A vehicle weather environment adaptive method based on online learning

Abstract

This invention belongs to the field of vehicle intelligent control and discloses a vehicle weather environment adaptive method based on online learning. The method includes: extracting and quantizing information from preprocessed real-time environmental images and weightedly fusing them with historical environmental information to obtain multi-frame fused environmental information, which is then input into a weather type prediction model; dynamically adjusting the vehicle's driving coefficient, braking coefficient, and steering coefficient based on the weather impact level, and correspondingly outputting driving torque, braking torque, and steering angle adapted to the weather conditions; further limiting the driving torque and steering angle based on a pre-constructed dynamic threshold relationship mapping table; calculating the difference between the recognition result and the prediction result, inputting the difference data into an online gradient network, minimizing the difference as the loss, calculating the adjustment gradient of the model parameters, and updating the model parameters online. This significantly improves the model's adaptability to sudden weather changes while maintaining the overall prediction accuracy stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle control, and more particularly to a vehicle weather environment adaptive method based on online learning. Background Technology

[0002] Most existing environmental perception systems in autonomous driving and assisted driving technologies focus on road target recognition and distance detection, using road condition assessment to address complex road conditions or sudden accident risks. Currently, mass-produced vehicles typically require drivers to manually adjust driving modes, usually based on pre-set fixed parameters to cope with limited weather scenarios under preset conditions. This presents a vehicle control mode where "the driver adapts to the weather, and the vehicle passively adapts to the weather," making it difficult for drivers to accurately perceive the impact of weather changes on vehicle driving, easily leading to problems such as delayed control response and insufficient driving stability.

[0003] Due to the significant uncertainty and diversity of weather conditions, the single fixed parameters preset by vehicles are difficult to adapt to diverse weather changes. Although some vehicles have basic functions such as speed limits in rainy weather and lights on in foggy weather, they lack the ability to accurately identify and dynamically adapt to weather conditions. They cannot cover complex and ever-changing weather gradients, resulting in insufficient vehicle handling stability and extended braking distances in extreme weather conditions. They are unable to adapt to weather differences in different regions and at different times, which can easily lead to safety hazards. Summary of the Invention

[0004] The purpose of this invention is to provide a quantitative coupling and online optimization method to dynamically update vehicle control coefficients to adapt to driving needs under different weather conditions. Its core lies in adaptively coupling the current environmental information with the vehicle control system and training the prediction model in real time through an online feedback network.

[0005] In a first aspect, embodiments of the present invention provide a vehicle weather environment adaptive method based on online learning, comprising:

[0006] Real-time environmental images are acquired, and the preprocessed real-time environmental images are sequentially passed through an encoder, a latent space, and a decoder for information extraction and quantization processing to output real-time environmental information. The real-time environmental information is then weighted and fused with historical environmental information to obtain multi-frame fused environmental information.

[0007] The multi-frame fused environmental information is input into a pre-built weather type prediction model, which outputs weather identification results, weather impact level, and environmental information prediction results for the next moment; wherein, the weather impact level is a value obtained by mapping real-time environmental information and used to comprehensively evaluate the severity of the vehicle driving environment, and its range is from 0 to 1.

[0008] Based on the weather impact level, the vehicle's drive coefficient, braking coefficient, and steering coefficient are dynamically adjusted. Combined with the driver's input requests from the drive pedal, brake pedal, and steering wheel, the driving torque, braking torque, and steering angle adapted to the current weather conditions are calculated and output. Based on a pre-built dynamic threshold relationship mapping table between the output driving torque and the maximum vehicle speed, and a dynamic threshold relationship mapping table between the output steering angle and the weather impact level, the driving torque is adaptively limited to be less than the maximum driving torque, and the steering angle is adaptively limited to be less than the maximum steering angle.

[0009] Calculate the difference data between the multi-frame fused environmental information and the predicted environmental information at the next moment;

[0010] After fusing the differential data and cloud-based meteorological data, the data is input into an online gradient network. The online gradient network compares the current weather identification result with the current prediction result cached by the system, minimizes the difference between the two as the loss, calculates the adjustment gradient of the model parameters, and updates the parameters of the encoder, decoder, and weather type prediction model online based on the adjustment gradient.

[0011] Optionally, the preprocessed real-time environment image is sequentially passed through an encoder, a latent space, and a decoder for information extraction and quantization processing, and the real-time environment information is output.

[0012] The real-time environmental images acquired by the camera are preprocessed, including: filtering out noise caused by rain, snow and reflected light in the real-time environmental images based on image processing algorithms, while enhancing the details and edge information of the images; finding image regions related to weather judgment through region of interest extraction algorithms and making decisions, and outputting the preprocessed real-time environmental images;

[0013] The real-time environment image is input into the encoder, and the pixel matrix of the real-time environment image is converted into a high-dimensional feature vector. The high-dimensional feature vector is used to characterize the spatial features of the real-time environment image, and key environmental features are extracted from the high-dimensional feature vector in the latent space.

[0014] The key environmental features are reduced in dimensionality and normalized by a decoder, and then mapped and quantized into real-time environmental information. At the same time, the real-time environmental information is cached in the system as historical environmental information.

[0015] Optionally, the multi-frame fused environmental information is a first environmental information quantity obtained by weighted fusion of historical environmental information quantities from time (tk) to time t;

[0016] The calculation of the difference data between the multi-frame fused environmental information and the predicted environmental information at the next moment includes:

[0017] Among them, the multi-frame fused environmental information at time (t-1) is classified and predicted based on the weather type prediction model to obtain the environmental information prediction result at time t, which is used as the second environmental information quantity;

[0018] The error between the first environmental information quantity and the second environmental information quantity is calculated to obtain differential data.

[0019] Optionally, the differential data and the cloud-based meteorological data are fused and then input into an online gradient network. The online gradient network compares the current weather identification result with the current prediction result cached by the system, minimizes the difference between the two as the loss, calculates the adjustment gradient of the model parameters, and updates the parameters of the encoder, decoder, and weather type prediction model online based on the adjustment gradient.

[0020] The system pre-connects to a cloud-based meteorological service and obtains cloud-based meteorological data after passing through an environmental information network, wherein the environmental information network is used to extract and identify weather-related feature information;

[0021] The differential data and cloud meteorological data are weighted and fused based on the Kalman filter algorithm to filter out random noise in the differential data and obtain comprehensive data, wherein the weight of the cloud meteorological data is less than the weight of the differential data.

[0022] The online gradient network constructs a loss function based on the comprehensive data with the goal of improving the accuracy of weather forecasting, and calculates the gradient of the loss function with respect to the model parameters through the backpropagation algorithm to obtain the loss gradient; the model parameters include weight parameters and bias parameters.

[0023] Based on a preset learning rate, the loss gradient is backpropagated using a gradient descent algorithm to update the weights and bias parameters of the encoder, decoder, and weather type prediction model.

[0024] Optionally, based on the weather impact level, the vehicle's drive coefficient, braking coefficient, and steering coefficient are dynamically adjusted. Combined with the driver's input requests via the drive pedal, brake pedal, and steering wheel, the driving torque, braking torque, and steering angle adapted to the current weather conditions are calculated and output, including:

[0025] Based on the drive pedal input request and the pedal drive force curve, the initial drive torque is obtained within the motor performance safety limits; combined with the drive coefficient determined by the weather influence level and the tire adhesion coefficient, the initial drive torque is optimized to obtain the final output drive torque;

[0026] Based on the brake pedal input request and the pedal braking force curve, the initial braking torque is obtained within the safety limits of the braking system; combined with the braking coefficient determined by the weather impact level, the initial braking torque is optimized to obtain the final output braking torque;

[0027] Based on the steering wheel input request and steering force curve, the initial steering angle is obtained within the steering performance safety limit; combined with the steering coefficient determined by the weather impact level, the initial steering angle is optimized to obtain the final output steering angle.

[0028] Optionally, the weather impact level is mapped from environmental visibility and road adhesion coefficient to a value between 0 and 1.

[0029] Among them, environmental visibility is determined by rainfall and snowfall, fog concentration gradient and light intensity gradient; road adhesion coefficient is determined by road quality, pavement water mark distribution, water depth, snow depth or ice thickness.

