A vehicle anti-skid control method and vehicle

Abstract

This application provides a vehicle anti-skid control method and vehicle, relating to the field of vehicle anti-skid technology. The method includes acquiring time-series data including road surface images and vehicle driving data from historical moments to the current moment; determining a slip probability sequence based on the time-series data, indicating whether the vehicle is in a state of no slip, about to slip, or already slipped from the historical moment to the current moment; determining whether each pixel in the road surface image belongs to a water film region; mapping the determination results of all pixels in the road surface image to binary pixel values ​​to obtain a binary mask image; generating control parameters based on the slip probability sequence and the binary mask image; and controlling vehicle driving based on the control parameters. This application can generate accurate vehicle control parameters in real time and quickly respond to the command to control the vehicle, effectively suppressing vehicle slippage on low-friction surfaces and improving vehicle driving safety and passability on low-friction surfaces.

Description

Technical Field

[0001] This application relates to the field of vehicle anti-skid technology, and more specifically, to a vehicle anti-skid control method and a vehicle in the field of vehicle anti-skid technology. Background Technology

[0002] On low-friction surfaces such as wet grass, mossy ground, mud, and icy roads, vehicle drive wheels are highly susceptible to "stationary slippage" due to insufficient ground friction. This occurs when the vehicle starts or travels at low speeds, the drive wheels spin freely, preventing the vehicle from moving forward effectively. This phenomenon not only leads to difficulty starting and loss of power output control but can also cause serious safety risks such as steering instability, skidding, and even rollovers, especially on slopes or narrow roads. Furthermore, frequent slippage accelerates tire wear and reduces the efficiency of the electric drive system. Therefore, improving the anti-skid capability of vehicles on low-friction surfaces has become an urgent technical problem to be solved. Summary of the Invention

[0003] This application provides a vehicle anti-skid control method and a vehicle. This application can generate accurate vehicle control parameters in real time and quickly respond to the command to control the vehicle. It can not only effectively suppress the occurrence of vehicle skidding on low-friction coefficient road surfaces, but also improve the driving safety and passability of the vehicle on low-friction coefficient road surfaces.

[0004] Firstly, a vehicle anti-skid control method is provided, comprising: acquiring first time-series data and a first road surface image while the vehicle is in motion, wherein the first time-series data includes vehicle driving data from historical moments to the current moment; based on vehicle driving data at each moment in the first time-series data and the moments preceding that moment, using a temporal convolutional network in a parameter generation model to evaluate the confidence of the vehicle being in a first to third slip state at each moment, obtaining a slip probability sequence corresponding to the first time-series data, wherein the slip probability sequence includes the probability of the vehicle being in a first to third slip state at each moment, the first slip state being no slip, the second slip state being about to slip, and the third slip state being slipped; using a U-shaped network in the parameter generation model to determine whether each pixel in the first road surface image belongs to a water film region, mapping the determination results of all pixels to binary pixel values, obtaining a binary mask image corresponding to the first road surface image; using a multilayer perceptron network in the parameter generation model to map the slip probability sequence and the binary mask image to a first control parameter for vehicle control; and controlling vehicle driving based on the first control parameter.

[0005] Based on the above technical solution, after obtaining the slip probability sequence through a temporal convolutional network and the binary mask image through a U-shaped network, the multilayer perceptron network combines the slip probability sequence and the binary mask image to predict the risk of skidding during future vehicle operation. Since the slip probability sequence accurately describes the probability distribution of a vehicle being in a "no slip," "about to slip," or "already slipped" state at each moment from the past to the present, and the binary mask image clearly reflects whether each area on the road ahead is a water film region through pixel values, the system can predict the risk of skidding during future vehicle operation. When it is predicted that the vehicle is about to pass through a water film region and the probability of slipping increases significantly, the system can generate active anti-skid control parameters in advance. When it is predicted that the vehicle is about to pass through a safe adhesion region that is not a water film region and the probability of slipping remains low, sufficient power output can be maintained to ensure vehicle passability. Therefore, based on the slip probability sequence and binary mask image, the system can generate accurate vehicle control parameters in real time and quickly respond to the command to control the vehicle. This not only effectively suppresses the occurrence of vehicle slippage on low-friction surfaces, but also improves the driving safety and passability of vehicles on low-friction surfaces.

[0006] In one possible implementation, the temporal convolutional network includes N causal dilated convolutional modules and a first output layer. The first to Nth causal dilated convolutional modules are connected sequentially, and the Nth causal dilated convolutional module is connected to the first output layer. In any two adjacent causal dilated convolutional modules, the input of the preceding causal dilated convolutional module is connected to the input of the following causal dilated convolutional module, where N≥6. Based on the vehicle driving data at each time step and the time steps preceding that time step in the first temporal data, the temporal convolutional network in the parameter generation model is used to evaluate the confidence that the vehicle is in the first to third slip states at each time step, obtaining the slip probability sequence corresponding to the first temporal data. This includes: inputting the first temporal data into the first causal dilated convolutional module, obtaining the input of the first causal dilated convolutional module... The process involves: merging the output of the first causal dilated convolutional module with the first time-series data, inputting this data into the second causal dilated convolutional module, and obtaining the output of the second causal dilated convolutional module; merging the output of the second causal dilated convolutional module with its input data, inputting this data into the third causal dilated convolutional module, and so on, until the output of the Nth causal dilated convolutional module and its input data are merged and input into the first output layer; and then performing linear transformation and probabilistic processing on the fused data of the Nth causal dilated convolutional module's output and input data to obtain a sliding probability sequence.

[0007] In one possible implementation, the U-shaped network in the parameter generation model is used to determine whether each pixel in the first road surface image belongs to the water film region, and the determination results of all pixels are mapped to binary pixel values ​​to obtain the binary mask image corresponding to the first road surface image. This includes: using the U-shaped network to perform color enhancement on the first road surface image to obtain an enhanced image; determining whether each pixel in the enhanced image belongs to the water film region, and mapping the determination results of all pixels in the enhanced image to binary pixel values ​​to obtain the binary mask image.

[0008] In one possible implementation, the U-shaped network includes an encoder, a decoder, and a second output layer connected in sequence. The encoder includes M encoding modules, with the first to Mth encoding modules connected sequentially. The number of convolutional kernels in the first to Mth encoding modules increases sequentially. Each encoding module from the first to the (M-1)th encoding module includes two convolutional activation blocks, an activation function, and a pooling layer, where M ≥ 2. The decoder includes M-1 decoding modules, with the first to the (M-1)th decoding modules connected sequentially. The number of convolutional kernels in the first to the (M-1)th decoding modules decreases sequentially. Each decoding module includes two convolutional activation blocks and... Activation function; the Mth encoding module is connected to the 1st decoding module, the (M-1)th decoding module is connected to the second output layer, and the outputs from the (M-1)th decoding module to the 1st decoding module are connected one-to-one with the outputs from the 1st encoding module to the (M-1)th encoding module; determining whether each pixel in the enhanced image belongs to the water film region, mapping the determination results of all pixels in the enhanced image to binary pixel values, and obtaining a binary mask image includes: inputting the enhanced image into the 1st encoding module to obtain the output result of the (M-1)th decoding module; using the second output layer to perform linear transformation and probabilistic processing on the output result of the (M-1)th decoding module to obtain a binary mask image.

[0009] In one possible implementation, mapping the slip probability sequence and binary mask image to a first control parameter for vehicle control using a multilayer perceptron network in the parameter generation model includes: using a multilayer perceptron network to perform feature stitching on the slip probability sequence and binary mask image to obtain stitched features; using a multilayer perceptron network to map the stitched features to selection probabilities of multiple candidate anti-skid modes; determining the target anti-skid mode from the multiple candidate anti-skid modes based on the selection probabilities; and obtaining the first control parameter based on the target anti-skid mode.

[0010] In one possible implementation, the multilayer perceptron network includes K fully connected layers and a third output layer. The first to the Kth fully connected layers are sequentially connected through an activation function layer, and the Kth fully connected layer is connected to the third output layer through an activation function layer, where K ≥ 3. Mapping the stitched features to the selection probabilities of multiple candidate anti-skid modes based on the multilayer perceptron network includes: inputting the stitched features into the first fully connected layer, where the K fully connected layers and the activation function layer score the matching degree between the multiple candidate anti-skid modes and the road ahead of the vehicle based on the stitched features, obtaining matching degree scores for the multiple candidate anti-skid modes; and inputting the matching degree scores of the multiple candidate anti-skid modes into the third output layer through the activation function layer between the Kth fully connected layer and the third output layer, where the third output layer performs probabilistic processing on the matching degree scores of the multiple candidate anti-skid modes to obtain the selection probabilities of the multiple candidate anti-skid modes.

