Vehicle tire pressure adjusting method and vehicle

Abstract

The invention relates to the technical field of vehicle control, and provides a vehicle tire pressure adjusting method and a vehicle. Vehicle running data and a road topographic image in front of the vehicle are obtained, so that the vehicle trapping risk of the vehicle is predicted, and vehicle trapping probability information of the vehicle is obtained; according to the method, the vehicle trapping risk of the vehicle can be predicted in advance, the accuracy of vehicle trapping risk prediction is ensured through environmental perception and operation data, and the vehicle trapping risk prediction accuracy is improved. The tire pressure adjusting quantity is determined in time according to the vehicle sinking probability information and the vehicle operation data, so that the tire pressure adjusting accuracy is ensured, the purpose of timely intervention before the vehicle sinking is achieved, the vehicle sinking caused by mismatching of the tire pressure under the complex road conditions such as sand and mud is avoided, the vehicle escaping capability is improved, and the vehicle safety is improved. And the trafficability of the vehicle under various complex road conditions is improved.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and in particular to a method for regulating tire pressure and a vehicle. Background Technology

[0002] Currently, most vehicles are equipped with a tire pressure regulation system. Existing vehicle tire pressure regulation systems mostly use fixed values ​​for adjustment. By monitoring the tire pressure in real time, the system actively adjusts the tire pressure when the pressure is lower or higher than the set standard value.

[0003] However, existing tire pressure regulation systems cannot predict the risk of getting stuck, which often causes vehicles to get stuck in complex road conditions such as sand and mud due to mismatched tire pressure. Summary of the Invention

[0004] In view of this, the purpose of this application is to propose a vehicle tire pressure adjustment method and a vehicle that can predict the risk of getting stuck and adjust the tire pressure in a timely manner, avoid the vehicle getting stuck due to tire pressure mismatch, improve the vehicle's ability to get out of trouble, and cope with various extreme and complex road conditions.

[0005] To achieve the above objectives, this application provides a method for adjusting vehicle tire pressure, comprising: Acquire vehicle operation data and road terrain images in front of the vehicle; Based on the vehicle operation data and the road terrain image, determine the probability information of the vehicle getting stuck; Based on the vehicle getting stuck probability information and the vehicle operation data, the tire pressure adjustment amount of each tire in the vehicle is determined, and the tire pressure is adjusted according to the tire pressure adjustment amount.

[0006] Optionally, based on the vehicle operation data and the road terrain image, the probability information of the vehicle getting stuck is determined, including: Based on the road terrain image and the pre-trained terrain recognition model, determine the current terrain feature vector; Based on the current terrain feature vector, the vehicle operation data, and the pre-trained vehicle getting stuck risk prediction model, the probability information of the vehicle getting stuck is determined.

[0007] Optionally, based on the current terrain feature vector, the vehicle operation data, and a pre-trained vehicle-trapping risk prediction model, the probability information of the vehicle getting stuck is determined, including: Based on the current terrain feature vector and the current operating parameters in the vehicle operating data, the current frame data is constructed, and based on the historical terrain feature vector and the historical operating parameters in the vehicle operating data, historical frame data is constructed. The current frame data and the historical frame data are input into a pre-trained vehicle getting stuck risk prediction model to obtain the vehicle's probability of getting stuck.

[0008] Optionally, based on the vehicle getting stuck probability information and the vehicle operation data, the tire pressure adjustment amount of each tire in the vehicle is determined, including: Based on the vehicle getting stuck probability information, the current operating parameters in the vehicle operating data, and the current terrain feature vector, a current comprehensive state is constructed. The current overall state is input into the policy network of a pre-trained tire pressure regulation model, so that the policy network determines the probability distribution of each tire pressure policy under the current overall state, and determines the target tire pressure policy based on the probability distribution, wherein the target tire pressure policy includes the tire pressure adjustment amount of each tire.

[0009] Optionally, the network parameters in the strategy network are obtained by training the value network in the tire pressure regulation model, and the value network is used to determine the strategy value of executing the preset tire pressure strategy under the preset comprehensive state; The strategy value includes at least one of the following: vehicle pass reward, frequent voltage adjustment penalty, smooth voltage adjustment reward, energy consumption penalty, and rapid response reward.

[0010] Optionally, in determining the current terrain feature vector, the following may also be included: Based on the road terrain image and the pre-trained terrain recognition model, a terrain segmentation map is determined, wherein the terrain segmentation map is used to describe the terrain category corresponding to each pixel in the road terrain image; After determining the probability information of the vehicle getting stuck based on the vehicle operation data and the road terrain image, the method further includes: If the current probability of getting stuck in the vehicle is higher than a preset probability threshold in the vehicle getting stuck probability information, the vehicle getting stuck risk warning information and the terrain segmentation map are displayed on the vehicle interface.