[0030] When the weather impact level is close to 0, it indicates a severe environment. Driving stability can be improved by reducing the driving coefficient, increasing the braking coefficient, and reducing the steering coefficient.

[0031] When the weather impact level is close to 1, indicating a favorable environment, the vehicle's normal handling performance can be restored by increasing the drive coefficient, decreasing the braking coefficient, and increasing the steering coefficient.

[0032] Optionally, pre-built dynamic threshold relationship mapping tables between output driving torque and maximum vehicle speed, and between output steering angle and weather impact level, include:

[0033] The first mapping relationship is based on the real-time coupling of weather impact level and vehicle power system. The maximum speed limit is set for different weather impact levels and mapped to the maximum driving torque threshold to adapt to road adhesion and reserve reaction time for sudden situations in low visibility.

[0034] The second mapping relationship is based on the real-time coupling of weather impact level and steering system, setting the maximum allowable steering angle threshold for different weather impact levels to adapt to the road surface lateral adhesion and steering sensitivity under different weather conditions.

[0035] Optionally, the weather type prediction model employs a long short-term memory network to learn weather change trends from the multi-frame fused environmental information, and combines the contextual features of cloud meteorological data to output weather identification and short-term prediction results.

[0036] Secondly, embodiments of the present invention provide a vehicle weather environment adaptive device based on online learning, comprising:

[0037] The environmental image data processing module is used to acquire real-time environmental images, and sequentially pass the pre-processed real-time environmental images through an encoder, a latent space, and a decoder to extract information and perform quantization processing, outputting real-time environmental information; and then weightedly fuses the real-time environmental information with historical environmental information to obtain multi-frame fused environmental information.

[0038] The weather type detection module is used to input the multi-frame fused environmental information into a pre-built weather type prediction model and output the weather identification result, the weather impact level, and the environmental information prediction result for the next moment; wherein, the weather impact level is a value obtained by mapping the real-time environmental information volume and is used to comprehensively evaluate the severity of the vehicle driving environment, and its range is from 0 to 1;

[0039] The weather-mode vehicle control module is used to dynamically adjust the vehicle's drive coefficient, braking coefficient, and steering coefficient based on the weather impact level. It also calculates and outputs drive torque, braking torque, and steering angle adapted to the current weather conditions, taking into account the driver's input requests via the drive pedal, brake pedal, and steering wheel. Furthermore, based on pre-built dynamic threshold relationship mapping tables between output torque and maximum vehicle speed, and between output steering angle and weather impact level, it adaptively limits the drive torque to be less than the maximum drive torque and the steering angle to be less than the maximum steering angle.

[0040] The online parameter optimization module is used to calculate the difference data between the multi-frame fused environmental information and the predicted environmental information at the next moment; after fusing the difference data with cloud meteorological data, it is input into the online gradient network; the online gradient network calculates the adjustment gradient of the model parameters by comparing the weather recognition result at the current moment with the prediction result at the current moment cached by the system, minimizing the difference between the two as the loss, and updates the parameters of the encoder, decoder and weather type prediction model online based on the adjustment gradient.

[0041] Thirdly, embodiments of the present invention provide an electronic device, including:

[0042] At least one processor;

[0043] Memory for storing the at least one processor-executable instruction;

[0044] The at least one processor is configured to execute the instructions to implement the method as described in the first aspect.

[0045] The technical solution provided in this invention acquires real-time environmental images, and sequentially processes the pre-processed real-time environmental images through an encoder, latent space, and decoder for information extraction and quantization, outputting real-time environmental information. This real-time environmental information is then weighted and fused with historical environmental information to obtain multi-frame fused environmental information. This multi-frame fused environmental information, along with cloud-based meteorological data acquired in real-time, is input into a pre-constructed weather type prediction model, outputting weather identification results, weather impact level, and environmental information prediction results for the next moment. Based on the weather impact level, the vehicle's drive coefficient, braking coefficient, and steering coefficient are dynamically adjusted. Combined with the driver's input requests via the drive pedal, brake pedal, and steering wheel, the driving torque and braking torque adapted to the current weather conditions are calculated and output. And steering angle; and based on the pre-built dynamic threshold relationship mapping table between output driving force and maximum vehicle speed, and the dynamic threshold relationship mapping table between output steering angle and weather impact level, adaptively limit the driving torque to be less than the maximum driving torque, and adaptively limit the steering angle to be less than the maximum steering angle; calculate the difference data between multi-frame fused environmental information and the environmental information prediction result at the next moment; after fusing the difference data and cloud meteorological data, input it into the online gradient network; the online gradient network calculates the adjustment gradient of model parameters by comparing the weather recognition result at the current moment with the prediction result at the current moment cached by the system, with the loss being the minimization of the difference between the two, and updates the parameters of the encoder, decoder and weather type prediction model online based on the adjustment gradient.

[0046] The technical solution provided by this invention quantifies weather environment information and couples the mapped weather impact level with the vehicle control system in real time. It adaptively adjusts the drive coefficient, braking coefficient, and steering coefficient in the current vehicle control system, achieving for the first time the adaptive coupling of the vehicle control system with weather gradient changes. This method innovates the single weather mode control in traditional driving strategies into a smooth switching of vehicle control parameters according to the weather impact level. At the same time, the online gradient network automatically trains the weather type detection model online, eliminating the model's dependence on periodic offline retraining. It learns real-time data and updates model parameters, enabling rapid capture of local climate and improving prediction accuracy. While ensuring the stability of overall prediction accuracy, it significantly improves the model's adaptability to sudden weather changes. Attached Figure Description

[0047] Figure 1 This is the overall framework of the weather-adaptive vehicle control system of the present invention.

[0048] Figure 2 This is a flowchart illustrating a vehicle weather environment adaptive method based on online learning according to the present invention.

[0049] Figure 3 This is a flowchart illustrating the weather type detection module of the present invention.

[0050] Figure 4 This is a schematic diagram of the adaptive control logic of the drive system of the present invention.

[0051] Figure 5 This is a schematic diagram of the adaptive control logic of the braking system of the present invention.

[0052] Figure 6 This is a schematic diagram of the adaptive control logic of the steering system of the present invention.

[0053] Figure 7 This is a schematic diagram of a vehicle weather environment adaptive device based on online learning according to the present invention. Detailed Implementation

[0054] The present invention will be described in detail below through embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0055] In this invention, the terms "in one possible embodiment," "exemplary," or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "in one possible embodiment," "exemplary," or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "in one possible embodiment," "exemplary," or "for example" is intended to present the relevant concepts in a specific manner.

[0056] Most existing environmental perception systems in autonomous driving and assisted driving technologies focus on target recognition and distance detection, using road condition assessment to address complex road conditions or sudden accident risks. Currently, mass-produced vehicles typically require drivers to manually adjust driving modes, usually based on pre-set fixed parameters to cope with limited weather scenarios under preset conditions. This presents a vehicle control mode of "driver adapting to the weather, vehicle passively adapting to the weather," making it difficult for drivers to accurately perceive the impact of weather changes on vehicle driving, easily leading to problems such as delayed control response and insufficient driving stability.

[0057] For example, most mass-produced vehicles on the market currently rely on sensors to identify weather conditions and automatically switch to fixed driving modes, such as rain, snow, and fog modes. These modes employ preset, fixed parameters to formulate driving strategies for a limited number of weather types, which cannot cover complex and ever-changing weather gradients. While headlights automatically turn on when the ambient light dims and fog lights are manually activated in foggy conditions, the vehicle does not actively utilize enhanced sensor fusion algorithms to address reduced visibility. This requires drivers to accurately assess road conditions and weather and switch modes promptly.

[0058] When a vehicle is traveling on a seemingly dry highway and suddenly encounters a downpour, the rainwater mixes with dust and oil on the road surface to form an extremely slippery film. The driver, judging by the windshield that it's raining, instinctively lightly applies the brakes to slow down. However, due to the drastic decrease in road surface adhesion limits, the braking force calibrated for dry roads is still too great, easily causing the tires to lock up or triggering the ABS (Anti-lock Braking System) frequently, resulting in a longer braking distance than expected and increasing the risk of a rear-end collision. This response to sudden situations is isolated and mechanical, unable to handle the continuously changing gradient of conditions such as "light rain," "moderate rain," "heavy rain," "just wet road surface," "slippery road surface," and "water accumulation on the road." Furthermore, the vehicle lacks the subtle perception of how changes in weather affect its driving, failing to automatically and linearly adjust braking force for smooth deceleration and failing to provide the driver with a more stable steering feel.