[0011] In one possible implementation, after controlling the vehicle's movement based on the first control parameter, the vehicle anti-skid control method further includes: acquiring multiple combined data, each combined data including a second time-series data, a second road surface image, and a second control parameter, wherein the generation time of the multiple combined data is after the generation time of the first time-series data, the first road surface image, and the first control parameter, and the time period corresponding to the second time-series data in each combined data is different; for each combined data, inputting the second time-series data in the combined data into a temporal convolutional network to obtain the first intermediate layer feature of the temporal convolutional network, the first intermediate layer feature being the output result of the Nth causal dilated convolutional module; inputting the second road surface image of the combined data into a U-shaped network to obtain a U-shaped... The second intermediate layer feature of the network is the output of the (M-2)th decoding module. The first intermediate layer feature and the second intermediate layer feature are fused to obtain the fused prior feature corresponding to the combined data. The average value of the fused prior features corresponding to multiple combined data is determined to obtain the average prior feature. The similarity between the average prior feature and multiple preset road scene features is determined to obtain multiple similarities. If the minimum similarity among the multiple similarities is less than a preset threshold, the network parameters of the multilayer perceptron network are adjusted according to the multiple combined data to obtain the adjusted multilayer perceptron network. The adjusted multilayer perceptron network is used as the multilayer perceptron network, and the steps of acquiring the first time-series data and the first road surface image are performed when the vehicle is in motion.

[0012] In one possible implementation, the network parameters include the weights and bias vectors of the third output layer in the multilayer perceptron network. Adjusting the network parameters of the multilayer perceptron network based on multiple combined data includes: for each combined data, determining the parameter error between the second control parameter in the combined data and the reference control parameter corresponding to the combined data, obtaining the parameter errors corresponding to multiple combined data, where the reference control parameter is the control parameter triggered by the user operating the vehicle under the second time-series data and the second road surface image in the combined data; determining the first gradient and the second gradient corresponding to the multiple combined data based on the parameter errors corresponding to the multiple combined data, where the first gradient is the product of the parameter error and the weights, and the second gradient is the product of the parameter error and the bias vector; determining the first average gradient of the first gradient corresponding to the multiple combined data, and the second average gradient of the second gradient corresponding to the multiple combined data; subtracting the weights from the first product to obtain the adjusted weights, where the first product is the product of the first average gradient and the learning rate; subtracting the bias vector from the second product to obtain the adjusted bias vector, where the second product is the product of the second average gradient and the learning rate.

[0013] In one possible implementation, the network parameters include the weights and bias vectors of the third output layer in the multilayer perceptron network. Adjusting the network parameters of the multilayer perceptron network based on multiple combined data includes: for each combined data, fusing the second time-series data and the second road surface image in the combined data to obtain the fused data features corresponding to the combined data; determining the average data features of the fused data features corresponding to multiple combined data; obtaining the preset weights and preset bias vectors of the average data feature mapping; replacing the weights with preset weights, and replacing the bias vectors with preset bias vectors.

[0014] Secondly, a vehicle anti-skid control device is provided, the vehicle anti-skid control device comprising: The data acquisition module is used to acquire first time-series data and first road surface image when the vehicle is in motion. The first time-series data includes vehicle driving data from historical time to the current time. The first processing module is used to evaluate the confidence of the vehicle being in the first to third slip states at each time step based on the vehicle driving data at each time step in the first time series data and the time steps before that time step, using the temporal convolutional network in the parameter generation model, to obtain the slip probability sequence corresponding to the first time series data. The slip probability sequence includes the probability of the vehicle being in the first to third slip states at each time step, where the first slip state is no slip, the second slip state is about to slip, and the third slip state is already slipped. The second processing module is used to use the U-shaped network in the parameter generation model to determine whether each pixel in the first road surface image belongs to the water film region, and to map the judgment results of all pixels into binary pixel values ​​to obtain the binary mask image corresponding to the first road surface image. The third processing module is used to map the slip probability sequence and the binary mask image into the first control parameters for vehicle control using the multilayer perceptron network in the parameter generation model. The vehicle control module is used to control the vehicle's movement based on the first control parameters.

[0015] Thirdly, a vehicle is provided, including a memory and a processor. The memory is used to store executable program code, and the processor is used to call and run the executable program code from the memory, causing the vehicle to perform the vehicle anti-skid control method of the first aspect or any possible implementation thereof.

[0016] Fourthly, a computer program product is provided, comprising: computer program code, which, when executed on a computer, causes the computer to perform the vehicle anti-skid control method of the first aspect or any possible implementation thereof.

[0017] Fifthly, a computer-readable storage medium is provided that stores computer program code, which, when executed on a computer, causes the computer to perform the vehicle anti-skid control method of the first aspect or any possible implementation thereof. Attached Figure Description

[0018] Figure 1 A schematic flowchart of a vehicle anti-skid control method provided in an embodiment of this application is shown; Figure 2 An exemplary example diagram of a parameter generation model provided in this application is shown; Figure 3 An example schematic diagram of a binary mask image provided in an embodiment of this application is shown; Figure 4 An exemplary example diagram of a temporal convolutional network provided in this application is shown. Figure 5 An exemplary example diagram of a U-shaped network provided in this application embodiment is shown; Figure 6 An exemplary example diagram of a multilayer perceptron network provided in this application is shown; Figure 7 This paper shows a schematic diagram of the structure of a vehicle anti-skid control device provided in an embodiment of this application; Figure 8 A schematic diagram of the structure of a vehicle provided in an embodiment of this application is shown. Detailed Implementation

[0019] The technical solutions in this application will be clearly and thoroughly described below with reference to the accompanying drawings. In the description of the embodiments of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. "And / or" in the text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of the embodiments of this application, "multiple" refers to two or more than two.

[0020] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0021] On low-friction surfaces such as wet grass, mossy ground, mud, and icy roads, vehicle drive wheels are highly susceptible to "stationary slippage" due to insufficient ground friction. This occurs when the vehicle starts or travels at low speeds, the drive wheels spin freely, preventing the vehicle from moving forward effectively. This phenomenon not only leads to difficulty starting and loss of power output control but can also cause serious safety risks such as steering instability, skidding, and even rollovers, especially on slopes or narrow roads. Furthermore, frequent slippage accelerates tire wear and reduces the efficiency of the electric drive system. Therefore, improving the anti-skid capability of vehicles on low-friction surfaces has become an urgent technical problem to be solved.

[0022] Based on the above problems, this application provides a vehicle anti-skid control method and a vehicle, which aims to improve the anti-skid capability of the vehicle on low-friction road surfaces, so as to effectively suppress skidding and thus ensure the driving safety and passability of the vehicle on low-friction road surfaces.

[0023] The following is an embodiment of a vehicle anti-skid control method provided in this application specification.

[0024] Figure 1 A schematic flowchart of a vehicle anti-skid control method provided in an embodiment of this application is shown, such as... Figure 1 As shown, the vehicle anti-skid control method provided in this application embodiment can be executed by a vehicle domain controller, and the vehicle anti-skid control method includes the following schemes: S110: Acquire first time-series data and first road surface image while the vehicle is in motion.

[0025] In an exemplary embodiment, during vehicle operation, first time-series data and a first road surface image are acquired. The first time-series data includes vehicle driving data from historical moments to the current moment. This vehicle driving data is multi-dimensional data, including at least one of wheel spin rate, longitudinal acceleration, yaw rate, and wheel-end vibration information (such as vertical vibration acceleration, lateral vibration acceleration, longitudinal vibration acceleration, vibration frequency, etc.). The vehicle driving data in the first time-series data is sequential, i.e., arranged in chronological order. For example, t represents the current moment, and t-1, t-2, and so on, all belonging to historical moments. The historical moments are limited to all sampling moments from the initial moment of vehicle startup to the current moment. For example, if t-1 and t-2 are selected as historical moments, then the first time-series data includes vehicle driving data at t-2, t-1, and t.

[0026] The first road surface image is the image of the road surface in front of the vehicle, that is, the image of the road surface that the vehicle is about to pass but has not yet passed. The first road surface image is captured by the forward-facing camera. In order to balance computing resources and segmentation machine accuracy, the image resolution is set to a preset resolution, such as 640×480. In addition, for the subsequent judgment of the reflection of the water film on the road surface, the first road surface image retains the RGB three-channel information.

[0027] S120: Based on the vehicle driving data at each time step in the first time series data and the time steps before that time step, the temporal convolutional network in the parameter generation model is used to evaluate the confidence of the vehicle being in the first to third slip states at each time step, and the slip probability sequence corresponding to the first time series data is obtained.

[0028] A parameter generation model can be understood as a model used to generate vehicle anti-skid control parameters. For example... Figure 2 As shown, Figure 2 The illustration shows an exemplary example diagram of a parameter generation model provided by an embodiment of this application. The parameter generation model includes a Temporal Convolutional Network (TCN), a U-shaped Network (U-Net), and a Multilayer Perceptron (MLP). The outputs of the Temporal Convolutional Network and the U-shaped Network are connected to the input of the Multilayer Perceptron Network.

[0029] After acquiring the first time-series data, the first time-series data is input into a temporal convolutional network. The temporal convolutional network evaluates the confidence of the vehicle being in the first to third slip states at each time point from the historical time to the current time based on the vehicle driving data (multidimensional data) at each time point in the first time-series data and the time points before that time point. The confidence of the vehicle being in the first to third slip states at each time point is obtained. The confidence is used to represent the probability of the slip state, and a slip probability sequence with the same time length as the first time-series data is obtained. The slip probability sequence includes the probability of the vehicle being in the first to third slip states at each time point. The first slip state is no slip, the second slip state is about to slip, and the third slip state is already slipped.

