reasoning data of the initial reasoning unit is extracted from the interference enhancement data, and the reasoning data contains the change information of the road danger area within the reasoning unit length, and the target description features and forward slippery road condition data in the corresponding time period are extracted at the same time.

[0059] Step S1443, using the driving safety reasoning network, combined with the reasoning data in the initial reasoning unit, the initialized reasoning state, and the dangerous area information, execute the first round of reasoning, update the reasoning state during the reasoning process, and generate the first reasoning result, which describes the changing trend of the road dangerous area in the current reasoning unit.

[0060] Step S1444, according to the set inference unit length, moves to the next inference unit, and extracts the interference enhancement data, target description features and forward slippery road condition data in the next inference unit, and updates the inference state at the same time, and outputs the updated inference unit data and inference state.

[0061] Repeat the above steps and continuously update the reasoning unit data and reasoning status. Each reasoning result will be generated based on the previous reasoning result and the data of the current reasoning unit, so as to gradually build a complete dangerous area change sequence corresponding to the current cycle reasoning stage. The dangerous area change sequence includes all the reasoning results from the initial reasoning unit to the last reasoning unit, and describes the changes in the road dangerous area within the entire set time interval.

[0062] Step S1445, evaluating the danger zone change sequence to determine whether there is a potential danger situation. If the evaluation result indicates that there is a potential danger situation, triggering a safety warning mechanism, the safety warning mechanism issues a warning to the driver in a variety of ways.

[0063] In this embodiment, when processing the task of driving safety analysis on slippery roads, it is necessary to first obtain the interference features to be injected, set the inference unit length, and set the number of cycles. In the current scenario, the interference feature to be injected is the sudden crosswind interference, which is a factor that may affect vehicle driving based on a comprehensive analysis of the meteorological data and road topography in the area. The inference unit length is set to 2 minutes. This time length is determined based on multiple factors such as the frequency of vehicle state changes on slippery roads and data processing efficiency. It can capture changes in vehicle states and dangerous areas more accurately without excessively increasing the amount of calculation. The number of cycles is set to 3. This value is determined based on previous experience in driving safety analysis on similar slippery roads and an estimate of the complexity of the current data. A sufficient number of cycles helps to more comprehensively and accurately infer the sequence of changes in dangerous areas.

[0064] Then, the interference features to be injected (crosswind interference), the set inference unit length (2 minutes), the previously obtained dangerous area information (such as the dangerous area conditions of the winding mountain road bends and downhill sections), the target description features (the vehicle's speed fluctuates greatly in a specific time period and the steering and braking operations are frequent) and the forward slippery road condition data (rainfall, friction coefficient, water depth and other data in the previous time period) are input into the driving safety inference network. For the first cycle reasoning stage, the driving safety inference network begins to analyze. For example, in the first 2 minutes, based on the dangerous area information at the bend of the winding mountain road (the bend radius is small and it is in the downhill section), combined with the speed fluctuation in the vehicle target description features (the speed fluctuates greatly between 40-60 kilometers per hour) and the forward slippery road condition data (the friction coefficient is 0.35 and the water depth is 2 cm), and considering the crosswind interference, the inference network predicts that the risk of vehicle skidding at the bend has increased from the original basic risk value (assumed to be 30%) to 40%. For downhill sections, due to frequent vehicle braking and slippery conditions, coupled with the influence of crosswind, the risk of rear-end collision increased from 25% to 35%. In this way, the reasoning network generated the initial dangerous area change sequence corresponding to the first cyclic reasoning stage, recording the changes in the danger level of the curve and downhill section within these 2 minutes.

[0065] Next, the initial sequence of dangerous area changes is subjected to interference enhancement. For example, for the side slip risk at the bend, according to the established interference enhancement algorithm, the side slip risk that originally rose to 40% is further enhanced to 45%, and the rear-end collision risk on the downhill section is enhanced from 35% to 40%. At the same time, according to the requirements of the current cyclic reasoning stage, the set reasoning unit length is adjusted from 2 minutes to 3 minutes. This adjustment is based on the need to consider the comprehensive impact of various factors over a longer period of time as the cyclic reasoning deepens, in order to more accurately infer changes in dangerous areas.

[0066] On this basis, the driving safety reasoning network initializes the reasoning state, which contains many intermediate variables and parameters. For example, it includes the basic danger level of each current dangerous area (curve, downhill section, etc.) (the initial risk value obtained based on the previous road condition and vehicle state analysis), the current basic driving state parameters of the vehicle (such as average speed, average steering angle, etc.), and the weight coefficients of previous data (dangerous area information, target description features, forward slippery road condition data, etc.). These intermediate variables and parameters will be continuously updated and adjusted in the subsequent reasoning process.

[0067] Since the inference unit length corresponding to the current loop inference stage is 3 minutes, the server extracts the inference data of the initial inference unit within these 3 minutes from the interference enhancement data. For example, during these 3 minutes, the risk of skidding at bends has continued to rise from 45%, possibly reaching 50%, and the risk of rear-end collision on downhill sections has increased from 40% to 45%. At the same time, the target description features corresponding to these 3 minutes were extracted, and it was found that the vehicle speed fluctuations were still large, and the steering operation, although the frequency was reduced, increased in amplitude. From the previous forward slippery road condition data, it was obtained that the rainfall did not change significantly during these 3 minutes, but the depth of water accumulation increased to 2.2 cm due to factors such as splashing caused by vehicle driving.