[0059] Therefore, it is necessary to propose a new method that adaptively couples the vehicle control system with weather gradient changes based on current weather information and predicted environmental information, thereby significantly improving the model's adaptability to identifying and predicting sudden weather changes while ensuring the accuracy of weather prediction.

[0060] To ensure the plan is clearly described, the following will first combine... Figure 1 The overall technical solution of the embodiments of the present invention will be described.

[0061] Figure 1 The overall framework of the weather-adaptive vehicle control system provided in an embodiment of the present invention is shown. Figure 1 As shown, the entire system constitutes a complete technical closed loop from environmental perception to vehicle control, including online optimization. The specific process and module correspondence are as follows:

[0062] Step 1: Input multi-source feature information.

[0063] After the system starts, it receives feature information from two types of data sources in parallel:

[0064] 1. Images: Real-time environmental video streams captured by vehicle-mounted cameras serve as the raw input for visual perception.

[0065] 2. Cloud-based weather service. Real-time weather data received through the vehicle-to-everything (V2X) module provides macro-level weather information.

[0066] Step 2: Weather type detection and pattern quantification.

[0067] The feature information from the two aforementioned sources is simultaneously input to the core weather type detection module. This module uses an embedded neural network model (such as the encoder-decoder structure and prediction network, which will be detailed later) to fuse and analyze the input data, outputting a judgment result on the current weather type (such as sunny, rainy, foggy, or snowy). It also maps the overall impact of the weather on driving to a continuous weather pattern score. Figure 1 As shown, the score value ranges from 0 (representing severe weather) to 1 (representing good weather), realizing the conversion from discrete weather types to continuous control parameters.

[0068] Step 3: Control command distribution and execution.

[0069] The weather pattern value output by the weather type detection module is used as a core control variable and is distributed in real time to the various actuator control systems of the vehicle. For example... Figure 1 As shown, these systems mainly include:

[0070] First, VCU (Vehicle Control Unit): Adjusts drive torque output according to weather patterns.

[0071] Second, EPS (Electric Power Steering): Adjusts steering assist characteristics and angle limits according to weather patterns.

[0072] Third, EHB (Electro-hydraulic Braking System): Adjusts the braking response curve and braking force distribution according to weather patterns.

[0073] Fourth, AD / ADAS controller (Autonomous Driving / Advanced Driver Assistance System Controller): Receives weather patterns as input for higher-level decision-making and coordinates the overall autonomous driving strategy.

[0074] Step 4: Iteration of system online and automated operation.

[0075] Figure 1 The document clearly identifies the core features of this system: online operation and automation. To achieve this goal, the system includes an independent online optimization loop, specifically comprising:

[0076] (1) The weather type detection module not only outputs the current weather pattern, but also outputs a prediction of the short-term future weather conditions.

[0077] (2) The system calculates the prediction error by comparing the predicted value with the subsequent actual weather conditions (i.e., the result of the newly input feature information being identified by the detection module). This error is fed into the online gradient network (which can be regarded as...). Figure 1 (Part or related component of the "Weather Type Detection" module) generates adjustment gradients for model parameters through the backpropagation algorithm.

[0078] (3) The calculated gradients are used to update the parameters (such as weights and biases) of the internal neural network model of the weather type detection module online and automatically. This process continues while the vehicle is in motion, enabling the model to adapt to different regional and climatic characteristics without manual intervention or offline training. Figure 1 The arrows indicate "online and automated".

[0079] In summary, the embodiments of the present invention, through Figure 1 The clearly demonstrated closed-loop framework of "feature information input → weather type detection (output weather pattern) → vehicle control execution (VOU / EPS / EHB / ADAS) → feedback-based online automated update" enables the vehicle control system to adjust to the weather environment in real time, accurately and adaptively, and endows the system with the ability to learn and continuously optimize itself.

[0080] After describing the overall technical solution of the embodiments of the present invention, the following will describe in detail a vehicle weather environment adaptive method based on online learning provided by the embodiments of the present invention.

[0081] Figure 2 This is a flowchart illustrating a vehicle weather environment adaptive method based on online learning according to the present invention. Figure 2 As shown, the vehicle weather environment adaptive method based on online learning proposed in this invention includes the following steps:

[0082] S210: Acquire real-time environmental images, and sequentially pass the preprocessed real-time environmental images through an encoder, latent space, and decoder to extract information and perform quantization processing, outputting real-time environmental information; and then weightedly fuse the real-time environmental information with the historical environmental information to obtain multi-frame fused environmental information.

[0083] Specifically, the vehicle uses an optical lens (usually a CMOS or CCD image sensor) to continuously collect reflected light from the external environment, capturing real-time environmental images. The photosensitive unit within the lens converts the received light signal into a corresponding electrical signal (electron charge) according to its intensity, forming an analog electrical signal. The analog-to-digital converter within the image sensor samples and quantizes the analog electrical signal, converting continuous voltage values ​​into discrete values, ultimately generating a digital image signal—the real-time environmental image. The real-time environmental image is a matrix of pixels, where the value of each pixel represents the brightness of the grayscale image or the color component of the RGB color image at that point.

[0084] Rain and snow appear as moving, bright stripes or point-like noise in real-time environmental images. Water reflections from road surfaces and vehicle bodies create large areas of highlight. Preprocessing of real-time environmental images involves using a lightweight spatial median filtering algorithm with frequency domain filtering for initial noise reduction. This is achieved by replacing the center pixel value with the median of its neighboring pixels to filter out rain and snow noise. High-frequency noise from reflected light manifests in the spatial domain as local pixel brightness reaching the sensor's limit, resulting in complete loss of detail. Image inpainting is performed after thresholding. The pre-denoised real-time environmental image is then further processed by a lightweight deep learning model for refined noise reduction.

[0085] Unsharpened masks or high-pass filters can be used to enhance image details and edge information, making road textures and object outlines clearer, which is used to determine environmental visibility and road surface wetness / dryness. Regions of interest extraction algorithms can be used to locate image regions relevant to weather assessment based on color features, texture features, or semantic segmentation models.

[0086] A deep convolutional neural network encoder can be used to extract and compress features from the input image through a series of multi-layer convolution and downsampling operations, converting it into a high-dimensional feature vector. In the latent space, key environmental features are automatically learned and extracted from the high-dimensional feature vector. After the decoder performs upsampling or transposed convolution on the key environmental features, it undergoes a quantization process to reduce the dimensionality, thus obtaining the real-time environmental information.

[0087] The system stores the calculated real-time environmental information in a fixed-length buffer. When the next moment arrives, the buffered real-time information is automatically converted into historical environmental information. The multi-frame fused environmental information is arranged in chronological order, and the real-time environmental information is weighted and fused with the historical environmental information to represent the high-frequency, quantified real-time evolution process of the local environment of the vehicle in a short period of time.

[0088] S220 inputs multi-frame fused environmental information into a pre-built weather type prediction model and outputs weather identification results, weather impact level, and environmental information prediction results for the next moment.

[0089] Specifically, the multi-frame fused environmental information of each frame is spliced ​​or cross-attention is calculated to form a fused temporal sequence, which is then input into an LSTM (Long Short-Term Memory) network. The LSTM processes each step in the temporal sequence step by step and classifies the current weather based on the multi-frame fused environmental information, such as: sunny, cloudy, light rain, heavy rain, fog, light snow, etc. At the same time, based on the learned weather change trend, the system predicts the short-term weather conditions and outputs the corresponding confidence level.

[0090] The weather impact level is a value ranging from 0 to 1, obtained by normalizing and mapping real-time environmental information. It is used to comprehensively evaluate the severity of the vehicle driving environment.

[0091] When the weather impact level approaches 0, it indicates that the current environment is extremely bad, which may be a combination of "dense fog + ice surface", that is, "extremely low environmental visibility + low road adhesion coefficient", which means that the risk of vehicles losing control is extremely high at this time.

[0092] When the weather impact level approaches 1, it indicates that the current environment is good, which may be a combination of "sunny weather + dry asphalt road surface", that is, "high environmental visibility + high road adhesion coefficient". At this time, the normal operating performance of the vehicle can be restored.