[0030] The process for determining whether a vehicle is in the first to third slip states is as follows: For a vehicle in the first slip state: if the wheel slip rate is lower than the first slip rate threshold, and the deviations of the vertical vibration acceleration, lateral vibration acceleration and longitudinal vibration acceleration at the wheel end from their respective steady-state calibration baseline values ​​are all within the preset tolerance range, and the absolute value of the yaw rate is less than the preset amplitude threshold, then it is determined that the vehicle has not slipped, that is, the vehicle is in the first slip state.

[0031] For a vehicle in a second slip state: if the wheel slip rate is between a first slip rate threshold and a second slip rate threshold (where the second slip rate threshold is greater than the first slip rate threshold), and at least one of the vertical, lateral, or longitudinal vibration accelerations at the wheel end deviates from the corresponding steady-state calibration baseline value beyond a preset tolerance range and shows an upward trend; or, the absolute value of the yaw rate exceeds a preset amplitude threshold, and the Pearson correlation coefficient between the yaw rate and the longitudinal acceleration within a first preset time period is lower than a first correlation threshold, then it is determined that the vehicle is about to slip, i.e., the vehicle is in a second slip state.

[0032] For a vehicle in a third slip state: if the wheel slip rate exceeds the second slip rate threshold and the duration is not less than the second preset duration, and the instantaneous peak value of at least one of the vertical, lateral or longitudinal vibration accelerations at the wheel end exceeds the preset acceleration threshold, and the Pearson correlation coefficient between the yaw rate and the longitudinal acceleration within the first preset duration is lower than the second correlation threshold, then the vehicle is determined to have slipped, i.e., the vehicle is in a third slip state, where the first correlation threshold is greater than the second correlation threshold.

[0033] Assume that the historical time to the current time is t-4, t-3, t-2, t-1, and t, respectively. For example, the first time-series data includes vehicle driving data from t-4 to t. After inputting the first time-series data into a temporal convolutional network, the network evaluates the confidence that the vehicle is in the first to third slip states at t-4 based on the vehicle driving data at t-4; the network evaluates the confidence that the vehicle is in the first to third slip states at t-3 based on the vehicle driving data at t-4 and t-3; the network evaluates the confidence that the vehicle is in the first to third slip states at t-2 based on the vehicle driving data from t-4 to t-1; and the network evaluates the confidence that the vehicle is in the first to third slip states at t-1 based on the vehicle driving data from t-4 to t-2. The system uses vehicle driving data at time t-1 to assess the confidence that the vehicle is in the first to third slip states at time t-1. A temporal convolutional network uses vehicle driving data from time t-4 to time t to assess the confidence that the vehicle is in the first to third slip states at time t. This yields the confidence that the vehicle is in the first to third slip states from time t-4 to time t. Then, based on the confidence that the vehicle is in the first to third slip states from time t-4 to time t, a sequence of probabilities of the vehicle being in the first to third slip states from time t-4 to time t is generated, i.e., the slip probability sequence. The slip probability sequence is shown in Table 1. Table 1

[0034] S130: Use the U-shaped network in the parameter generation model to determine whether each pixel in the first road surface image belongs to the water film region, and map the judgment results of all pixels to binary pixel values ​​to obtain the binary mask image corresponding to the first road surface image.

[0035] S130 and S120 are performed in parallel. After acquiring the first road surface image, the first road surface image is input into a U-shaped network. The U-shaped network determines whether each pixel in the first road surface image belongs to a water film region, obtaining the judgment results of whether all pixels in the first road surface image belong to water film regions. That is, through the judgment results, it can be intuitively determined which areas in the first road surface image are water film regions (i.e., the water layer covering the tire and the slippery road surface due to the failure of water to drain in time, which prevents the tire from directly contacting the road surface) and which areas are non-water film regions (i.e., areas other than water film regions). After obtaining the judgment results of all pixels, the judgment results of all pixels are mapped to binary pixel values. Specifically, the binary pixel values ​​include "1" and "0". "1" represents a pixel belonging to the water film area, and "0" represents a pixel not belonging to the water film area. Then, the binary pixel values ​​of "1" and "0" are concatenated to obtain a binary mask image with the same size as the first road surface image. That is, each binary pixel value in the binary mask image is used to indicate whether the area corresponding to that binary pixel value in the first road surface image is a water film area. Figure 3 An example schematic diagram of a binary mask image provided in an embodiment of this application is shown, such as... Figure 3 As shown, for example, a binary pixel value of "1" indicates that the area it corresponds to in the first road surface image is a water film area, and a pixel value of "0" indicates that the area it corresponds to in the first road surface image is a non-water film area.

[0036] S140: The multilayer perceptron network in the parameter generation model is used to map the slip probability sequence and the binary mask image into the first control parameter for vehicle control; S150: Control vehicle movement based on the first control parameter.

[0037] The slip probability sequence provides the trend of a vehicle's slip state from historical moments to the current moment, encompassing three slip states: no slip, about to slip, and already slipped. This reflects the temporal dynamics of slip risk. The binary mask image identifies the presence and spatial distribution of water film areas on the road surface ahead, indicating spatial information about road adhesion conditions. While the slip probability sequence alone can identify the vehicle's current slip state, it lacks the ability to predict road conditions ahead, making precise early intervention difficult to avoid potential hazards. Conversely, while the binary mask image can pinpoint the location of water-covered sections and identify potentially low-adhesion sections ahead, it cannot accurately determine whether the vehicle is currently slipping or about to slip, potentially leading to false triggering or delayed response. Furthermore, in practice, precise vehicle control is essential to ensure user safety. Since neither slip probability sequences nor binary mask images can generate accurate vehicle control parameters, combining these two methods is crucial for generating timely and accurate vehicle control parameters. This ensures driving safety and stability even on low-adhesion surfaces. Therefore, after obtaining the slip probability sequence and binary mask image, they are input into a multilayer perceptron network. The multilayer perceptron network uses a nonlinear mapping method to map the slip probability sequence and binary mask image into the first control parameter for vehicle control.

[0038] Understandably, since the slip probability sequence explicitly records the probabilities of the first to third slip states of the vehicle at each time point from the historical moment to the current moment, and each pixel value in the binary mask image represents whether its corresponding area is a water film area, the combination of these two can predict whether the vehicle will slip when continuing to move forward, and how to control the vehicle when slipping occurs, and how to control the vehicle when it will not slip. Then, based on the slip probability sequence and the binary mask image, first control parameters (such as powertrain control parameters, braking system control parameters, steering system control parameters, user interaction commands, etc.) are generated to control the vehicle's movement. Finally, the vehicle is controlled based on these first control parameters to ensure that it smoothly and successfully passes through the road ahead.

[0039] Based on the technical solution formed in S110 to S130 above, after the temporal convolutional network obtains the slip probability sequence and the U-shaped network obtains the binary mask image, the slip probability sequence accurately describes the probability distribution of the vehicle being in "no slip," "about to slip," and "already slipped" at each moment from the historical moment to the current moment. The binary mask image clearly reflects whether each area on the road ahead is a water film area (i.e., a low-adhesion-risk area) through pixel values. The multilayer perceptron network combines the slip probability sequence and the binary mask image to achieve early prediction of the risk of skidding during future vehicle operation. When it is predicted that the vehicle is about to pass through a water film area and the probability of slipping increases significantly, the system can generate active anti-skid control parameters in advance. When it is predicted that the vehicle is about to pass through a safe adhesion area that is not a water film area and the probability of slipping remains low, sufficient power output can be maintained to ensure vehicle passability. Therefore, based on the slip probability sequence and binary mask image, the system can generate accurate vehicle control parameters in real time and quickly respond to the command to control the vehicle. This not only effectively suppresses the occurrence of vehicle slippage on low-friction surfaces, but also improves the driving safety and passability of vehicles on low-friction surfaces.

[0040] In one possible implementation, the temporal convolutional network includes N causal dilated convolutional modules and a first output layer. The first to Nth causal dilated convolutional modules are connected sequentially, and the Nth causal dilated convolutional module is connected to the first output layer. Adjacent causal dilated convolutional modules are connected using a residual connection mechanism, that is, the input of the previous causal dilated convolutional module is connected to the input of the next causal dilated convolutional module. By using the residual connection mechanism, the difficulty of deep training of the model can be alleviated.

[0041] The kernel size is the same for all causal dilated convolutional modules from the first to the Nth (e.g., 3). The number of channels increases sequentially from the first to the fourth causal dilated convolutional modules, from 16 to 128. The number of channels is the same for all causal dilated convolutional modules from the fifth to the Nth, where N ≥ 6. Each causal dilated convolutional module uses a network structure that combines dilated convolution with ReLU activation and Dropout regularization, i.e., dilated convolution + ReLU activation + Dropout regularization. By designing dilated convolution, the receptive field can be expanded without increasing computational complexity, and the model's attention to key moments can also be enhanced.