[0068] The driving safety reasoning network combines the reasoning data in the initial reasoning unit (risk changes in curves and downhill sections, changes in water depth, etc.), the initialized reasoning state (basic risk values ​​of each dangerous area, basic vehicle driving state parameters, etc.) and dangerous area information (geographical features of curves and downhill sections, etc.) to perform the first round of reasoning. During the reasoning process, for example, considering the large fluctuations in vehicle speed and the increase in steering amplitude, as well as the increase in water depth, the risk of skidding at the curve was re-evaluated and considered to have risen to 55% within these 3 minutes. For the downhill section, since the vehicle speed fluctuated but the overall speed did not show a significant downward trend, the rear-end collision risk rose to 50%. In this process, the reasoning network updates the reasoning state, such as adjusting the weight coefficient of the impact of vehicle speed fluctuations on the risk value of the dangerous area. The first reasoning result generated accurately describes the trend of the change in the degree of danger of the dangerous areas of the road (curves and downhill sections) within these 3 minutes.

[0069] According to the set 3-minute reasoning unit length, the server moves to the next reasoning unit. Interference enhancement data is extracted from the next 3-minute reasoning unit, and it is found that the risk of skidding at the curve has a slight downward trend due to some adjustment measures taken by the vehicle (analyzed from the target description features, such as the reduction of the speed fluctuation range), which may drop to 52%. The risk of rear-end collision on the downhill section is maintained at around 50% due to the good following distance of the subsequent vehicle (judged from the target description features). At the same time, the reasoning state is updated, for example, the relevant parameters for the risk assessment of the dangerous area are adjusted according to the new vehicle driving state, and the updated reasoning unit data (including risk change data of curves and downhill sections, etc.) and reasoning state (adjusted intermediate variables and parameters) are output.

[0070] Repeat the above steps continuously, and continuously update the reasoning unit data and reasoning status during the entire 3-round loop reasoning process. For example, in subsequent reasoning units, the risk value of the dangerous area may fluctuate continuously due to factors such as slight changes in weather conditions (information such as a slight drop in humidity is obtained from more comprehensive forward slippery road condition data) and adjustments in vehicle driving habits (continuous analysis from target description features). Through the generation method of each reasoning result based on the previous reasoning result and the current unit data, a complete dangerous area change sequence is gradually constructed. This sequence describes in detail how the degree of danger of dangerous areas on the road (curves and downhill sections) changes within the entire set 9 minutes (3 3-minute reasoning units), such as complex changes such as the risk of sideslip at the curve first increases, then decreases, and then increases.

[0071] Finally, the server evaluates the constructed sequence of changes in dangerous areas. For example, it was found that the risk of sideslipping on a curve reached 60% at a certain moment, and the risk of rear-end collision on a downhill section was close to 55%. According to the pre-set danger threshold (the risk of sideslipping on a curve exceeding 50% and the risk of rear-end collision on a downhill section exceeding 50% are considered potential dangerous situations), it is determined that there is a potential dangerous situation. The safety warning mechanism is then triggered, and a warning is issued to the driver through the vehicle's on-board system. For example, text prompts such as "The risk of sideslipping on a curve is extremely high, please slow down" and "Pay attention to the following distance, the risk of rear-end collision on a downhill section is high" are displayed on the dashboard, and the corresponding warning voice is played through the in-car voice system to remind the driver to take corresponding safety measures.

[0072] In a possible implementation, after step S140, a safety prompt operation may be performed according to the dangerous area change sequence to generate target driving safety prompt information.

[0073] After completing the reasoning of the dangerous area change sequence and evaluating the potential dangerous situation, the safety prompt operation is performed according to the dangerous area change sequence to generate the target driving safety prompt information. In addition to the immediate warning information triggered previously, more detailed and targeted driving safety prompt information will be further generated according to the entire dangerous area change sequence. For example, if the dangerous area change sequence shows that the risk of skidding at the curve remains at a high level for a period of time, the target driving safety prompt information generated by the server may include "Continuous curves ahead, please reduce the speed to less than 30 kilometers per hour in advance, avoid sudden braking and sharp turns". If the risk of rear-end collision on a downhill section suddenly increases at a certain moment, the prompt information may be "On a downhill section, please keep a safe distance of at least 50 meters from the vehicle in front and pay attention to the brake lights of the vehicle in front". These target driving safety prompt information can provide drivers with more specific and effective driving safety guidance, helping drivers better deal with potential dangers on slippery roads.

[0074] In a possible implementation, the method further includes:

[0075] Step A110, obtaining a training slippery road condition data stream, and extracting a training slippery road condition data segment from the training slippery road condition data stream according to the set time interval, wherein the training slippery road condition data segment includes a plurality of sample slippery road condition data and sample vehicle sensor data.

[0076] Step A120, inferring the road dangerous areas corresponding to the roads in the plurality of sample slippery road condition data through the road dangerous area prediction network, and generating sample dangerous area data.

[0077] Step A130, determining a sample reasoning unit length, and performing interference enhancement processing on the sample dangerous area data according to the sample reasoning unit length to generate interference enhanced dangerous area data.

[0078] Step A140, based on the interference enhanced dangerous area data, the sample inference unit length, the sample vehicle sensor data, the forward sample slippery road condition data corresponding to the training slippery road condition data segment, and the reference sample slippery road condition data, generate target inference dangerous area data through basic driving safety inference network inference.