[0093] S230 dynamically adjusts the vehicle's drive coefficient, braking coefficient, and steering coefficient based on the weather impact level. Combining the driver's input requests from the drive pedal, brake pedal, and steering wheel, it calculates and outputs drive torque, braking torque, and steering angle adapted to the current weather conditions. Based on a pre-built dynamic threshold relationship mapping table between output drive torque and maximum vehicle speed, and a dynamic threshold relationship table between output steering angle and weather impact level, it adaptively limits drive torque to less than the maximum drive torque and adaptively limits steering angle to less than the maximum steering angle.

[0094] Specifically, when the weather impact level decreases, indicating that the current weather is becoming more severe, reducing the drive coefficient means that with the same accelerator pedal opening, the engine torque output is significantly reduced, resulting in smoother vehicle acceleration and preventing wheel slippage due to excessive torque. Increasing the braking coefficient means that with the same brake pedal travel, the system intervenes in braking assistance earlier, optimizing braking distance on low-traction surfaces and preventing loss of control during emergency braking. Reducing the steering coefficient increases steering damping, making the steering wheel feel heavier, meaning that with the same steering wheel rotation angle, the wheel rotation angle is smaller, preventing vehicle skidding caused by excessive steering wheel rotation.

[0095] When the weather impact level increases, and the current weather tends to be good, the driving coefficient should be appropriately increased to provide aggressive throttle response and restore the vehicle's power demand; the braking coefficient should be reduced and the steering coefficient increased to restore the vehicle's normal performance.

[0096] S240 calculates the difference data between the multi-frame fused environmental information and the environmental information prediction result at the next moment; after fusing the difference data and cloud meteorological data, it is input into the online gradient network; the online gradient network calculates the adjustment gradient of the model parameters by comparing the weather recognition result at the current moment with the prediction result at the current moment cached by the system, minimizing the difference between the two as the loss, and updates the parameters of the encoder, decoder and weather type prediction model online based on the adjustment gradient.

[0097] Specifically, the Kalman filter algorithm is used to fuse the differential data and cloud meteorological data. Based on the trend prediction of the cloud meteorological data, the current deviation should be estimated, providing a prior estimate. The predicted value of the prior estimate is then weighted and fused with the actual observed differential data. The weight of the predicted value of the prior estimate is less than the weight of the differential data, which is used to filter out random noise in the differential data.

[0098] A loss function is constructed to measure the prediction accuracy of the weather type forecasting model. The model parameters include all weight parameters and bias parameters, which refer to the adjustable values ​​within the encoder, decoder, and prediction model. Based on the backpropagation algorithm, starting from the output of the loss function, the partial derivatives of the loss function with respect to each parameter are calculated backward using the chain rule. These partial derivatives together constitute the loss gradient, i.e., the adjustment gradient.

[0099] The gradient adjustment is fed back to the encoder, decoder, and weather type prediction model. Parameter adjustments are performed with a preset, robust step size and learning rate. The model parameters are in a continuous process of small optimization, which significantly improves the model's adaptability to sudden weather changes while ensuring the stability of the overall prediction accuracy.

[0100] This embodiment constructs a closed loop of "weather perception, strategy adjustment, and execution feedback," which smoothly responds to weather changes in real time without requiring manual intervention from the driver, enabling the vehicle to proactively adapt to the weather.

[0101] The technical solution provided in this invention acquires real-time environmental images, and sequentially processes the pre-processed real-time environmental images through an encoder, latent space, and decoder for information extraction and quantization, outputting real-time environmental information. This real-time environmental information is then weighted and fused with historical environmental information to obtain multi-frame fused environmental information. This multi-frame fused environmental information, along with cloud-based meteorological data acquired in real-time, is input into a pre-constructed weather type prediction model, outputting weather identification results, weather impact level, and environmental information prediction results for the next moment. Based on the weather impact level, the vehicle's drive coefficient, braking coefficient, and steering coefficient are dynamically adjusted. Combined with the driver's input requests via the drive pedal, brake pedal, and steering wheel, the driving torque and braking torque adapted to the current weather conditions are calculated and output. And steering angle; and based on the pre-built dynamic threshold relationship mapping table between output driving force and maximum vehicle speed, and the dynamic threshold relationship mapping table between output steering angle and weather impact level, adaptively limit the driving torque to be less than the maximum driving torque, and adaptively limit the steering angle to be less than the maximum steering angle; calculate the difference data between multi-frame fused environmental information and the environmental information prediction result at the next moment; after fusing the difference data and cloud meteorological data, input it into the online gradient network; the online gradient network calculates the adjustment gradient of model parameters by comparing the weather recognition result at the current moment with the prediction result at the current moment cached by the system, with the loss being the minimization of the difference between the two, and updates the parameters of the encoder, decoder and weather type prediction model online based on the adjustment gradient.

[0102] As can be seen, this invention, by quantifying weather environment information, couples the mapped weather impact level with the vehicle control system in real time, adaptively adjusting the drive coefficient, braking coefficient, and steering coefficient in the current vehicle control system. This achieves for the first time the adaptive coupling of the vehicle control system with weather gradient changes. This method innovates the single weather mode control in traditional driving strategies into a smooth switching of vehicle control parameters according to the weather impact level. At the same time, the online gradient network automatically trains the weather type detection model online, eliminating the model's dependence on periodic offline retraining. It learns real-time data and updates model parameters, enabling rapid capture of local climate and improving prediction accuracy. While ensuring the stability of overall prediction accuracy, it significantly improves the model's adaptability to sudden weather changes.

[0103] Based on the above embodiments, as one implementation of the present invention, the preprocessed real-time environment image is sequentially passed through an encoder, a latent space, and a decoder for information extraction and quantization processing, and the real-time environment information is output.

[0104] Figure 3 This is a flowchart illustrating the weather type detection module. Figure 3 As shown, the environmental image data processing in the weather type detection model includes the following steps, namely steps a1 to a3:

[0105] Step a1 involves preprocessing the real-time environmental image acquired by the camera, including: filtering out noise caused by rain, snow, and reflected light in the real-time environmental image based on image processing algorithms, while enhancing the details and edge information of the image; identifying image regions related to weather judgment using region of interest extraction algorithms and making decisions, and outputting the preprocessed real-time environmental image.

[0106] Initial noise reduction can be achieved using a lightweight spatial median filtering algorithm combined with a frequency domain filtering algorithm. For pixel coordinates (x, y), its brightness values ​​over N consecutive frames (e.g., 5 frames) are collected.

[0107] The brightness values ​​of N frames are sorted. Rain and snow typically appear as outliers in a pixel within a short period. The median of the sorted list is selected as the output value for N frames. Thresholding segmentation is performed using the Otsu algorithm, which automatically calculates the optimal threshold in the image. This threshold divides the image pixels into overexposed and normal regions, maximizing the variance of these two types of pixels, and generating a binary mask. Bilinear interpolation or weighted averaging can be used to gradually fill in the overexposed regions from their boundaries inwards. The value of each pixel to be repaired is determined by the weighted average of its surrounding known pixels.

[0108] One approach is to use a sharpening filter algorithm to enhance the details and edge information of an image, specifically by enhancing the high-frequency components. First, a Gaussian blur is applied to the original image via convolution with a Gaussian kernel, averaging the values ​​of each pixel and its surrounding pixels to preserve the overall contour and smooth areas. Subtracting the Gaussian-filtered image from the original image yields a detail image, and adding the detail image to the original image results in the sharpened image.

[0109] Specifically, algorithms for region of interest extraction are used to locate and determine weather-related image regions based on color features, texture features, or semantic segmentation models. Different weather phenomena show significant characteristics in different image regions, specifically as follows:

[0110] 1. The sky region is the main basis for judging cloud cover, precipitation, haze, and overall brightness. Based on color and texture features, combined with color threshold segmentation and texture analysis, the local variance is calculated. The region with small variance is the smooth sky region, and the boundary is determined by region growing.

[0111] 2. The road surface area is a key area for determining whether the road surface is dry, wet, icy, or snowy, and whether the road surface is uneven. A CNN model can be trained to segment each pixel in the image.