[0042] The above-mentioned vehicle driving data at each time step and the time steps before that time step in the first time series data are used to evaluate the confidence of the vehicle being in the first to third slip states at each time step using a temporal convolutional network in the parameter generation model, and the slip probability sequence corresponding to the first time series data is obtained by the following steps: The first time-series data is input into the first causal dilated convolution module to obtain the output result of the first causal dilated convolution module; The output of the first causal dilated convolution module is fused with the first time series data and then input into the second causal dilated convolution module to obtain the output of the second causal dilated convolution module. The output of the second causal dilated convolutional module is fused with the input data of the second causal dilated convolutional module and input into the third causal dilated convolutional module to obtain the output of the third causal dilated convolutional module. This process is repeated until the output of the Nth causal dilated convolutional module is fused with the input data of the Nth causal dilated convolutional module and input into the first output layer. The fused data of the output of the Nth causal dilated convolutional module and the input data of the Nth causal dilated convolutional module is linearly transformed and probabilistically processed by the first output layer to obtain the glide probability sequence.

[0043] Figure 4 An exemplary example diagram of a temporal convolutional network provided in this application is shown, such as... Figure 3 and Figure 4As shown, assuming N=6, the kernel size of the 6 causal dilated convolutional modules is fixed at 3, and the number of channels of the 1st to 6th causal dilated convolutional modules are 16, 32, 64, 128, 128 and 128 respectively. After acquiring the first time-series data, it is input into the first causal dilated convolutional module to obtain its output, which has a feature dimension of 16. Then, the output of the first causal dilated convolutional module is fused with the first time-series data to obtain first fused feature data with a feature dimension of 16. This first fused feature data is then input into the second causal dilated convolutional module to obtain its output, which has a feature dimension of 32. Next, the output of the second causal dilated convolutional module is fused with its input data (i.e., the first fused feature data) to obtain second fused feature data with a feature dimension of 32. This second fused feature data is then input into the third causal dilated convolutional module to obtain the third causal dilated convolutional module. The output of the third causal dilated convolutional module has a feature dimension of 32, and so on, until the output of the sixth causal dilated convolutional module is obtained, which has a feature dimension of 126. Then, the output of the sixth causal dilated convolutional module is fused with the input data of the sixth causal dilated convolutional module (i.e., the fifth fused feature data) to obtain the sixth fused feature data with a feature dimension of 128, and the sixth fused feature data is input into the first output layer. The first output layer can be understood as a classification head, which first performs a linear transformation on the sixth fused feature data, and then performs probabilistic processing on the linear transformation result to obtain the probability of the first to third slip states of the vehicle at each time step, thereby realizing the identification of the slip state of the vehicle at each time step and generating a slip probability sequence, as shown in Table 1 above.

[0044] One possible implementation is, such as Figure 3 As shown, the U-shaped network in the parameter generation model is used to determine whether each pixel in the first road surface image belongs to the water film region, and the determination results of all pixels are mapped to binary pixel values ​​to obtain the binary mask image corresponding to the first road surface image. The steps include: A U-shaped network is used to enhance the color of the first road surface image to obtain an enhanced image; Determine whether each pixel in the enhanced image belongs to the water film region, and map the determination results of all pixels in the enhanced image to binary pixel values ​​to obtain a binary mask image.

[0045] The U-shaped network designed in this application is a lightweight network used to segment water film areas in images, identify the distribution range of water film in areas such as grass and moss, and use the identified water film areas as a basis for judging whether there is a risk of vehicle skidding.

[0046] After inputting the first road surface image into the U-shaped network, the network first performs color enhancement on the image to obtain an enhanced image. Then, it determines whether each pixel in the enhanced image belongs to a water film region, obtaining the determination results for all pixels in the enhanced image. This allows for a direct visual identification of which regions in the enhanced image are water film regions and which are not. After obtaining the determination results for all pixels, these results are mapped to binary pixel values. Specifically, the binary pixel values ​​include "1" and "0". "1" represents pixels belonging to water film regions, and "0" represents pixels not belonging to water film regions. Then, the binary pixel values ​​for "1" and "0" are concatenated to obtain a binary mask image. By performing color enhancement on the first road surface image, the color contrast of various regions in the image can be improved, thereby enhancing the model's ability to recognize water films.

[0047] In one possible implementation, since the water film in the image has a certain degree of ambiguity, it is often accompanied by phenomena such as specular reflection, sudden brightness changes, and local color saturation changes. In order to improve the model's ability to recognize such ambiguous targets, the U-shaped network involved in this application adopts a symmetrical encoder-decoder structure and performs lightweight pruning on the original U-shaped network structure, retaining 4 layers of convolution downsampling paths. The number of convolution kernels in each layer is 16, 32, 64, and 128 respectively. A combination of 3×3 convolution kernels and ReLU activation function is used. The decoder part uses transposed convolution for upsampling, while retaining the skip-connection with the encoder to ensure the effective recovery of high-resolution information.

[0048] The U-shaped network designed in this application includes an encoder, a decoder, and a second output layer (including a 1×1 convolution and a sigmoid activation function) connected in sequence. The encoder includes M encoding modules, with the first to Mth encoding modules connected sequentially. The number of convolutional kernels in the first to Mth encoding modules increases sequentially. Each encoding module in the first to M-1th encoding modules includes two convolutional activation blocks, an activation function (ReLU), and a pooling layer (e.g., MaxPooling), where M≥2. The decoder includes M-1... The decoding modules are connected sequentially from the 1st to the (M-1th)th decoding modules, with the number of convolutional kernels decreasing sequentially. Each decoding module includes two convolutional activation blocks and an activation function. For example, a convolutional activation block is a 3×3 convolution. The Mth encoding module is connected to the 1st decoding module, and the (M-1th)th decoding module is connected to the second output layer. The outputs of the (M-1th)th to the 1st decoding module are connected in a one-to-one correspondence with the outputs of the 1st to the (M-1th)th encoding modules.

[0049] Figure 5 An exemplary example diagram of a U-shaped network provided in this application is shown, as illustrated in the figure. Figure 5 As shown, assuming M=4, the encoder includes four sequentially connected encoding modules (1 through 4), with convolutional kernels of 16, 32, 64, and 128 respectively. The decoder includes three sequentially connected decoding modules (1 through 3), with convolutional kernels of 64, 32, and 16 respectively. The output of the 3rd decoding module is connected to the output of the 1st encoding module, the output of the 2nd decoding module is connected to the output of the 2nd encoding module, the output of the 1st decoding module is connected to the output of the 3rd encoding module, and the output of the 4th encoding module is connected to the second output layer.

[0050] The above process determines whether each pixel in the enhanced image belongs to the water film region. The determination results for all pixels in the enhanced image are then mapped to binary pixel values ​​to obtain a binary mask image, including: The enhanced image is input into the first encoding module to obtain the output of the (M-1)th decoding module; The output of the (M-1)th decoding module is linearly transformed and probabilistically processed by the second output layer to obtain a binary mask image.

[0051] The U-shaped network designed in this application is a lightweight network used to segment water film areas in images, identify the distribution range of water film in areas such as grass and moss, and use the identified water film areas as a basis for judging whether there is a risk of vehicle skidding.

[0052] The enhanced image obtained above has the same resolution as the first road surface image, for example, 640×480. After obtaining the enhanced image, its resolution is preprocessed to 320×240 to obtain a preprocessed image. This preprocessed image is then input into the first encoding module to obtain the output of the third decoding module. The encoder is used for image downsampling. Each of the first to fourth encoding modules includes two 3×3 convolutions + ReLU + 2×2 max pooling. Max pooling can halve the spatial size of the image. The number of channels in the output of the first to fourth encoding modules are 16, 32, 64, and 128, respectively. The decoder is used for image upsampling and includes three decoding modules. Each decoding module first upsamples the image using transposed convolution (which doubles the image size), and then concatenates the outputs of its skip-connected encoding modules to obtain high-resolution features. These high-resolution features are then processed by two 3×3 convolutions + ReLU in the next decoding module to obtain the outputs of the three decoding modules. The number of channels in the outputs of the three decoding modules are 64, 32 and 16 respectively, meaning that the number of channels in the output of the third decoding module is 16.

[0053] After obtaining the output of the third decoding module, the output of the third decoding module is input into the second output layer. The second output layer uses a 1×1 convolution to compress the number of channels of the output of the third decoding module to 1. Then, the output of the third decoding module with 1 channel is probabilistically processed by the Sigmoid activation function, so that the output value of the output of the third decoding module is between [0,1], thereby generating a binary probability mask image with the same resolution as the first road surface image.

[0054] In one possible implementation, the multilayer perceptron network designed in this application is used to jointly map the slip probability sequence and the water film distribution features in the image into specific vehicle control parameters to achieve adaptive response to different slippery surfaces. The design goal of this multilayer perceptron network is to directly output vehicle control parameters that can be used to adjust anti-slip control through deep fusion analysis of the slip probability sequence and the binary mask image. These parameters may include target anti-slip modes such as drive torque distribution ratio, braking intervention level, and traction control strategy number, thereby providing feedforward control signals to the powertrain or chassis control system.