[0079] Step A150, optimizing the basic driving safety reasoning network according to the target inference dangerous area data and the sample dangerous area data, and generating a driving safety reasoning network.

[0080] In a possible implementation manner, before step A140, the method further includes:

[0081] The forward sample slippery road condition data corresponding to the training slippery road condition data segment is randomly extracted, and the forward sample slippery road condition data corresponding to the training slippery road condition data segment is adjusted according to the result of the random extraction.

[0082] Alternatively, the forward sample slippery road condition data, the reference sample slippery road condition data, and the sample state description sub-features corresponding to the sample vehicle sensor data are jointly randomly extracted, and the forward sample slippery road condition data, the reference sample slippery road condition data, and the sample state description sub-features corresponding to the sample vehicle sensor data are adjusted according to the result of the joint random extraction.

[0083] In this embodiment, the training slippery road condition data stream collects information from multiple different data sources, such as meteorological sensors, road sensors, and various sensor data related to vehicles distributed on slippery roads in different regions. The server extracts training slippery road condition data segments from this huge training slippery road condition data stream according to a preset time interval, such as a time interval set to 15 minutes.

[0084] In this 15-minute training wet road condition data segment, multiple sample wet road condition data and sample vehicle sensor data are included. The sample wet road condition data may cover information such as rainfall, road temperature, humidity, water depth, and road friction coefficient. For example, the sample wet road condition data at a certain moment shows that the rainfall is 3 mm per hour, the road temperature is 5 degrees Celsius, the humidity is 80%, the water depth is 1.5 cm, and the friction coefficient is 0.4. The sample vehicle sensor data contains information such as the vehicle's speed, acceleration, steering angle, and braking operation. For example, the sample vehicle sensor data of a vehicle in this 15-minute period shows that its speed fluctuates between 40-60 kilometers per hour, the acceleration has multiple positive and negative alternations, the steering angle has several small adjustments, and the braking operation is relatively frequent. These data together constitute the training wet road condition data segment, which provides a basic data source for subsequent network training and optimization.

[0085] The server inputs multiple sample slippery road condition data extracted from the training slippery road condition data fragment into the road condition dangerous area prediction network. This network has been preliminarily trained and can analyze the slippery road condition data to determine the corresponding dangerous area of ​​the road.

[0086] For example, for the previously mentioned sample slippery road condition data (information such as rainfall, road temperature, humidity, water depth, and road friction coefficient), the road condition dangerous area prediction network combines its own stored road geographic information (such as curve location, slope, etc.) for analysis. If a certain road section is a curve and the water depth is deep and the friction coefficient is low, the network will determine that this curve area is a dangerous area. If there is a downhill section and the vehicle braking operation is frequent (combined with the braking information in the sample vehicle sensor data), and considering the slippery road factor, this downhill section will also be marked as a dangerous area. The road condition dangerous area prediction network will organize the location information of these dangerous areas (for example, accurately indicating the location of curves and downhill sections through longitude and latitude coordinates), the degree of danger (based on a comprehensive evaluation of multiple factors, such as water depth and friction coefficient, the degree of danger may be quantified as a value between 0-1, such as 0.6 for a higher degree of danger) and other information to generate sample dangerous area data. These sample dangerous area data accurately describe the dangerous areas and their degree of danger on the road under these sample slippery road conditions, providing a key basis for subsequent further processing.

[0087] The server determines the sample reasoning unit length, which is determined based on the analysis of the training data and the frequency of changes in the driving conditions on slippery roads. Assume that the determined sample reasoning unit length is 3 minutes.

[0088] Based on this 3-minute sample reasoning unit length, the server performs interference enhancement processing on the sample dangerous area data. This interference enhancement processing is to simulate various complex situations and uncertain factors that may occur during actual driving, so as to make the training data more representative and comprehensive. For example, for a certain curved dangerous area, the original danger level in the sample dangerous area data is 0.6. According to the interference enhancement algorithm, the server takes into account the possible sudden crosswind interference (this is a common interference factor that affects the stability of the vehicle at the curve in actual driving) and enhances the danger level to 0.7. For the dangerous area on the downhill section, due to the possible change in vehicle load (although it may not be explicitly included in the sample vehicle sensor data, it is a possible interference factor in actual conditions), the danger level is enhanced from the original 0.5 to 0.6. In this way, the sample dangerous area data within each sample reasoning unit length is processed to generate interference enhanced dangerous area data. These interference enhanced dangerous area data can better reflect the situation of road dangerous areas in complex and changeable actual driving environments, which helps to improve the generalization ability of the basic driving safety reasoning network.

[0089] Before generating target reasoning dangerous area data through basic driving safety reasoning network reasoning, some data adjustment operations are required.

[0090] The server randomly extracts the forward sample slippery road condition data corresponding to the training slippery road condition data segment. For example, in a 15-minute training slippery road condition data segment, the forward sample slippery road condition data before each 3-minute sample inference unit length is randomly extracted. Assuming that the forward sample slippery road condition data contains historical data for several consecutive hours, the server randomly extracts a part of the data points. For example, there may be 100 data points representing the previous rainfall data, and 30 data points are randomly extracted. Then, based on the results of this random extraction, the forward sample slippery road condition data corresponding to the training slippery road condition data segment is adjusted. This may include operations such as recalculating the average value and adjusting the time series weight of the data. Through this random extraction and adjustment, the diversity of the data is increased, and the basic driving safety reasoning network is prevented from being overly dependent on specific data patterns.