[0112] 3. The distant view area is used to quantify environmental visibility. The range of the distant view area is determined by combining the lower boundary of the sky area and the far end of the road area.

[0113] Step a2 involves inputting the real-time environment image into the encoder, converting the pixel matrix of the real-time environment image into a high-dimensional feature vector, which is used to characterize the spatial features of the real-time environment image, and extracting key environmental information from the high-dimensional feature vector in the latent space.

[0114] Specifically, a deep convolutional neural network encoder can be used. Through a series of multi-layer convolution and downsampling operations, the convolutional kernel slides on the image. By using local connectivity and weight sharing, it focuses on the gradient of light intensity, the gradient concentration of fog areas, and the amount of rainfall and snowfall, as well as road quality, the distribution of water marks on the road surface, water depth, snow depth, or ice thickness. These are internalized into the encoder's weight parameters. High-dimensional feature vectors are extracted layer by layer. The downsampling operation can increase the receptive field to capture global contextual information. Feature extraction and compression are performed on the input image. In the latent space, key environmental information is extracted from the high-dimensional feature vectors.

[0115] Step a3 involves using a decoder to reduce the dimensionality and normalize the key environmental features, mapping and quantizing the key environmental features into real-time environmental information, and caching the real-time environmental information in the system as historical environmental information.

[0116] Specifically, the decoder upsamples key environmental features to generate a larger feature map, reconstructing feature details. Then, a specific activation function is used to map and quantize physical quantities, outputting real-time environmental information (e.g., [environmental visibility = 850 meters, road adhesion coefficient = 0.7, weather type = light rain]). The real-time environmental information calculated at each moment is stored at the tail of a queue. When the queue is full, new data is automatically pushed out the oldest data at the head of the queue, constructing historical environmental information.

[0117] This implementation sequentially passes the preprocessed real-time environment image through an encoder, latent space, and decoder for information extraction and quantization, resulting in the following two significant benefits:

[0118] Firstly, preprocessing real-time environmental images can effectively remove noise such as rain, snow, and reflections, greatly improving the accuracy of feature extraction, preventing erroneous information from being amplified in subsequent stages, improving image clarity in low-light and foggy scenes, and removing redundant information such as billboards and pedestrians, focusing on areas related to weather judgment.

[0119] Secondly, the structure of the "encoder and decoder" transforms qualitative descriptions into quantitative descriptions, quantifies environmental information, and can be coupled with the vehicle control system to achieve precise control of the vehicle's drive coefficient, braking coefficient, and steering coefficient.

[0120] exist Figure 2 Based on the illustrated embodiment, as an implementation method of this invention, the real-time environmental information and historical environmental information are weighted and fused to obtain multi-frame fused environmental information, which may include steps b1 to b3:

[0121] Step b1: The multi-frame fused environmental information is the first environmental information obtained by weighted fusion of historical environmental information from time (tk) to time t.

[0122] In this method, the Kalman filter algorithm is used to fuse data at a time step of (k+1), assigning greater weight to historical environmental information that is closer to time t and has higher confidence.

[0123] Step b2, calculate the difference data between the multi-frame fused environmental information and the environmental information prediction result at the next time step, including: classifying and predicting the multi-frame fused environmental information at time (t-1) based on the weather type prediction model, and obtaining the environmental information prediction result at time t, which is used as the second environmental information quantity.

[0124] In this process, the multi-frame fused environmental information at time (t-1) is input into the LSTM prediction network, and the model learns the environmental evolution trend over the past k time steps to predict the environmental information at time t.

[0125] Step b3: Calculate the error between the first environmental information quantity and the second environmental information quantity to obtain differential data.

[0126] This implementation utilizes multi-frame fusion of environmental information to calculate differential data, which filters out errors caused by random fluctuations in sensor readings and decoder predictions. By using historical environmental information from consecutive frames to weightedly fuse real-time environmental information, it effectively smooths out random noise and obtains more reliable environmental information. Simultaneously, the system uses its own generated, more reliable fused information to verify the model's prediction results and determine the model's prediction error.

[0127] exist Figure 2 Based on the illustrated embodiment, as one implementation of this invention, the differential data and cloud-based meteorological data are fused and then input into an online gradient network, which may include the following steps, including steps c1 to c4:

[0128] Step c1: Connect to the cloud-based meteorological service in advance and obtain cloud-based meteorological data after passing through an environmental information network. The environmental information network is used to extract and identify weather-related feature information.

[0129] Specifically, the data interface of the National Meteorological Administration is connected in advance to obtain raw macro-meteorological data, which mainly includes the following:

[0130] First, it covers numerical forecast grid data within a 100-kilometer radius around the vehicle, including data such as temperature, humidity, air pressure, wind direction and speed, precipitation type and intensity, visibility, and cloud cover.

[0131] Second, radar reflectivity mosaic, which displays the precipitation area, intensity, direction of movement and speed in real time;

[0132] Third, satellite cloud images display the distribution and development trends of large-scale cloud systems in real time.

[0133] The Environmental Information Network is a lightweight deep neural network deployed on regional edge computing nodes. It receives meteorological data from multiple sources and in multiple spatiotemporal dimensions, performs spatiotemporal alignment and normalization, and can use the Transformer encoder architecture to capture key weather elements, such as the movement, development or dissipation trends of precipitation core areas and fog areas, and finally outputs structured and semantic weather feature vectors.

[0134] Step c2 involves weighted fusion of the differential data and cloud-based meteorological data using the Kalman filter algorithm to filter out random noise in the differential data and obtain comprehensive data, wherein the weight of the cloud-based meteorological data is less than the weight of the differential data.

[0135] Specifically, cloud-based meteorological data provides prior estimates to adjust for uncertainties in process noise. The prior estimate refers to the predicted value of the current deviation. At the same time, the covariance matrix of the prior estimate is calculated. The covariance matrix of the prior estimate is dynamically adjusted using cloud-based meteorological data. The Kalman gain is calculated by fusing the differential data. The weights of the Kalman gain are dynamically allocated to the differential data to obtain a comprehensive data that filters out noise and incorporates macro weather change trends.

[0136] Step c3: The online gradient network constructs a loss function based on comprehensive data with the goal of improving the accuracy of weather forecasting, and calculates the gradient of the loss function with respect to the model parameters through the backpropagation algorithm to obtain the loss gradient.

[0137] Step c4: Based on the preset learning rate, the loss gradient is backpropagated using the gradient descent algorithm to update the weights and bias parameters of the encoder, decoder, and weather type prediction model.

[0138] The method involves inputting comprehensive data into an online gradient network, using weather prediction accuracy as the loss function, calculating the loss gradient of each layer in reverse, calculating the partial derivatives with respect to all weights and biases, and following the chain rule in calculus to traverse the network in reverse. Finally, it outputs the set of partial derivatives of the loss function with respect to all trainable weights and biases in the network.

[0139] By using the online gradient network described in this implementation to provide real-time feedback parameter tuning, a paradigm shift is achieved from passive stability control based on weather conditions to feedforward stability control based on environmental perception. This dynamically optimizes model parameters and effectively overcomes the performance degradation problem of fixed-parameter models.

[0140] exist Figure 2 Based on the illustrated embodiment, as one implementation of this invention, the vehicle's drive coefficient, braking coefficient, and steering coefficient are dynamically adjusted based on the weather impact level. Combined with the driver's input requests via the drive pedal, brake pedal, and steering wheel, the driving torque, braking torque, and steering angle adapted to the current weather conditions are calculated and output. This may include the following three steps:

[0141] Step d1: Based on the drive pedal input request and the pedal drive force curve, obtain the initial drive torque within the motor performance safety limits; combine the drive coefficient determined by the weather influence level and the tire adhesion coefficient to optimize the initial drive torque and obtain the final output drive torque.

[0142] Specifically, the vehicle control system receives the raw voltage signal from the drive pedal sensor, which is proportional to the driver's acceleration request, i.e., the drive pedal input request. This drive pedal input request is then compared in real-time with the vehicle's performance limitations, such as the motor's external characteristic curve (peak torque and speed relationship) and the battery system's maximum discharge power. Within the motor's performance safety limits, the initial drive torque is obtained by analyzing the current pedal opening, while a vehicle state observer is used to estimate the initial coefficient of friction between the tires and the road surface. By querying the drive coefficient control curve and determining the drive coefficient based on the weather impact level, the initial drive force is optimized to obtain the drive torque.