[0055] The above-mentioned multilayer perceptron network in the parameter generation model maps the slip probability sequence and binary mask image into first control parameters for vehicle control, including the following steps: A multilayer perceptron network is used to stitch together the features of the slip probability sequence and the binary mask image to obtain the stitched features. Map the spliced ​​features to the selection probability of multiple candidate anti-slip patterns; The target anti-slip mode is determined from multiple candidate anti-slip modes based on selection probability; The first control parameter is obtained based on the target anti-slip mode.

[0056] After inputting the slip probability sequence and binary mask image into a multilayer perceptron network, the network is used to stitch the slip probability sequence and binary mask image together to obtain stitched features related to vehicle slip state and road environment conditions. The stitching process includes: calculating a first data set based on the slip probability sequence, which includes the average probability of the first to third slip states, the maximum gradient of the probability change of the first slip state, the maximum gradient of the probability change of the second slip state, the maximum gradient of the probability change of the third slip state, and the slip trend slope; calculating a second data set based on the binary mask image, which includes the proportion of the water film region, edge complexity, and the area of ​​the largest connected region; and normalizing the first and second data sets before stitching them together to obtain the stitched features.

[0057] After obtaining the stitched features, the multilayer perceptron network maps these features to selection probabilities for multiple candidate anti-skid modes. The highest probability among these probabilities is then used as the target probability, and the candidate anti-skid mode corresponding to the target probability is selected to obtain the target anti-skid mode. This, in turn, yields the corresponding vehicle control parameters, resulting in the first control parameters. The multiple candidate anti-skid modes are pre-set modes for controlling vehicle anti-skid, such as a first mode, a second mode, and a third mode. The first mode is suitable for slight skidding, the second mode for heavy skidding, and the third mode for severe skidding. The anti-skid level of the third mode is greater than that of the second mode, which is greater than that of the first mode. The selection probabilities can be mapped to obtain the corresponding anti-skid level. For example, a selection probability less than the minimum value of a preset interval corresponds to the anti-skid level of the first mode; a selection probability within the preset interval corresponds to the anti-skid level of the second mode; and a selection probability greater than the maximum value of the preset interval corresponds to the anti-skid level of the third mode.

[0058] In one possible implementation, the multilayer perceptron network includes K fully connected layers and a third output layer (including a softmax function). The first to the Kth fully connected layers are sequentially connected through an activation function layer (including a ReLU function). The Kth fully connected layer is connected to the third output layer through an activation function layer. A fully connected layer and an activation function layer constitute a hidden layer, where K ≥ 3. The above-mentioned selection probabilities for mapping the spliced ​​features to multiple candidate anti-slip patterns based on the multilayer perceptron network include: The stitched features are input into the first fully connected layer, and K fully connected layers and activation function layers score the matching degree between multiple candidate anti-skid modes and the road in front of the vehicle based on the stitched features, so as to obtain the matching degree scores corresponding to multiple candidate anti-skid modes. The matching scores of multiple candidate anti-slip patterns are input into the third output layer through the activation function layer between the Kth fully connected layer and the third output layer. The third output layer then performs probabilistic processing on the matching scores of multiple candidate anti-slip patterns to obtain the selection probability of multiple candidate anti-slip patterns.

[0059] Figure 6 An exemplary example diagram of a multilayer perceptron network provided in this application is shown, as illustrated in the figure. Figure 6 As shown, assuming K=3, the multilayer perceptron network includes three fully connected layers and a third output layer. The first fully connected layer is connected to the second fully connected layer through the first activation function layer, the second fully connected layer is connected to the third fully connected layer through the second activation function layer, and the third fully connected layer is connected to the third output layer through the third activation function layer. The first fully connected layer and the first activation function layer constitute the first hidden layer, the second fully connected layer and the second activation function layer constitute the second hidden layer, and the third fully connected layer and the third activation function layer constitute the third hidden layer. The dimensions of the first, second, and third hidden layers decrease sequentially, for example, 64, 32, and 16 respectively.

[0060] After obtaining the spliced ​​features, the spliced ​​features are input into the first fully connected layer. These three fully connected layers and three activation function layers score the matching degree between multiple candidate anti-skid modes and the road ahead of the vehicle based on the spliced ​​features, and obtain the matching degree scores corresponding to multiple candidate anti-skid modes. That is, the output of the third activation function layer is the matching degree scores corresponding to multiple candidate anti-skid modes. Then, the matching degree scores corresponding to multiple candidate anti-skid modes are input into the third output layer. The third output layer performs probabilistic processing on the matching degree scores corresponding to multiple candidate anti-skid modes, thereby unifying the matching degree scores corresponding to multiple candidate anti-skid modes to the range [0,1], and obtaining the selection probability of multiple candidate anti-skid modes output by the third output layer. For example, the selection probability of multiple candidate anti-skid modes output by the third output layer is: the selection probability of the first mode is P1, the selection probability of the second mode is P2, and the selection probability of the third mode is P3.

[0061] In one possible implementation, after controlling the vehicle's movement based on the first control parameter, the vehicle anti-skid control method further includes the following steps: Multiple combined data are acquired. Each combined data includes a second time series data, a second road surface image, and a second control parameter. The generation time of the multiple combined data is after the generation time of the first time series data, the first road surface image, and the first control parameter. The time period corresponding to the second time series data in each combined data is different. For each combined data, the second temporal data in the combined data is input into the temporal convolutional network to obtain the first intermediate layer feature of the temporal convolutional network. The first intermediate layer feature is the output of the Nth causal dilated convolutional module. The second road surface image in the combined data is input into the U-shaped network to obtain the second intermediate layer features of the U-shaped network. The second intermediate layer features are the output results of the (M-2)th decoding module. The features of the first intermediate layer and the features of the second intermediate layer are fused to obtain the fused prior features corresponding to the combined data; Determine the average value of the fused prior features corresponding to multiple combined data to obtain the average prior features; Determine the similarity between the average prior features and multiple preset road scene features to obtain multiple similarities; If the minimum similarity among multiple similarities is less than a preset threshold, the network parameters of the multilayer perceptron network are adjusted according to multiple combined data to obtain the adjusted multilayer perceptron network. The adjusted multilayer perceptron network is used as a multilayer perceptron network, and the steps of acquiring first time-series data and first road surface image are performed when the vehicle is in motion.

[0062] After controlling the vehicle's movement based on the first control parameter, it is determined whether the vehicle has entered a new road scenario. A new road scenario can be understood as a road scenario that is not a preset road scenario. A preset road scenario refers to the road scenario in which training data for training parameter generation models is collected, that is, the training data is all collected under the preset road scenario.

[0063] Determining whether a vehicle has entered a new road scenario involves acquiring multiple sets of combined data. Each set includes a second time-series data set, a second road surface image, and a second control parameter. The second time-series data set has the same data type as the first time-series data set, differing only in their acquisition time. Each set of combined data sets is generated after time t, and the second time-series data set within each set corresponds to a different time period. For example, the combined data set might consist of three sets: a first set, a second set, and a third set. The first set corresponds to the time period from time t to time t+1, the second set to time t+1 to time t+2, and the third set to time t+2 to time t+3.

[0064] For each combined data Di, the second temporal data in Di is input into a temporal convolutional network, and the output of the Nth causal dilated convolutional module in the temporal convolutional network is obtained to obtain the first intermediate layer feature of the temporal convolutional network; the second road surface image in Di is input into a U-shaped network, and the output of the (M-2)th decoding module in the U-shaped network is obtained to obtain the second intermediate layer feature of the U-shaped network; then the first intermediate layer feature and the second intermediate layer feature are fused to obtain the fused prior feature corresponding to the combined data Di (i.e., the fused feature of the first intermediate layer feature and the second intermediate layer feature), and thus the fused prior features corresponding to multiple combined data can be obtained.

[0065] After obtaining the fused prior features corresponding to each of the multiple combined data sets, the average of these fused prior features is calculated to obtain the average prior feature. Then, the Euclidean distance between the average prior feature and multiple preset road scene features is calculated to obtain the similarity between the average prior feature and the multiple preset road scene features, i.e., multiple similarity scores are obtained. The multiple preset road scene features are road scene features corresponding to multiple preset roads obtained during model training, and they serve as parameters for determining whether a vehicle has entered a new road scene.

[0066] After obtaining multiple similarities, the minimum similarity is compared with a preset threshold (e.g., 0.9). If the minimum similarity is less than the preset threshold, it means that the vehicle has entered a new road scene (i.e., the vehicle has entered a road scene it has not entered before). In this case, the k-shot prior transfer mechanism is triggered, which adjusts the network parameters of the multilayer perceptron network based on multiple combined data. This improves the dynamic adjustment capability of the parameter generation model under different road conditions, different grass types, and different degrees of slipperiness, thereby improving the response speed and generalization performance of the parameter generation model in new road scenes.

[0067] After obtaining the adjusted multilayer perceptron network, the adjusted multilayer perceptron network is used as the multilayer perceptron network. Then, the steps of acquiring the first time-series data and the first road surface image while the vehicle is in motion are restarted to achieve precise anti-skid control of the vehicle in the new road scenario.