[0091] Another way is that the server performs joint random extraction of sample state description sub-features corresponding to the forward sample slippery road condition data, the reference sample slippery road condition data, and the sample vehicle sensor data. For example, the rainfall, road temperature and other data in the forward sample slippery road condition data, the reference sample slippery road condition data (which may be data of other similar sections or time periods for comparison and reference), and the sample state description sub-features corresponding to the sample vehicle sensor data (such as speed fluctuation, steering angle change, etc.) are regarded as a whole data set. A certain proportion of data is randomly extracted from this whole data set. Then, based on the results of the joint random extraction, the sample state description sub-features corresponding to the forward sample slippery road condition data, the reference sample slippery road condition data, and the sample vehicle sensor data are adjusted. For example, the extracted data is re-normalized, and the association weights between different data are adjusted to improve the randomness and representativeness of the data, so that the basic driving safety reasoning network can better cope with various different input data combinations.

[0092] After completing the above data adjustment operation, the server starts to generate target reasoning danger zone data through the basic driving safety reasoning network.

[0093] First, the server aligns multiple time-continuous sample slippery road condition data and corresponding sample vehicle sensor data in the training slippery road condition data segment in chronological order. For example, the sample slippery road condition data (rainfall, road surface temperature, etc.) and sample vehicle sensor data (speed, acceleration, etc.) within each 3-minute sample inference unit length are accurately aligned according to the timestamp to ensure that the road condition data and vehicle sensor data at the same time can correspond to each other, and the aligned sample data set is output.

[0094] Then, the road condition dangerous area prediction network is used to process the sample slippery road condition data in the aligned sample data set to extract the dangerous area information corresponding to the road at each time point and obtain the sample dangerous area data set. In this sample dangerous area data set, each data point contains a timestamp and the corresponding dangerous area information (such as the dangerous area location and danger level mentioned above).

[0095] Next, according to the preset interference feature generation rules, the interference enhancement processing is performed on the dangerous area information in the sample dangerous area data set, and the interference enhanced dangerous area data set is output. For example, according to the set interference type (such as sudden increase in wind speed interference, sudden vehicle failure interference, etc.) and intensity parameters (such as wind speed size, the impact of the failure on vehicle operation, etc.), each dangerous area data is randomly disturbed or noise is added to generate interference enhanced dangerous area data. At the same time, according to the set sample reasoning unit length (3 minutes), the interference enhanced dangerous area data is divided into multiple reasoning units, each reasoning unit contains 3 minutes of time series data.

[0096] Afterwards, the driving state description network is used to process the sample vehicle sensor data in the aligned sample data set to extract driving state description features. Each data point in these driving state description features contains a timestamp and a corresponding state description feature (such as speed fluctuation, steering operation features, etc.).

[0097] Based on the aligned sample data set, the server extracts the forward sample slippery road condition data and the reference sample slippery road condition data at each time point.

[0098] Finally, the basic driving safety reasoning network is used to combine the current input sample dangerous area data set, driving state description features, forward sample slippery road condition data and reference sample slippery road condition data to perform the first round of reasoning and output the first reasoning result of the initial dangerous area change sequence. Then, based on this first reasoning result, the interference enhanced dangerous area data of the next reasoning unit and the corresponding sample state description sub-features, the next data unit of the forward sample slippery road condition data, and the next data unit of the reference sample slippery road condition data, according to the set reasoning unit length (3 minutes), the data of the current reasoning unit is updated, and the next round of reasoning is combined with the updated data to output the next reasoning result of the dangerous area change sequence, so as to gradually build a complete dangerous area change sequence, complete the reasoning of the entire time series, and finally output the target reasoning dangerous area data.

[0099] The server constructs the target training cost parameter based on the error between the target reasoning dangerous area data and the sample dangerous area data. For example, if the predicted danger level of a curve dangerous area in the target reasoning dangerous area data is 0.7, and the actual danger level of the curve in the sample dangerous area data is 0.65, then the difference between the two is calculated as part of the error. All such errors are combined to construct the target training cost parameter. This target training cost parameter reflects the degree of deviation between the current prediction result of the basic driving safety reasoning network and the actual result.

[0100] Then, the basic driving safety inference network is optimized based on this training cost parameter. For example, by adjusting the weight coefficients and neuron connection parameters in the network, the basic driving safety inference network can reduce the prediction error in subsequent inferences. After multiple such optimization processes, the basic driving safety inference network continuously adjusts its own parameters and finally generates a driving safety inference network. This driving safety inference network can more accurately infer the changes in dangerous areas of slippery road condition data under vehicle sensor data, thereby providing more reliable support for actual slippery road driving safety warnings.

[0101] In a possible implementation, step A140 includes:

[0102] Step A141, aligning a plurality of time-continuous sample slippery road condition data and corresponding sample vehicle sensor data in the training slippery road condition data segment in time order, and outputting the aligned sample data set.

[0103] Step A142, using the road condition danger zone prediction network to process the sample slippery road condition data in the aligned sample data set to extract the danger zone information corresponding to the road at each time point, and obtain a sample danger zone data set, wherein each data point in the sample danger zone data set contains a timestamp and corresponding danger zone information.