[0143] For example, in sunny and dry road conditions, the weather impact level is 1.0 and the tire adhesion coefficient is 0.8. Therefore, the drive coefficient γ is 0.8 to adapt to normal road conditions and ensure drive efficiency.

[0144] In rainy and slippery road conditions, with a weather impact level of 0.6 and a tire adhesion coefficient of 0.8, the drive coefficient γ is 0.48 to adapt to rainy and slippery road conditions, reduce the speed of torque output, and prevent tire slippage.

[0145] In icy and extremely slippery road conditions, with a weather impact level of 0.3 and a tire adhesion coefficient of 0.8, the driving coefficient γ is 0.24 to adapt to the current adverse weather road conditions.

[0146] Step d2: Based on the brake pedal input request and the pedal braking force curve, obtain the initial braking torque within the safety limits of the braking system; combine the braking coefficient determined by the weather impact level to optimize the initial braking torque and obtain the final output braking torque.

[0147] Specifically, the vehicle control system receives the raw voltage signal from the brake pedal sensor, which is proportional to the driver's deceleration request, i.e., the brake pedal input request. Based on the vehicle system's braking capacity limitations, the brake pedal input request is mapped to an initial braking torque in the pedal braking force curve. By querying the braking coefficient control curve and determining the braking coefficient according to the weather impact level, the initial braking torque is optimized to obtain the final braking torque. For example, in rainy or snowy weather, when the weather impact level drops to 0.3, the braking coefficient β is increased to extend the braking distance, adapting to the reduced road adhesion.

[0148] Step d3: Based on the steering wheel input request and steering force curve, obtain the initial steering angle within the steering performance safety limit; combine the steering coefficient determined by the weather influence level to optimize the initial steering angle and obtain the final output steering angle.

[0149] Specifically, the vehicle control system receives the raw voltage signal from the steering wheel sensor, which is proportional to the driver's steering request, i.e., the steering wheel input request. Based on the vehicle system's steering performance safety limitations, the steering wheel input request is mapped to an initial steering angle in the steering force curve. By querying the steering coefficient control curve and determining the steering coefficient based on the weather impact level, the initial steering angle is optimized to obtain the final steering angle. For example, in rainy or snowy weather, with a weather impact level of 0.3, the steering coefficient α is reduced to increase steering damping and prevent oversteering due to slippery road surfaces.

[0150] This embodiment provides a gradual and predictable adjustment to the vehicle's dynamic performance control. First, the vehicle's drive motor and load motor are used as internal performance limits. By sensing the weather impact level under the current vehicle driving condition in real time, the drive, braking, and steering coefficients are smoothly adjusted in advance to prevent the vehicle from skidding due to sudden acceleration and rear-end collisions. This achieves smooth control of the vehicle by adapting it to the current weather conditions.

[0151] exist Figure 2 Based on the illustrated embodiment, as one implementation of the present invention, the weather impact level is mapped from environmental visibility and road adhesion coefficient to a value between 0 and 1.

[0152] Among them, environmental visibility is determined by rainfall and snowfall, fog concentration gradient and light intensity gradient; road adhesion coefficient is determined by road quality, road surface watermark distribution, water depth, snow depth or ice thickness.

[0153] Specifically, precipitation and snowfall refer to the liquid depth of rain and snow (after melting) per unit time and under specific conditions. Rain and snow reduce atmospheric visibility through absorption, scattering, and shading effects on light. Fog concentration gradient refers to the rate of change in fog concentration; light intensity gradient refers to the degree of light attenuation in the environment. Based on real-time environmental information, the actual visibility distance and the degree of decrease in visual contrast are estimated, resulting in a visibility index. The lower the value, the worse the current environmental visibility. The real-time environmental information output includes road quality, road surface watermark distribution, water depth, snow depth, or ice thickness. Road quality refers to determining whether the current road surface is flat or muddy, and calculating the maximum potential adhesion coefficient of the current road surface, i.e., the road adhesion coefficient. After comprehensively calculating environmental visibility and the road adhesion coefficient, a normalized mapping is performed to a value in the range of 0 to 1.

[0154] When the weather impact level approaches 0, it indicates that the current environment is extremely harsh, possibly a combination of "dense fog + icy surface", that is, "extremely low environmental visibility + low road adhesion coefficient", which means that the risk of losing vehicle control is extremely high. Driving stability can be improved by reducing the driving coefficient, increasing the braking coefficient, and reducing the steering coefficient.

[0155] When the weather impact level approaches 1, it indicates that the current environment is good, which may be a combination of "sunny weather + dry asphalt road surface", that is, "high environmental visibility + high road adhesion coefficient". The normal handling performance of the vehicle can be restored by increasing the driving coefficient, decreasing the braking coefficient, and increasing the steering coefficient.

[0156] The system described in this implementation continuously and smoothly adjusts control parameters under different weather conditions, transforming the vehicle control system from a static driving mode to an active prevention mode in response to weather conditions, thus making the driver's control of the vehicle more stable.

[0157] exist Figure 2 Based on the illustrated embodiment, as one implementation of the present invention, a dynamic threshold relationship mapping table between the output driving torque and the maximum vehicle speed, and a dynamic threshold relationship mapping table between the output steering angle and the weather impact level are pre-constructed.

[0158] The first mapping relationship is based on the real-time coupling of weather impact level and vehicle power system. The maximum speed limit is set for different weather impact levels and mapped to the maximum driving torque threshold to adapt to road adhesion and reserve reaction time for emergencies in low visibility.

[0159] The system sets maximum speed limits for different weather impact levels, creating a chain control mechanism of "weather impact level, maximum speed limit, and mapped maximum driving torque threshold." When the weather impact level decreases, it means reduced visibility. Based on the visible distance perceived by the camera and the driver's typical reaction time, the system calculates the maximum speed at which the vehicle can safely stop within the visible distance and maps it to the maximum driving torque threshold. This maximum driving torque threshold is then sent to the vehicle controller (VOU). Therefore, regardless of how deeply the driver presses the accelerator or how the adaptive drive coefficient is adjusted, the system's final driving torque output will be strictly limited to this maximum driving torque threshold, fundamentally preventing rear-end collisions due to insufficient visibility and providing crucial reaction time for emergencies.

[0160] The second mapping relationship is based on the real-time coupling of weather impact level and steering system, setting the maximum allowable steering angle threshold for different weather impact levels to adapt to the road surface lateral adhesion and steering sensitivity under different weather conditions.

[0161] Specifically, by calculating the theoretical maximum front wheel steering angle at the tire's adhesion limit under different road adhesion coefficients, the weather impact level is coupled with the steering system in real time. This maximum front wheel steering angle is then sent to the electric power steering (EPS) system. When the steering angle requested by the driver or the autonomous driving system exceeds this threshold, the EPS will limit the angle, ensuring that the actual steering angle remains within a safe and stable range. This effectively avoids skidding and fishtailing accidents caused by oversteering (such as turning the steering wheel too sharply during emergency maneuvers) on low-traction surfaces.

[0162] By setting the maximum driving torque threshold and the maximum permissible steering angle threshold that match the real-time environmental risks, it effectively prevents drivers from speeding due to misjudgment of environmental visibility or road conditions, and avoids skidding and fishtailing due to oversteering in emergency avoidance situations.

[0163] exist Figure 2 Based on the illustrated embodiment, as one implementation of the present invention, the weather type prediction model employs a long short-term memory network to learn weather change trends from the multi-frame fused environmental information, and combines the contextual features of cloud meteorological data to output weather identification and short-term prediction results.

[0164] Specifically, the contextual features of cloud-based meteorological data serve as the background environment, which is then concatenated or cross-attention calculated with the multi-frame fused environmental information of each frame. The resulting fused temporal sequence is input into an LSTM (Long Short-Term Memory) network for forward propagation to learn weather evolution patterns. The gated recurrent units in the LSTM process each step of the temporal sequence step by step, classifying specific weather states based on the multi-frame fused environmental information, such as: sunny, cloudy, light rain, heavy rain, fog, light snow, etc., while simultaneously predicting short-term weather conditions and outputting corresponding confidence scores.