[0068] In one possible implementation, the network parameters include the weights and bias vectors of the third output layer in the multilayer perceptron network. Adjusting the network parameters of the multilayer perceptron network based on multiple combined data includes the following steps: For each combination of data, determine the parameter error between the second control parameter in the combination of data and the reference control parameter corresponding to the combination of data, and obtain the parameter error corresponding to multiple combination of data. Based on the parameter errors corresponding to multiple combined data, the first gradient and the second gradient corresponding to the multiple combined data are determined. The first gradient is the product of the parameter error and the weight, and the second gradient is the product of the parameter error and the bias vector. Determine the first average gradient of the first gradient corresponding to multiple combinations of data, and the second average gradient of the second gradient corresponding to multiple combinations of data; The adjusted weights are obtained by subtracting the weights from the first product, where the first product is the product of the first average gradient and the learning rate. The adjusted bias vector is obtained by subtracting the bias vector from the second product, which is the product of the second average gradient and the learning rate.

[0069] This application proposes two parameter fine-tuning methods for adjusting the weights and bias vectors of the third output layer. The first method involves calculating the difference between the second control parameter in the combined data Di and the corresponding reference control parameter for the combined data Di, thus obtaining the parameter error for the combined data Di. The reference control parameter is pre-set and refers to the control parameter triggered by the user's vehicle operation under the second time-series data and the second road surface image in the combined data Di. Based on the calculation method of the parameter error for the combined data Di, the parameter errors for multiple combined data can be obtained. Then, the weights of the third output layer are multiplied by the parameter error of the combined data Di to obtain the first gradient for the combined data Di; and the bias vector of the third output layer is multiplied by the parameter error of the combined data Di to obtain the second gradient for the combined data Di. In this way, the first and second gradients corresponding to multiple combined data are obtained.

[0070] After obtaining the first gradient and the second gradient corresponding to multiple combinations of data, the average value of the first gradient corresponding to each of the multiple combinations of data is calculated to obtain the first average gradient; and the average value of the second gradient corresponding to each of the multiple combinations of data is calculated to obtain the second average gradient.

[0071] For adjusting the weights of the third output layer, the weights are denoted as w1, the first average gradient as g1, the adjusted weights as w2, and the learning rate as a, where w2 = w1 - g1 × a. For adjusting the bias vector of the third output layer, the bias vector is denoted as b1, the second average gradient as g2, and the adjusted bias vector as b2, where b2 = b1 - g2 × a. This achieves fine-tuning of the weights and bias vectors.

[0072] In one possible implementation, the network parameters include the weights and bias vectors of the third output layer in the multilayer perceptron network. Adjusting the network parameters of the multilayer perceptron network based on multiple combined data includes the following steps: For each combined data, the second time-series data and the second road surface image in the combined data are fused to obtain the fused data features corresponding to the combined data; Determine the average data feature of the fused data features corresponding to multiple combined data; Obtain the preset weights and preset bias vectors of the average data feature mapping; The weights are replaced with preset weights, and the bias vectors are replaced with preset bias vectors.

[0073] For the second parameter adjustment method of the weights and bias vectors of the third output layer, the second time-series data and the second road surface image from multiple combined data are input into a pre-set parameter generation network, which outputs the adjusted weights and bias vectors. The processing of the parameter generation network is as follows: the second time-series data and the second road surface image in the combined data Di are fused to obtain the fused data features corresponding to the combined data Di, thereby obtaining the fused data features corresponding to multiple combined data. Then, the average value of the fused data features corresponding to multiple combined data is calculated to obtain the average data features. Then, the preset weights and preset bias vectors of the average data feature mapping are obtained. That is, the preset weights of the average data feature mapping are the adjusted weights, and the preset bias vectors of the average data feature mapping are the adjusted bias vectors. Then, the preset weights are used to replace the weights, and the preset bias vectors are used to replace the bias vectors, thus realizing the fine-tuning of the adjusted weights and bias vectors.

[0074] It is worth noting that when adjusting the weights and bias vectors of the third output layer in a multilayer perceptron network, it is necessary to freeze all other layers in the multilayer perceptron network except for the third output layer, the U-shaped network, and the temporal convolutional network.

[0075] The training of the parameter generation model in this application is described below.

[0076] For temporal convolutional networks, the training samples come from driving data in vehicle test tracks and natural roads, including slip samples under various climates, slopes, and tire conditions. This ensures that the temporal convolutional network has good generalization performance. The loss function of the temporal convolutional network adopts the weighted cross-entropy loss function, which focuses on penalizing misjudgments of "about to slip" states, thereby improving the system's sensitivity to potentially dangerous states.

[0077] For the U-shaped network, the training samples are drawn from a wet grass image database, which includes images with varying lighting, viewing angles, grass types, climates, and background complexity. Each image in the database carries a marker indicating the presence of a water film boundary. The U-shaped network's loss function employs a weighted combination of Dice Loss and Binary Cross Entropy, thus addressing the issue of foreground-background imbalance caused by the small area of ​​the water film region.

[0078] For a multilayer perceptron network, the training data includes control parameters before and after a slip event, and these control parameters are matched one-to-one with the corresponding ground image and slip probability sequence. The loss function of a multilayer perceptron network includes mean squared error or cross-entropy.

[0079] This application integrates temporal convolutional networks, U-shaped networks, and multilayer perceptron networks into a unified training framework, and achieves deep collaborative optimization of the three-level functional model of "perception-prediction-control" through multi-source data sharing, loss function fusion, and cross-module feedback linkage mechanism.

[0080] Traditional multi-model systems often employ a phased, independent training strategy. While this ensures the convergence of each module, it suffers from insufficient overall system optimization, high data redundancy, and uncontrollable error propagation across modules. To overcome these shortcomings, this application designs an end-to-end collaborative training structure, establishing a unified feature flow and loss coordination mechanism while maintaining the independent controllability of each model architecture.

[0081] Specifically, time-series data and road surface images are used as dual-modal input sources, processed by a temporal convolutional network and a U-shaped network, respectively. The temporal convolutional network outputs a slip probability sequence (covering three states: "not slipping," "about to slip," and "slipped"), and its supervision target uses a weighted cross-entropy loss. The U-shaped network generates a binary mask image of the water film region, and its loss function is a weighted combination of Dice Loss and binary cross-entropy. The outputs of both are not only used for local supervised learning in their respective modules but are also fused into multimodal features, which are then input into a subsequent multilayer perceptron network for regressing specific anti-skid control parameters. The loss function of the multilayer perceptron network is selected based on the task nature, choosing either mean squared error or cross-entropy; that is, the total loss function of the parameter generation model is αL. TCN +βL U-Net +γL MLP α, β, and γ are hyperparameters used to control the contribution weights of the losses of the temporal convolutional network, U-shaped network, and multilayer perceptron network to the total loss, enabling the training process to adaptively balance the optimization focus of the three and effectively prevent overfitting of one module or learning stagnation of another module.

[0082] Furthermore, a feedback-based loss coupling mechanism is introduced. When the output deviation of the multilayer perceptron network increases significantly, or the slip state prediction error rate rises, the system automatically increases L. TCN With L U-Net The weights are used to drive the temporal convolutional network and the U-shaped network to improve the output accuracy, forming a closed-loop optimization logic to control the error and enhance perception.

[0083] The entire joint training process employs end-to-end backpropagation and embeds learnable attention gating structures into the feature fusion channels of the temporal convolutional network and the U-shaped network. This dynamically adjusts the relative importance of temporal features and spatial semantic features to the input of the multilayer perceptron network, achieving adaptive optimization of feature-level multimodal fusion. This architecture not only improves the anti-skid response speed and control accuracy of the system on low-adhesion road surfaces but also achieves end-to-end collaborative optimization of environmental perception, state prediction, and control decision-making through deep collaborative training, significantly improving the consistency of the entire link response and overall control performance.

[0084] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0085] Figure 7 This application provides a schematic diagram of the structure of a vehicle anti-skid control device according to an embodiment of the present application. Figure 7 As shown, the vehicle anti-skid control device 700 includes: The data acquisition module 710 is used to acquire first time-series data and first road surface image when the vehicle is in motion. The first time-series data includes vehicle driving data from historical time to the current time. The first processing module 720 is used to evaluate the confidence of the vehicle being in the first to third slip states at each time step based on the vehicle driving data at each time step and the time step before that time step in the first time series data, using the temporal convolutional network in the parameter generation model, to obtain the slip probability sequence corresponding to the first time series data. The slip probability sequence includes the probability of the vehicle being in the first to third slip states at each time step, where the first slip state is no slip, the second slip state is about to slip, and the third slip state is already slipped. The second processing module 730 is used to use the U-shaped network in the parameter generation model to determine whether each pixel in the first road surface image belongs to the water film region, and to map the judgment results of all pixels into binary pixel values ​​to obtain the binary mask image corresponding to the first road surface image. The third processing module 740 is used to map the slip probability sequence and the binary mask image into first control parameters for vehicle control using a multilayer perceptron network in the parameter generation model. The vehicle control module 750 is used to control the vehicle's movement based on the first control parameters.