[0104] Step A143, according to the preset interference feature generation rules, interference enhancement processing is performed on the dangerous area information in the sample dangerous area data set, and an interference enhanced dangerous area data set is output, wherein, according to the set interference type and intensity parameters, each dangerous area data is randomly disturbed or noise is added to generate interference enhanced dangerous area data, and at the same time, according to the set sample inference unit length, the interference enhanced dangerous area data is divided into multiple inference units, each inference unit contains time series data of the sample inference unit length.

[0105] Step A144, using the driving state description network to process the sample vehicle sensor data in the aligned sample data set to extract driving state description features, each data point in the driving state description features includes a timestamp and a corresponding state description feature.

[0106] Step A145, based on the aligned sample data set, extracting the forward sample slippery road condition data and the reference sample slippery road condition data at each time point.

[0107] Step A146, using the basic driving safety reasoning network, combined with the currently input sample dangerous area data set, driving state description features, forward sample slippery road condition data and reference sample slippery road condition data, perform the first round of reasoning and output the first reasoning result of the initial dangerous area change sequence.

[0108] Step A147, based on the first reasoning result, the interference enhanced dangerous area data of the next reasoning unit and the corresponding sample state description features, the next data unit of the forward sample slippery road condition data, and the next data unit of the reference sample slippery road condition data, update the data of the current reasoning unit according to the set reasoning unit length, and perform the next round of reasoning in combination with the updated data, output the next reasoning result of the dangerous area change sequence, thereby gradually constructing a complete dangerous area change sequence, completing the reasoning of the entire time series, and finally outputting it as the target reasoning dangerous area data.

[0109] In a possible implementation, step A150 includes:

[0110] Step A151, constructing a target training cost parameter according to the error between the target inference dangerous area data and the sample dangerous area data.

[0111] Step A152: Optimize the basic driving safety reasoning network according to the training cost parameter to generate a driving safety reasoning network.

[0112] In this embodiment, first, the server aligns multiple time-continuous sample slippery road condition data and corresponding sample vehicle sensor data in the training slippery road condition data segment in time order, and outputs the aligned sample data set. In this process, the server will obtain a large number of sample slippery road condition data and sample vehicle sensor data contained in the training slippery road condition data segment. The sample slippery road condition data contains information such as rainfall, road surface temperature, humidity, water depth, and road friction coefficient, which are collected at certain time intervals. The sample vehicle sensor data contains information such as vehicle speed, acceleration, steering angle, braking operation, etc., and is also a sequence data with time stamps. The server will accurately match the sample slippery road condition data and the corresponding sample vehicle sensor data at each moment according to the timestamp, for example, combining the road condition data such as rainfall and road surface temperature at a certain moment with the sensor data such as vehicle speed and steering angle at the same moment to form a complete data pair. By processing all time-continuous data in this way, the server outputs an aligned sample data set. This data set ensures that in subsequent analysis, the road condition information and vehicle status information at each time point can be accurately matched, providing a basic guarantee for accurate reasoning about dangerous areas.

[0113] Next, the server uses the road condition dangerous area prediction network to process the sample slippery road condition data in the aligned sample data set to extract the dangerous area information corresponding to the road at each time point and obtain the sample dangerous area data set. During the processing, the road condition dangerous area prediction network will analyze the sample slippery road condition data at each time point in combination with its internal pre-set rules and algorithms. For example, for the sample slippery road condition data at a certain moment, where the rainfall is 3 mm per hour, the road surface temperature is 5 degrees Celsius, the humidity is 80%, the water depth is 1.5 cm, and the friction coefficient is 0.4, the network will make a comprehensive judgment based on the geographical information of the road (such as the location of the curve, the slope, etc.). If at a certain curve, due to the depth of water accumulation and the lower friction coefficient, the network will determine that the curve is a dangerous area and determine its degree of danger. The determination of this degree of danger may be based on a predefined evaluation model that comprehensively considers the weights of various factors. In this sample dangerous area dataset, each data point contains a timestamp and corresponding dangerous area information. For example, the dangerous area information corresponding to a timestamp may be "curve (longitude and latitude coordinates), danger level 0.6". Such a dataset completely records the dangerous area conditions of the road at each time point.

[0114] Then, the server performs interference enhancement processing on the dangerous area information in the sample dangerous area data set according to the preset interference feature generation rules, and outputs the interference enhanced dangerous area data set. In this process, the server operates according to the set interference type and intensity parameters. Interference types may include sudden meteorological change interference (such as strong winds, sudden increase in rainstorm intensity), temporary road condition interference (such as small potholes or scattered objects on the road), etc. The intensity parameter is a numerical value that quantifies the degree of impact of these interferences on the dangerous area. For example, for the previously determined dangerous area of ​​the curve, according to the set interference type of strong wind interference and the intensity parameter set to medium intensity, the server will randomly disturb or add noise to its danger level data. Assuming that the original curve danger level is 0.6, it may become 0.7 after interference enhancement processing. At the same time, according to the set sample reasoning unit length (assuming 3 minutes), the interference enhanced dangerous area data is divided into multiple reasoning units. Each reasoning unit contains 3 minutes of time series data. Such a division helps to analyze step by step according to the time unit in the subsequent reasoning process, which is more in line with the consideration of the changes in dangerous areas in different time periods during actual driving.