[0165] The LSTM prediction network described in this implementation transforms the vehicle's reactive driving strategy in response to weather into predictive decisions, enabling rapid capture of local climate conditions and timely mapping to the vehicle control system.

[0166] The following will combine Figures 4 to 6 The specific implementation algorithm and control logic of the embodiments of the present invention will be further described in detail.

[0167] Figure 4 The adaptive control logic of this invention in a drive system is illustrated. The specific implementation process is as follows:

[0168] The system receives pedal input requests. x Then, the drive pedal travel is obtained, and the initial drive torque is obtained by mapping the pedal drive force curve. The request x First, the motor performance limiting module is checked to ensure the initial driving torque. Not exceeding the maximum torque of the motor, i.e. x ≤ Additionally, the request should also refer to the tire adhesion coefficient curve to prevent excessive torque from causing the drive wheels to slip.

[0169] The driving coefficient control module dynamically calculates the driving coefficient γ based on the weather pattern, where γ = q· The mapping rule is as follows: when the weather mode approaches 0 (severe weather), γ decreases to reduce the response rate and peak value of the drive torque and avoid sudden acceleration that could cause the drive wheels to spin freely; when the weather mode approaches 1 (good weather), γ returns to the baseline value.

[0170] The maximum speed control module dynamically calculates the safe maximum speed under the current environment based on the weather pattern, and then calculates the maximum longitudinal driving force threshold allowed by the powertrain system based on this. .

[0171] The system will handle drive requests that have passed the motor speed performance limit. Multiplying by the driving coefficient γ, we obtain the weather-adapted driving request γ·x (i.e., The requested value must also satisfy γ·x ≤ The final safety constraints. Ultimately, the system output torque (drive) is managed through both coefficient adjustment and vehicle speed threshold.

[0172] By introducing the driving coefficient γ and the maximum driving force threshold based on weather patterns On the one hand, γ suppresses slippage during start-up and acceleration on low-traction surfaces; on the other hand, through... This ensures that the vehicle will not exceed the safe speed allowed by the current visibility and road conditions under any circumstances, thus preventing the risk of speeding.

[0173] Figure 5 The adaptive control logic of this invention in a braking system is illustrated. The specific implementation process is as follows:

[0174] The system receives pedal input requests. y This allows us to obtain the brake pedal travel, and then map the initial braking torque using the pedal braking force curve. The brake pedal input request y is first verified by the brake performance limiting module to ensure the initial braking torque. The braking force shall not exceed the system's maximum braking force, i.e., y ≤ At the same time, the system acquires the weather pattern value output by the weather type detection module.

[0175] The braking coefficient control module dynamically calculates the braking coefficient β based on the weather pattern. The mapping rule is as follows: when the weather pattern approaches 0 (severe weather, such as rain, snow, or icy roads), β increases to advance and enhance the braking response and compensate for the extended braking distance caused by the decrease in road adhesion; when the weather pattern approaches 1 (good weather), β returns to the reference value.

[0176] The system will process braking requests that have been subject to basic limits. Multiplying this by the braking coefficient β yields the weather-adapted braking torque request β·y. Finally, the system outputs this output torque (braking).

[0177] By introducing a braking coefficient β controlled by weather patterns, this embodiment achieves intelligent pre-compensation for braking torque. On low-traction surfaces, the system automatically increases braking force output, enabling the driver to obtain a deceleration that better matches the expected actual stopping distance when pressing the brake pedal for the same distance. This effectively compensates for the reduced braking performance caused by slippery road surfaces, shortens the effective braking distance, and improves driving safety.

[0178] Figure 6 The adaptive control logic of this invention in a steering system is illustrated. The specific implementation process is as follows:

[0179] The system receives the steering wheel input request z, then obtains the steering wheel travel, and maps it to the steering force curve to obtain the initial steering angle. The request is first checked by the performance limiting module to ensure its base value. The system does not exceed its mechanical limits, i.e., z ≤ At the same time, the system acquires the weather pattern value output by the weather type detection module.

[0180] The steering coefficient control module dynamically calculates the steering coefficient α based on the weather pattern. The mapping rule is as follows: when the weather pattern approaches 0 (severe weather), α decreases to reduce steering sensitivity and response speed; when the weather pattern approaches 1 (good weather), α increases or returns to the reference value of 1 to ensure steering agility.

[0181] The maximum turning angle control module dynamically calculates the maximum turning angle under the current conditions based on the weather pattern. The maximum turning angle is set based on the estimated road surface lateral adhesion coefficient, with the principle of physically avoiding sideslip.

[0182] The system will redirect the request after basic restrictions. Multiply by the steering coefficient α to obtain the weather-adapted steering request. (i.e., αz). It must satisfy αz ≤ Safety constraints. Ultimately, the system output angle φ is adjusted in two ways (coefficient adjustment and threshold limiting).

[0183] By introducing weather-based control of the steering coefficient α and dynamic maximum steering angle This embodiment enables the steering system to adapt to road surface adhesion conditions. In adverse weather conditions, the system automatically increases steering damping, reduces steering gain, and limits the maximum steering angle, effectively suppressing the risk of oversteering and instability caused by excessive driver operation or slippery road surfaces, thus improving steering safety and controllability.

[0184] Secondly, embodiments of the present invention also provide a vehicle weather environment adaptive device 70 based on online learning, such as... Figure 7 As shown, it includes:

[0185] The environmental image data processing module 710 is used to acquire real-time environmental images, and sequentially pass the pre-processed real-time environmental images through an encoder, a latent space and a decoder to extract information and perform quantization processing, and output real-time environmental information; and weightedly fuse the real-time environmental information with the historical environmental information to obtain multi-frame fused environmental information.

[0186] The weather type detection module 720 is used to input the multi-frame fused environmental information into a pre-built weather type prediction model and output the weather identification result, the weather impact level, and the environmental information prediction result for the next moment; wherein, the weather impact level is a value obtained by mapping the real-time environmental information and used to comprehensively evaluate the severity of the vehicle driving environment, and its range is from 0 to 1.

[0187] The weather mode vehicle control module 730 is used to dynamically adjust the vehicle's drive coefficient, braking coefficient, and steering coefficient based on the weather impact level. It calculates and outputs drive torque, braking torque, and steering angle adapted to the current weather conditions, taking into account the driver's input requests from the drive pedal, brake pedal, and steering wheel. Based on a pre-built dynamic threshold relationship mapping table between output torque and maximum vehicle speed, and a dynamic threshold relationship mapping table between output steering angle and weather impact level, it adaptively limits the drive torque to be less than the maximum drive torque and the steering angle to be less than the maximum steering angle.

[0188] The online parameter optimization module 740 is used to calculate the difference data between the multi-frame fused environmental information and the predicted environmental information at the next moment; after fusing the difference data and the cloud meteorological data, it is input into the online gradient network; the online gradient network calculates the adjustment gradient of the model parameters by comparing the weather recognition result at the current moment with the prediction result at the current moment cached by the system, minimizing the difference between the two as the loss, and updates the parameters of the encoder, decoder and weather type prediction model online based on the adjustment gradient.

[0189] Thirdly, embodiments of the present invention also provide an electronic device, comprising:

[0190] At least one processor;

[0191] Memory for storing the at least one processor-executable instruction;

[0192] The at least one processor is configured to execute the instructions to implement the method as described in the first aspect.