[0086] In one possible implementation, the temporal convolutional network includes N causal dilated convolutional modules and a first output layer. The first to Nth causal dilated convolutional modules are connected sequentially, and the Nth causal dilated convolutional module is connected to the first output layer. The input of one causal dilated convolutional module is connected to the input of the next causal dilated convolutional module in any two adjacent modules, where N ≥ 6. A first processing module 720 is specifically used to input the first temporal data into the first causal dilated convolutional module to obtain its output; to fuse the output of the first causal dilated convolutional module with the first temporal data, and then input it into the second causal dilated convolutional module. The first causal dilated convolutional module is used to obtain the output of the second causal dilated convolutional module. The output of the second causal dilated convolutional module is then fused with the input data of the second causal dilated convolutional module and input to the third causal dilated convolutional module to obtain its output. This process is repeated until the output of the Nth causal dilated convolutional module is fused with its input data and input to the first output layer. The first output layer then performs a linear transformation and probabilistic processing on the fused data of the Nth causal dilated convolutional module's output and input data to obtain a glide probability sequence.

[0087] In one possible implementation, the second processing module 730 is specifically used to perform color enhancement on the first road surface image using a U-shaped network to obtain an enhanced image; determine whether each pixel in the enhanced image belongs to the water film region, and map the determination results of all pixels in the enhanced image to binary pixel values ​​to obtain a binary mask image.

[0088] In one possible implementation, the U-shaped network includes an encoder, a decoder, and a second output layer connected in sequence. The encoder includes M encoding modules, with the first to Mth encoding modules connected sequentially. The number of convolutional kernels in the first to Mth encoding modules increases sequentially. Each encoding module from the first to the (M-1)th encoding module includes two convolutional activation blocks, an activation function, and a pooling layer, where M ≥ 2. The decoder includes M-1 decoding modules, with the first to the (M-1)th decoding modules connected sequentially. The number of convolutional kernels in the first to the (M-1)th decoding modules increases sequentially. The order of the decoding modules decreases sequentially. Each decoding module includes two convolutional activation blocks and an activation function. The Mth encoding module is connected to the 1st decoding module, and the (M-1)th decoding module is connected to the second output layer. The outputs of the (M-1)th decoding module to the 1st decoding module are connected one-to-one with the outputs of the 1st encoding module to the (M-1)th encoding module. The second processing module 730 is specifically used to input the enhanced image into the 1st encoding module to obtain the output result of the (M-1)th decoding module. The second output layer is used to perform linear transformation and probabilistic processing on the output result of the (M-1)th decoding module to obtain a binary mask image.

[0089] In one possible implementation, the third processing module 740 is specifically used to perform feature stitching on the slip probability sequence and the binary mask image using a multilayer perceptron network to obtain stitched features; map the stitched features to the selection probabilities of multiple candidate anti-slip modes; determine the target anti-slip mode from the multiple candidate anti-slip modes based on the selection probabilities; and obtain the first control parameter based on the target anti-slip mode.

[0090] In one possible implementation, the multilayer perceptron network includes K fully connected layers and a third output layer. The first to the Kth fully connected layers are sequentially connected through an activation function layer, and the Kth fully connected layer is connected to the third output layer through an activation function layer, where K ≥ 3. The third processing module 740 is specifically used to input the spliced ​​features into the first fully connected layer, so that the K fully connected layers and the activation function layer score the matching degree between multiple candidate anti-skid modes and the road ahead of the vehicle based on the spliced ​​features, thereby obtaining matching degree scores corresponding to multiple candidate anti-skid modes. The matching degree scores corresponding to multiple candidate anti-skid modes are input into the third output layer through the activation function layer between the Kth fully connected layer and the third output layer, so that the third output layer performs probabilistic processing on the matching degree scores corresponding to multiple candidate anti-skid modes, thereby obtaining the selection probability of multiple candidate anti-skid modes.

[0091] In one possible implementation, after controlling the vehicle's movement based on the first control parameters, the vehicle anti-skid control device further includes: The parameter update unit is used to acquire multiple combined data sets. Each combined data set includes a second time-series data set, a second road surface image, and a second control parameter. The generation time of the multiple combined data sets is after the generation time of the first time-series data set, the first road surface image, and the first control parameter. The time period corresponding to the second time-series data set in each combined data set is different. For each combined data set, the second time-series data set is input into a temporal convolutional network to obtain the first intermediate layer features of the temporal convolutional network. The first intermediate layer features are the output of the Nth causal dilated convolutional module. The second road surface image is input into a U-shaped network to obtain the second intermediate layer features of the U-shaped network. The features are the output results of the (M-2)th decoding module; the first intermediate layer features and the second intermediate layer features are fused to obtain the fused prior features corresponding to the combined data; the average value of the fused prior features corresponding to multiple combined data is determined to obtain the average prior features; the similarity between the average prior features and multiple preset road scene features is determined to obtain multiple similarities; if the minimum similarity among the multiple similarities is less than a preset threshold, the network parameters of the multilayer perceptron network are adjusted according to the multiple combined data to obtain the adjusted multilayer perceptron network; the adjusted multilayer perceptron network is used as the multilayer perceptron network, and the steps of acquiring the first time-series data and the first road surface image are performed when the vehicle is in motion.

[0092] In one possible implementation, the network parameters include the weights and bias vectors of the third output layer in the multilayer perceptron network. The parameter update unit is specifically used to, for each combination of data, determine the parameter error between the second control parameter in the combination data and the reference control parameter corresponding to the combination data, obtaining multiple parameter errors corresponding to the combination data. The reference control parameter is the control parameter triggered by the user operating the vehicle under the second time-series data and the second road surface image in the combination data. Based on the parameter errors corresponding to the multiple combination data, determine the first gradient and the second gradient corresponding to the multiple combination data. The first gradient is the product of the parameter error and the weights, and the second gradient is the product of the parameter error and the bias vector. Determine the first average gradient of the first gradient corresponding to the multiple combination data, and the second average gradient of the second gradient corresponding to the multiple combination data. Subtract the weights from the first product to obtain the adjusted weights. The first product is the product of the first average gradient and the learning rate. Subtract the bias vector from the second product to obtain the adjusted bias vector. The second product is the product of the second average gradient and the learning rate.

[0093] In one possible implementation, the network parameters include the weights and bias vectors of the third output layer in the multilayer perceptron network. The parameter update unit is specifically used to, for each combined data, fuse the second time-series data and the second road surface image in the combined data to obtain the fused data features corresponding to the combined data; determine the average data features of the fused data features corresponding to multiple combined data; obtain the preset weights and preset bias vectors of the average data feature mapping; replace the weights with preset weights; and replace the bias vectors with preset bias vectors.

[0094] It should be noted that the vehicle anti-skid control device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the vehicle anti-skid control method. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the vehicle anti-skid control device and the vehicle anti-skid control method embodiments provided in the above embodiments belong to the same concept. Therefore, for details not disclosed in the device embodiments of this application, please refer to the embodiments of the vehicle anti-skid control method described above in this application, which will not be repeated here.

[0095] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0096] Figure 8 This application provides a schematic diagram of the structure of a vehicle according to an embodiment of the present application. Figure 8 As shown, the vehicle 800 includes a memory 801 and a processor 802. The memory 801 stores executable program code 8011, and the processor 802 is used to call and execute the executable program code 8011 to perform a vehicle anti-skid control method.

[0097] This embodiment can divide the vehicle into functional modules according to the above method example. For example, each function can be assigned to a separate module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware. It should be noted that the module division in this embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0098] When each functional module is divided according to its corresponding function, the vehicle may include: a data acquisition module, a first processing module, a second processing module, a third processing module, a vehicle control module, etc. It should be noted that all relevant content of each step involved in the above method embodiments can be referenced from the functional description of the corresponding functional module, and will not be repeated here.

[0099] The vehicle provided in this embodiment is used to execute the above-described vehicle anti-skid control method, and therefore can achieve the same effect as the above-described implementation method.

[0100] When using integrated units, the vehicle may include a processing module and a storage module. The processing module is used to control and manage the vehicle's movements. The storage module is used to support the vehicle in executing relevant program code and data.

[0101] The processing module may be a processor or a controller, which can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. The processor may also be a combination of functions that implement computing capabilities, such as a combination of one or more microprocessors, a combination of digital signal processing (DSP) and a microprocessor, etc., and the storage module may be a memory.

[0102] This embodiment also provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, the computer executes the above-described related method steps to implement a vehicle anti-skid control method in the above embodiment.

[0103] This embodiment also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned steps to implement a vehicle anti-skid control method as described in the above embodiment.

[0104] In addition, the vehicle provided in the embodiments of this application may specifically be a chip, component or module. The vehicle may include a connected processor and a memory. The memory is used to store instructions. When the vehicle is running, the processor can call and execute the instructions to make the chip execute a vehicle anti-skid control method in the above embodiments.

[0105] In this embodiment, the vehicle, computer-readable storage medium, computer program product, or chip are all used to execute the corresponding vehicle anti-skid control method provided above. Therefore, the beneficial effects that can be achieved can be referred to the beneficial effects in the corresponding vehicle anti-skid control method provided above, and will not be repeated here.