[0115] Afterwards, the server uses the driving state description network to process the sample vehicle sensor data in the aligned sample data set to extract driving state description features. The driving state description network will conduct an in-depth analysis of the sample vehicle sensor data at each time point. For example, at a certain moment, the vehicle speed is 50 kilometers per hour, the acceleration is -1 m / s2, the steering angle is 10 degrees, and the braking operation is in a light-stepping state. The driving state description network will combine this information to extract the corresponding driving state description features. This feature may include relevant information such as the vehicle's driving stability and the frequency of operation, and each data point contains a timestamp and a corresponding state description feature, such as "at a certain timestamp, the vehicle's driving stability is medium and the frequency of operation is low." These driving state description features can accurately reflect the driving status of the vehicle at different time points, and provide important basis for subsequent reasoning with dangerous area information.

[0116] Based on the aligned sample data set, the server extracts the forward sample slippery road condition data and reference sample slippery road condition data at each time point. The forward sample slippery road condition data refers to the road condition data before the current time point, such as the sample slippery road condition data within 10 minutes before the current 3-minute reasoning unit. These data can reflect the development trend of the previous slippery road conditions. The reference sample slippery road condition data refers to other related sample data, which may be road condition data from similar road sections or under similar weather conditions, and is used to provide additional reference information. The server accurately extracts this data from the aligned sample data set through specific algorithms and rules to provide more comprehensive input information for the basic driving safety reasoning network.

[0117] Finally, the server uses the basic driving safety reasoning network to combine the currently input sample dangerous area data set, driving state description features, forward sample slippery road condition data and reference sample slippery road condition data to perform the first round of reasoning and output the first reasoning result of the initial dangerous area change sequence. In the first round of reasoning, the basic driving safety reasoning network will comprehensively consider all input data. For example, combined with the sample dangerous area data set (including the location and danger level of each dangerous area), driving state description features (vehicle driving state), forward sample slippery road condition data (previous road condition development trend) and reference sample slippery road condition data (reference for similar situations) in the current 3-minute reasoning unit, the network will calculate according to its internal complex algorithm structure. Assume that in this reasoning unit, after combining these data, the danger level of a certain curve dangerous area is calculated by the network, and the result is that the danger level rises from 0.7 to 0.8 within this 3 minutes. This is the first reasoning result of the initial dangerous area change sequence, which describes the change trend of the dangerous area in this specific reasoning unit.

[0118] Then, based on the first reasoning result, the interference enhanced dangerous area data of the next reasoning unit and the corresponding sample state description features, the next data unit of the forward sample slippery road condition data, and the next data unit of the reference sample slippery road condition data, the data of the current reasoning unit is updated according to the set reasoning unit length, and the next round of reasoning is performed in combination with the updated data, and the next reasoning result of the dangerous area change sequence is output, so as to gradually build a complete dangerous area change sequence, complete the reasoning of the entire time series, and finally output the target reasoning dangerous area data. For example, when entering the next 3-minute reasoning unit, the server will obtain the interference enhanced dangerous area data of the next reasoning unit (assuming that the danger level of the curve dangerous area has changed to 0.85), the corresponding sample state description features (some changes have occurred in the vehicle driving state, such as increased speed fluctuations), the next data unit of the forward sample slippery road condition data (the previous road conditions have developed to the situation at this moment, such as a slight increase in rainfall) and the next data unit of the reference sample slippery road condition data (the reference situation of a similar road section at this moment). According to the set reasoning unit length (3 minutes), the server will update the data of the current reasoning unit, combine the previous reasoning results (the change in the danger level in the first reasoning result) with the newly acquired data, and then perform the next round of reasoning through the basic driving safety reasoning network. Assume that after this round of reasoning, it is concluded that the danger level of the curve danger zone rises to 0.9 within these 3 minutes. This is the next reasoning result of the danger zone change sequence. By repeating this process, data is updated and reasoned for each reasoning unit, and a complete danger zone change sequence is gradually constructed. This sequence fully describes the changes in the danger zone in the entire time series, and finally this complete danger zone change sequence is output as the target reasoning danger zone data.

[0119] When optimizing the basic driving safety reasoning network based on the target reasoning dangerous area data and the sample dangerous area data, first, the target training cost parameter is constructed based on the error between the target reasoning dangerous area data and the sample dangerous area data. In this process, the server will conduct a detailed comparative analysis of the target reasoning dangerous area data and the sample dangerous area data. For example, the dangerous area data at each time point or each reasoning unit are compared. Assuming that at a certain time point, the danger level of a certain curve in the sample dangerous area data is 0.75, and the danger level of the curve in the target reasoning dangerous area data is 0.8, then the server will calculate the difference between the two (0.8-0.75=0.05) as part of the error. For all dangerous area data points or reasoning units, the errors are calculated in this way, and these errors are combined to construct the target training cost parameter. This target training cost parameter can comprehensively reflect the degree of deviation between the predicted results and the actual results of the basic driving safety reasoning network in the current reasoning process, and it is an important indicator for measuring network performance.

[0120] Then, the basic driving safety reasoning network is optimized according to this training cost parameter to generate a driving safety reasoning network. During the optimization process, the server adjusts the internal parameters of the basic driving safety reasoning network according to the target training cost parameter. For example, the basic driving safety reasoning network may contain multiple neuron layers, and the connection weights between neurons are one of the key factors affecting the network output results. If the target training cost parameter shows a large error, the server will adjust these weights through a specific optimization algorithm (such as a gradient descent algorithm). Suppose the weight of a neuron connection is currently 0.5, and after calculation, it is found that reducing this weight can reduce the error in the target training cost parameter, then the server will adjust this weight to a smaller value (such as 0.4). At the same time, in addition to weight adjustment, the optimization of other network parameters may also be involved, such as the activation function parameters of neurons. By gradually adjusting these parameters, the basic driving safety reasoning network continuously optimizes its own performance, so that in the subsequent reasoning process, it can more accurately infer the changes in dangerous areas that are closer to the actual situation based on the input data (such as sample dangerous area data sets, driving state description features, forward sample slippery road condition data and reference sample slippery road condition data, etc.), and finally generate a driving safety reasoning network. This network can provide more reliable and accurate dangerous area reasoning results in application scenarios such as slippery road driving safety warning.