[0193] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A vehicle weather environment adaptive method based on online learning, characterized in that, include: The real-time environment image is acquired, and the preprocessed real-time environment image is sequentially passed through the encoder, latent space and decoder to extract information and quantize the data, and the real-time environment information is output. The real-time environmental information is then weighted and fused with the historical environmental information to obtain multi-frame fused environmental information. The multi-frame fused environmental information is input into a pre-built weather type prediction model, which outputs weather identification results, weather impact level, and environmental information prediction results for the next moment. The weather impact level is a value obtained by mapping real-time environmental information and is used to comprehensively evaluate the severity of the vehicle driving environment. Its range is from 0 to 1. Based on the weather impact level, the vehicle's drive coefficient, braking coefficient, and steering coefficient are dynamically adjusted. Combined with the driver's input requests from the drive pedal, brake pedal, and steering wheel, the driving torque, braking torque, and steering angle adapted to the current weather conditions are calculated and output. Based on a pre-built dynamic threshold relationship mapping table between output driving torque and maximum vehicle speed, and a dynamic threshold relationship mapping table between output steering angle and weather impact level, the driving torque is adaptively limited to be less than the maximum driving torque, and the steering angle is adaptively limited to be less than the maximum steering angle. Calculate the difference data between the multi-frame fused environmental information and the predicted environmental information at the next moment; After fusing the differential data and cloud-based meteorological data, the data is input into an online gradient network. The online gradient network compares the current weather identification result with the system-cached prediction result for the current time, minimizing the difference between the two as the loss, calculates the adjustment gradient of the model parameters, and updates the parameters of the encoder, decoder, and weather type prediction model online based on the adjustment gradient, including: The system pre-connects to a cloud-based meteorological service and obtains cloud-based meteorological data after passing through an environmental information network, which is used to extract and identify weather-related feature information. The differential data and the cloud-based meteorological data are weighted and fused based on the Kalman filter algorithm to filter out random noise in the differential data and obtain comprehensive data, wherein the weight of the cloud-based meteorological data is less than the weight of the differential data. The online gradient network constructs a loss function based on the comprehensive data with the goal of improving the accuracy of weather forecasting, and calculates the gradient of the loss function with respect to the model parameters through the backpropagation algorithm to obtain the loss gradient; the model parameters include weight parameters and bias parameters. Based on a preset learning rate, the loss gradient is backpropagated using a gradient descent algorithm to update the weights and bias parameters of the encoder, decoder, and weather type prediction model.

2. The method according to claim 1, characterized in that, The preprocessed real-time environment image is sequentially passed through an encoder, latent space, and decoder for information extraction and quantization, outputting real-time environment information, including: The real-time environmental image acquired by the camera is preprocessed, including: filtering out noise caused by rain, snow and reflected light in the real-time environmental image based on image processing algorithms, while enhancing the details and edge information of the image; finding the image region related to weather judgment through the region of interest extraction algorithm and making a decision, and outputting the preprocessed real-time environmental image; The real-time environment image is input into the encoder, and the pixel matrix of the real-time environment image is converted into a high-dimensional feature vector. The high-dimensional feature vector is used to characterize the spatial features of the real-time environment image, and key environmental features are extracted from the high-dimensional feature vector in the latent space. The key environmental features are reduced in dimensionality and normalized by a decoder, and then mapped and quantized into real-time environmental information. At the same time, the real-time environmental information is cached in the system as historical environmental information.

3. The method according to claim 1, characterized in that, The multi-frame fused environmental information is the first environmental information quantity obtained by weighted fusion of historical environmental information quantities from time (tk) to time t; The calculation of the difference data between the multi-frame fused environmental information and the predicted environmental information at the next moment includes: Based on the weather type prediction model, the multi-frame fused environmental information at time (t-1) is classified and predicted to obtain the environmental information prediction result at time t, which is used as the second environmental information quantity. The error between the first environmental information quantity and the second environmental information quantity is calculated to obtain differential data.

4. The method according to claim 1, characterized in that, Based on the aforementioned weather impact level, the vehicle's drive coefficient, braking coefficient, and steering coefficient are dynamically adjusted. Combined with the driver's input requests via the drive pedal, brake pedal, and steering wheel, the driving torque, braking torque, and steering angle adapted to the current weather conditions are calculated and output, including: Based on the drive pedal input request and the pedal drive force curve, the initial drive torque is obtained within the motor performance safety limits; combined with the drive coefficient determined by the weather influence level and the tire adhesion coefficient, the initial drive torque is optimized to obtain the final output drive torque; Based on the brake pedal input request and the pedal braking force curve, the initial braking torque is obtained within the safety limits of the braking system; combined with the braking coefficient determined by the weather impact level, the initial braking torque is optimized to obtain the final output braking torque; Based on the steering wheel input request and steering force curve, the initial steering angle is obtained within the steering performance safety limit; combined with the steering coefficient determined by the weather impact level, the initial steering angle is optimized to obtain the final output steering angle.

5. The method according to claim 1, characterized in that, The weather impact level is a value between 0 and 1, which is a combination of environmental visibility and road adhesion coefficient. Among them, environmental visibility is determined by rainfall and snowfall, fog concentration gradient and light intensity gradient; road adhesion coefficient is determined by road quality, road surface water mark distribution, water depth, snow depth or ice thickness. When the weather impact level is close to 0, it indicates a severe environment. Driving stability can be improved by reducing the driving coefficient, increasing the braking coefficient, and reducing the steering coefficient. When the weather impact level is close to 1, indicating a favorable environment, the vehicle's normal handling performance can be restored by increasing the drive coefficient, decreasing the braking coefficient, and increasing the steering coefficient.

6. The method according to claim 1, characterized in that, The pre-built dynamic threshold relationship mapping tables between output driving torque and maximum vehicle speed, and between output steering angle and weather impact level, include: The first mapping relationship is based on the real-time coupling of weather impact level and vehicle power system. The maximum speed limit is set for different weather impact levels and mapped to the maximum driving torque threshold to adapt to road adhesion and reserve reaction time for sudden situations in low visibility. The second mapping relationship is based on the real-time coupling of weather impact level and steering system, setting the maximum allowable steering angle threshold for different weather impact levels to adapt to the road surface lateral adhesion and steering sensitivity under different weather conditions.

7. The method according to claim 1, characterized in that, The weather type prediction model employs a long short-term memory network to learn weather change trends from the multi-frame fused environmental information and, combined with the contextual features of cloud-based meteorological data, outputs weather identification and short-term prediction results.

8. A vehicle weather environment adaptive device based on online learning, characterized in that, include: The environmental image data processing module is used to acquire real-time environmental images, and sequentially pass the pre-processed real-time environmental images through an encoder, a latent space, and a decoder to extract information and perform quantization processing, outputting real-time environmental information; and then weightedly fuses the real-time environmental information with historical environmental information to obtain multi-frame fused environmental information. The weather type detection module is used to input the multi-frame fused environmental information into a pre-built weather type prediction model and output the weather identification result, the weather impact level, and the environmental information prediction result for the next moment; wherein, the weather impact level is a value obtained by mapping the real-time environmental information volume and is used to comprehensively evaluate the severity of the vehicle driving environment, and its range is from 0 to 1; The weather-mode vehicle control module is used to dynamically adjust the vehicle's drive coefficient, braking coefficient, and steering coefficient based on the weather impact level. It also calculates and outputs drive torque, braking torque, and steering angle adapted to the current weather conditions, taking into account the driver's input requests via the drive pedal, brake pedal, and steering wheel. Furthermore, based on pre-built dynamic threshold relationship mapping tables between output torque and maximum vehicle speed, and between output steering angle and weather impact level, it adaptively limits the drive torque to be less than the maximum drive torque and the steering angle to be less than the maximum steering angle. An online parameter optimization module is used to calculate the difference data between the multi-frame fused environmental information and the predicted environmental information at the next moment; after fusing the difference data with cloud-based meteorological data, it is input into an online gradient network; the online gradient network calculates the adjustment gradient of the model parameters by comparing the weather identification result at the current moment with the prediction result cached by the system, minimizing the difference between the two as the loss, and updates the parameters of the encoder, decoder, and weather type prediction model online based on the adjustment gradient, including: The system pre-connects to a cloud-based meteorological service and obtains cloud-based meteorological data after passing through an environmental information network, which is used to extract and identify weather-related feature information. The differential data and the cloud-based meteorological data are weighted and fused based on the Kalman filter algorithm to filter out random noise in the differential data and obtain comprehensive data, wherein the weight of the cloud-based meteorological data is less than the weight of the differential data. The online gradient network constructs a loss function based on the comprehensive data with the goal of improving the accuracy of weather forecasting, and calculates the gradient of the loss function with respect to the model parameters through the backpropagation algorithm to obtain the loss gradient; the model parameters include weight parameters and bias parameters. Based on a preset learning rate, the loss gradient is backpropagated using a gradient descent algorithm to update the weights and bias parameters of the encoder, decoder, and weather type prediction model.

9. An electronic device, characterized in that, include: At least one processor; Memory for storing the at least one processor-executable instruction; The at least one processor is configured to execute the instructions to implement the method as described in any one of claims 1-7.

Images