[0106] Through the above description of the embodiments, those skilled in the art will understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0107] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0108] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A vehicle anti-skid control method characterized by, The vehicle anti-skid control method includes: While the vehicle is in motion, first time-series data and first road surface image are acquired, the first time-series data including vehicle driving data from historical time to the current time; Based on the vehicle driving data at each time step in the first time series data and the time steps before that time step, the confidence of the vehicle being in the first to third slip states at each time step is evaluated using a temporal convolutional network in the parameter generation model, and a slip probability sequence corresponding to the first time series data is obtained. The slip probability sequence includes the probability of the vehicle being in the first to third slip states at each time step, where the first slip state is no slip, the second slip state is about to slip, and the third slip state is already slipped. The U-shaped network in the parameter generation model is used to determine whether each pixel in the first road surface image belongs to the water film region, and the determination results of all pixels are mapped to binary pixel values ​​to obtain the binary mask image corresponding to the first road surface image. The multilayer perceptron network in the parameter generation model is used to map the slip probability sequence and the binary mask image into first control parameters for vehicle control. The vehicle is controlled to move based on the first control parameter.

2. The vehicle slip control method according to claim 1, characterized by, The temporal convolutional network includes N causal dilated convolutional modules and a first output layer. The first to Nth causal dilated convolutional modules are connected sequentially. The Nth causal dilated convolutional module is connected to the first output layer. The input of the preceding causal dilated convolutional module is connected to the input of the following causal dilated convolutional module in two adjacent causal dilated convolutional modules. N≥6. The method, based on vehicle driving data at each time step and previous time steps in the first time series data, uses a temporal convolutional network in the parameter generation model to evaluate the confidence that the vehicle is in the first to third slip states at each time step, and obtains the slip probability sequence corresponding to the first time series data, including: The first time-series data is input into the first causal dilated convolution module to obtain the output result of the first causal dilated convolution module; The output of the first causal dilated convolution module is fused with the first time series data and input into the second causal dilated convolution module to obtain the output of the second causal dilated convolution module. The output of the second causal dilated convolution module is fused with the input data of the second causal dilated convolution module and input into the third causal dilated convolution module to obtain the output of the third causal dilated convolution module. This process is repeated until the output of the Nth causal dilated convolution module is fused with the input data of the Nth causal dilated convolution module and input into the first output layer. The first output layer performs a linear transformation and probabilistic processing on the fused data of the output of the Nth causal dilated convolution module and the input data of the Nth causal dilated convolution module to obtain the glide probability sequence.

3. The vehicle slip control method according to claim 2, characterized by, The step of using the U-shaped network in the parameter generation model to determine whether each pixel in the first road surface image belongs to the water film region, and mapping the determination results of all pixels to binary pixel values ​​to obtain the binary mask image corresponding to the first road surface image includes: The U-shaped network is used to enhance the color of the first road surface image to obtain an enhanced image; Determine whether each pixel in the enhanced image belongs to the water film region, and map the determination results of all pixels in the enhanced image to binary pixel values ​​to obtain the binary mask image.

4. The vehicle slip control method according to claim 3, characterized by, The U-shaped network includes an encoder, a decoder, and a second output layer connected in sequence. The encoder includes M encoding modules, with the first to the Mth encoding modules connected in sequence. The number of convolutional kernels in the first to the Mth encoding modules increases sequentially. Each encoding module in the first to the (M-1)th encoding modules includes two convolutional activation blocks, an activation function, and a pooling layer, where M ≥ 2. The decoder includes M-1 decoding modules, with the first to the (M-1)th decoding modules connected sequentially. The number of convolutional kernels in the first to the (M-1)th decoding modules decreases sequentially. Each decoding module includes two convolutional activation blocks and an activation function. The Mth encoding module is connected to the 1st decoding module, the (M-1)th decoding module is connected to the second output layer, and the outputs from the (M-1)th decoding module to the 1st decoding module are connected in a one-to-one correspondence with the outputs from the 1st encoding module to the (M-1)th encoding module. The step of determining whether each pixel in the enhanced image belongs to the water film region, and mapping the determination results of all pixels in the enhanced image to binary pixel values ​​to obtain the binary mask image includes: The enhanced image is input into the first encoding module to obtain the output of the (M-1)th decoding module; The output of the (M-1)th decoding module is linearly transformed and probabilistically processed using the second output layer to obtain the binary mask image.

5. The vehicle anti-skid control method according to claim 4, characterized in that, The process of mapping the slip probability sequence and the binary mask image to first control parameters for vehicle control using the multilayer perceptron network in the parameter generation model includes: The multilayer perceptron network is used to perform feature concatenation on the slip probability sequence and the binary mask image to obtain concatenated features; The splicing features are mapped to the selection probabilities of multiple candidate anti-slip modes; The target anti-slip mode is determined from the plurality of candidate anti-slip modes based on the selection probability; The first control parameter is obtained based on the target anti-slip mode.

6. The vehicle anti-skid control method according to claim 5, characterized in that, The multilayer perceptron network includes K fully connected layers and a third output layer. The first to the Kth fully connected layers are connected sequentially through an activation function layer. The Kth fully connected layer is connected to the third output layer through the activation function layer, where K ≥ 3. The selection probability of mapping the spliced ​​features into multiple candidate anti-slip patterns based on the multilayer perceptron network includes: The stitching features are input into the first fully connected layer, so that the K fully connected layers and the activation function layer score the matching degree between the multiple candidate anti-skid modes and the road in front of the vehicle based on the stitching features, and obtain the matching degree scores corresponding to the multiple candidate anti-skid modes; The matching scores corresponding to the multiple candidate anti-slip patterns are input into the third output layer through the activation function layer between the Kth fully connected layer and the third output layer. The third output layer then performs probabilistic processing on the matching scores corresponding to the multiple candidate anti-slip patterns to obtain the selection probability of the multiple candidate anti-slip patterns.

7. The vehicle anti-skid control method according to claim 6, characterized in that, After controlling the vehicle's movement based on the first control parameter, the vehicle anti-skid control method further includes: Multiple combined data are acquired. Each combined data includes a second time-series data, a second road surface image, and a second control parameter. The generation time of the multiple combined data is after the generation time of the first time-series data, the first road surface image, and the first control parameter. The time period corresponding to the second time-series data in each combined data is different. For each combination of data, the second temporal data in the combination of data is input into the temporal convolutional network to obtain the first intermediate layer feature of the temporal convolutional network. The first intermediate layer feature is the output result of the Nth causal dilated convolutional module. The second road surface image in the combined data is input into the U-shaped network to obtain the second intermediate layer feature of the U-shaped network. The second intermediate layer feature is the output result of the (M-2)th decoding module. The first intermediate layer features and the second intermediate layer features are fused to obtain the fused prior features corresponding to the combined data; The average value of the fused prior features corresponding to the multiple combined data is determined to obtain the average prior features; Determine the similarity between the average prior features and multiple preset road scene features to obtain multiple similarities; If the minimum similarity among the multiple similarities is less than a preset threshold, the network parameters of the multilayer perceptron network are adjusted according to the multiple combined data to obtain the adjusted multilayer perceptron network. The adjusted multilayer perceptron network is used as the multilayer perceptron network, and the steps of acquiring the first time-series data and the first road surface image when the vehicle is in motion are performed.

8. The vehicle anti-skid control method according to claim 7, characterized in that, The network parameters include the weights and bias vectors of the third output layer in the multilayer perceptron network, and adjusting the network parameters of the multilayer perceptron network according to the multiple combined data includes: For each combination of data, the parameter error between the second control parameter in the combination of data and the reference control parameter corresponding to the combination of data is determined to obtain the parameter error corresponding to the multiple combination data. The reference control parameter is the control parameter triggered by the user operating the vehicle under the second time-series data and the second road surface image in the combination of data. Based on the parameter errors corresponding to the multiple combined data, a first gradient and a second gradient corresponding to the multiple combined data are determined. The first gradient is the product of the parameter error and the weight, and the second gradient is the product of the parameter error and the bias vector. Determine the first average gradient of the first gradient corresponding to the plurality of combined data, and the second average gradient of the second gradient corresponding to the plurality of combined data; The adjusted weights are obtained by subtracting the weights from the first product, where the first product is the product of the first average gradient and the learning rate. The adjusted bias vector is obtained by subtracting the bias vector from the second product, where the second product is the product of the second average gradient and the learning rate.

9. The vehicle anti-skid control method according to claim 7, characterized in that, The network parameters include the weights and bias vectors of the third output layer in the multilayer perceptron network, and adjusting the network parameters of the multilayer perceptron network according to the multiple combined data includes: For each combined data, the second time-series data and the second road surface image in the combined data are fused to obtain the fused data features corresponding to the combined data. Determine the average data feature of the fused data features corresponding to the multiple combined data; Obtain the preset weights and preset bias vectors of the average data feature mapping; The weights are replaced with the preset weights, and the bias vector is replaced with the preset bias vector.

10. A vehicle, characterized in that, The vehicles include: Memory, used to store executable program code; A processor is configured to call and run the executable program code from the memory, causing the vehicle to perform the vehicle anti-skid control method as described in any one of claims 1 to 9.

Images