[0121] Figure 2 The hardware structure of the AI-based slippery road driving safety warning system 100 for implementing the above-mentioned AI-based slippery road driving safety warning method provided by an embodiment of the present invention is shown, as Figure 2As shown, the AI-based slippery road driving safety warning system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0122] 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 AI-based slippery road driving safety warning system 100 uses to execute or use to complete the exemplary method described in the present invention.

[0123] 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 AI-based slippery road driving safety warning method as described in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected via the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0124] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned AI-based slippery road driving safety warning system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0125] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are stored. When a processor executes the computer executable instructions, the above-mentioned AI-based slippery road driving safety warning method is implemented.

[0126] 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. An AI-based early warning method for driving safety on slippery roads, characterized in that: The method comprises: Acquire candidate slippery road condition data, and extract dangerous area information corresponding to the road in the candidate slippery road condition data through a road condition dangerous area prediction network; Acquire vehicle sensor data, and extract driving state description features of the vehicle sensor data through a driving state description network; Extracting a target description feature from the driving state description feature according to a set time interval, and acquiring a plurality of forward slippery road condition data corresponding to the target description feature; According to the interference features to be injected, the inference unit length corresponding to the loop reasoning stage, the dangerous area information, the target description features, and the forward slippery road condition data, a driving safety inference network is used to infer the dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data according to at least two rounds of loop reasoning stages, so as to perform a slippery road driving safety warning based on the dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data.

2. The AI-based slippery road driving safety warning method according to claim 1 is characterized in that: According to the set time interval, the target description feature is extracted from the driving state description feature, including: According to a set time interval, the driving state description feature is decomposed into a plurality of state description sub-features associated in time and space domains; A target description feature is determined from the plurality of state description sub-features.

3. The AI-based slippery road driving safety warning method according to claim 2 is characterized in that: Acquiring a plurality of forward slippery road condition data corresponding to the target description feature, including: Determine a forward state description sub-feature located before the target description feature based on the state node order between the plurality of state description sub-features; A plurality of forward slippery road condition data are obtained from the forward dangerous area change sequence corresponding to the forward state description sub-feature.

4. The AI-based slippery road driving safety warning method according to claim 1 is characterized in that: Based on the interference feature to be injected, the inference unit length corresponding to the loop inference stage, the dangerous area information, the target description feature, and the forward slippery road condition data, a dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data is inferred by using a driving safety inference network in at least two rounds of loop inference stages, including: Obtain the interference features to be injected, set the inference unit length, and set the number of loop rounds; Based on the interference feature to be injected, the set inference unit length, the dangerous area information, the target description feature, and the forward slippery road condition data, the driving safety inference network is used to infer the initial dangerous area change sequence corresponding to the first loop inference stage; Performing interference enhancement processing on the initial dangerous area change sequence to generate interference enhancement data, and adjusting the set reasoning unit length to the reasoning unit length corresponding to the current loop reasoning stage; Based on the interference enhancement data, the inference unit length corresponding to the current loop inference stage, the dangerous area information, the target description feature, and the forward slippery road condition data, the driving safety inference network is used to infer the dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data according to the loop inference stage of the set number of loops; The step of using the driving safety reasoning network to infer the dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data according to the set number of loop reasoning stages according to the interference enhancement data, the reasoning unit length corresponding to the current loop reasoning stage, the dangerous area information, the target description features, and the forward slippery road condition data includes: Initializing a reasoning state using the driving safety reasoning network, the reasoning state including all intermediate variables and parameters required in the reasoning process; According to the inference unit length corresponding to the current loop inference stage, extracting the inference data of the initial inference unit from the interference enhancement data, the inference data including the change information of the road danger area within the inference unit length, and extracting the target description features and forward slippery road condition data within the corresponding time period; Using the driving safety reasoning network, combining the reasoning data in the initial reasoning unit, the initialized reasoning state, and the dangerous area information, performing a first round of reasoning, updating the reasoning state during the reasoning process, and generating a first reasoning result, the first reasoning result describing a change trend of the road dangerous area in the current reasoning unit; According to the set inference unit length, move to the next inference unit, extract the interference enhancement data, target description features and forward slippery road condition data in the next inference unit, update the inference state at the same time, and output the updated inference unit data and inference state; Repeat the above steps and continuously update the reasoning unit data and reasoning state. Each reasoning result is generated based on the previous reasoning result and the data of the current reasoning unit, so as to gradually construct a complete dangerous area change sequence corresponding to the current cycle reasoning stage; the dangerous area change sequence includes all the reasoning results from the initial reasoning unit to the last reasoning unit, and describes the changes of the road dangerous area within the entire set time interval; The danger zone change sequence is evaluated to determine whether there is a potential danger situation. If the evaluation result indicates that there is a potential danger situation, a safety warning mechanism is triggered, and the safety warning mechanism warns the driver in a variety of ways.

5. The AI-based slippery road driving safety warning method according to claim 1 is characterized in that: After the dangerous area change sequence of the candidate slippery road condition data under the vehicle sensor data is inferred by using a driving safety inference network in at least two rounds of loop reasoning stages according to the interference feature to be injected, the inference unit length corresponding to the loop reasoning stage, the dangerous area information, the target description feature, and the forward slippery road condition data, the method further includes: A safety prompt operation is performed according to the dangerous area change sequence to generate target driving safety prompt information.

6. The AI-based slippery road driving safety warning method according to claim 1 is characterized in that: The method further comprises: Acquire a training slippery road condition data stream, and extract a training slippery road condition data segment from the training slippery road condition data stream according to the set time interval, wherein the training slippery road condition data segment includes a plurality of sample slippery road condition data and sample vehicle sensor data; Inferring the road danger area corresponding to the road in the plurality of sample slippery road condition data through the road condition danger area prediction network, and generating sample danger area data; Determine a sample reasoning unit length, and perform interference enhancement processing on the sample dangerous area data according to the sample reasoning unit length to generate interference enhanced dangerous area data; Generate target reasoning danger zone data through basic driving safety reasoning network reasoning according to the interference enhanced danger zone data, the sample reasoning unit length, the sample vehicle sensor data, the forward sample slippery road condition data corresponding to the training slippery road condition data segment, and the reference sample slippery road condition data; According to the target inference dangerous area data and the sample dangerous area data, the basic driving safety inference network is optimized to generate a driving safety inference network.

7. The AI-based slippery road driving safety warning method according to claim 6 is characterized in that: Before generating target inference danger zone data through inference through a basic driving safety inference network based on the interference enhanced danger zone data, the sample inference unit length, the sample vehicle sensor data, the forward sample slippery road condition data corresponding to the training slippery road condition data segment, and the reference sample slippery road condition data, the method further includes: Randomly extracting the forward sample slippery road condition data corresponding to the training slippery road condition data segment, and adjusting the forward sample slippery road condition data corresponding to the training slippery road condition data segment according to the random extraction result; Alternatively, the forward sample slippery road condition data, the reference sample slippery road condition data, and the sample state description sub-features corresponding to the sample vehicle sensor data are jointly randomly extracted, and the forward sample slippery road condition data, the reference sample slippery road condition data, and the sample state description sub-features corresponding to the sample vehicle sensor data are adjusted according to the result of the joint random extraction.

8. The AI-based slippery road driving safety warning method according to claim 6 is characterized in that: According to the interference enhanced dangerous area data, the sample reasoning unit length, the sample vehicle sensor data, the forward sample slippery road condition data corresponding to the training slippery road condition data segment, and the reference sample slippery road condition data, the target reasoning dangerous area data is generated by reasoning through the basic driving safety reasoning network, including: Aligning a plurality of time-continuous sample slippery road condition data and corresponding sample vehicle sensor data in the training slippery road condition data segment in time order, and outputting an aligned sample data set; Processing the sample slippery road condition data in the aligned sample data set using the road condition dangerous area prediction network to extract dangerous area information corresponding to the road at each time point, and obtaining a sample dangerous area data set, wherein each data point in the sample dangerous area data set includes a timestamp and corresponding dangerous area information; According to the preset interference feature generation rule, interference enhancement processing is performed on the dangerous area information in the sample dangerous area data set, and an interference enhanced dangerous area data set is output, wherein, according to the set interference type and intensity parameters, each dangerous area data is randomly disturbed or noise is added to generate interference enhanced dangerous area data, and at the same time, according to the set sample reasoning unit length, the interference enhanced dangerous area data is divided into a plurality of reasoning units, each reasoning unit contains time series data of the sample reasoning unit length; Processing the sample vehicle sensor data in the aligned sample data set using a driving state description network to extract driving state description features, wherein each data point in the driving state description features includes a timestamp and a corresponding state description feature; Based on the aligned sample data set, extracting forward sample slippery road condition data and reference sample slippery road condition data at each time point; Using the basic driving safety reasoning network, combined with the currently input sample dangerous area data set, driving state description features, forward sample slippery road condition data and reference sample slippery road condition data, a first round of reasoning is performed to output the first reasoning result of the initial dangerous area change sequence; Based on the first reasoning result, the interference enhanced dangerous area data of the next reasoning unit and the corresponding sample state description features, the next data unit of the forward sample slippery road condition data, and the next data unit of the reference sample slippery road condition data, the data of the current reasoning unit is updated according to the set reasoning unit length, and the next round of reasoning is performed in combination with the updated data to output the next reasoning result of the dangerous area change sequence, thereby gradually constructing a complete dangerous area change sequence, completing the reasoning of the entire time series, and finally outputting it as the target reasoning dangerous area data.

9. The AI-based slippery road driving safety warning method according to claim 6 is characterized in that: According to the target inference dangerous area data and the sample dangerous area data, the basic driving safety inference network is optimized to generate a driving safety inference network, including: constructing a target training cost parameter according to an error between the target inference dangerous area data and the sample dangerous area data; The basic driving safety reasoning network is optimized according to the training cost parameter to generate a driving safety reasoning network.

10. An AI-based slippery road driving safety warning system, characterized in that: The AI-based slippery road driving safety warning system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the AI-based slippery road driving safety warning method described in any one of claims 1 to 9 above.

Images