[0011] Optionally, before adjusting the tire pressure according to the tire pressure adjustment amount, the method further includes: In response to the current confidence level in the vehicle getting stuck probability information being lower than a preset confidence threshold, the system determines whether the vehicle is in a preset vehicle getting stuck condition based on the vehicle operation data. In response to the vehicle being in a preset stuck condition, the step of adjusting the tire pressure according to the tire pressure adjustment amount is performed; The preset vehicle-trapping condition is defined as follows: the vehicle speed is lower than a preset speed threshold, the vehicle's driving gradient is greater than a preset gradient threshold, and the road surface adhesion coefficient is lower than a preset adhesion coefficient threshold.

[0012] Optionally, the training steps for the terrain recognition model, the vehicle getting stuck risk prediction model, and the tire pressure regulation model include: A terrain recognition model was trained based on a terrain-annotated dataset, and a vehicle-trapping risk prediction model was trained based on a vehicle-trapping risk-annotated dataset. Freeze the terrain recognition model and the vehicle getting stuck risk prediction model, and train the tire pressure regulation model based on the terrain feature vector output by the terrain recognition model and the vehicle getting stuck probability output by the vehicle getting stuck risk prediction model. Unfreeze a portion of the terrain recognition model and the vehicle getting stuck risk prediction model, and adjust the parameters in the unfrozen structure according to the strategy network in the tire pressure regulation model.

[0013] Optionally, the parameters in the defrosting structure are adjusted according to the strategy network in the tire pressure regulation model, including: Based on the error signal fed back by the strategy network, the corresponding terrain mapping vector and vehicle getting stuck mapping probability are determined. The error signal is fed back by the strategy network when the cumulative value is lower than a preset value. The cumulative value is the sum of the strategy values ​​of each tire pressure strategy under the terrain feature vector and the vehicle getting stuck probability information. Based on the terrain mapping vector, the terrain feature vector, the vehicle getting stuck mapping probability, and the vehicle getting stuck probability, a loss value is determined, and the parameters in the unfreezing structure are adjusted according to the loss value.

[0014] As can be seen from the above, the vehicle tire pressure adjustment method provided in this application acquires vehicle operating data and road terrain images in front of the vehicle, and then predicts the risk of the vehicle getting stuck based on the vehicle operating data and road terrain images to obtain the vehicle's probability of getting stuck. This probability information is then combined with the vehicle operating data to determine the tire pressure adjustment amount for each tire in the vehicle. Tire pressure is then adjusted according to this adjustment amount, achieving timely tire pressure adjustment before the vehicle gets stuck. This method can combine vehicle operating data and road terrain images to predict the risk of the vehicle getting stuck in advance. The accuracy of the risk prediction is ensured through environmental perception and operating data. Furthermore, the tire pressure adjustment amount is determined in a timely manner based on the probability information and vehicle operating data, ensuring the accuracy of tire pressure adjustment. This achieves the goal of timely intervention before the vehicle gets stuck, preventing the vehicle from getting stuck in complex road conditions such as sand and mud due to tire pressure mismatch, improving the vehicle's ability to get out of trouble, and enhancing the vehicle's passability in various complex road conditions. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1This is a flowchart of a vehicle tire pressure adjustment method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a vehicle tire pressure regulating device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0018] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0019] Before providing a detailed description of the vehicle tire pressure adjustment method provided in the embodiments of this application, the technical problem solved by this method will be explained first. In related technologies, vehicle tire pressure adjustment systems often use fixed values ​​or set benchmark values ​​based on simple rules. These systems cannot identify terrain types or predict the risk of getting stuck, resulting in vehicles often getting stuck due to mismatched tire pressure in complex road conditions such as sand and mud.

[0020] Therefore, in order to solve the above problems, this application provides a vehicle tire pressure adjustment method. This method predicts the risk of a vehicle getting stuck by using vehicle operation data and road terrain images. Then, by combining the predicted probability information of getting stuck with the vehicle operation data, the tire pressure adjustment amount is determined, so as to realize timely intervention of tire pressure before getting stuck, avoid the vehicle getting stuck due to tire pressure mismatch, improve the vehicle's ability to get out of trouble, and be able to cope with various extreme and complex road conditions.

[0021] Figure 1This is a flowchart of a vehicle tire pressure regulation method provided in an embodiment of this application. This method is applicable to situations where vehicle operating data and road terrain images are collected in real time during vehicle operation to regulate tire pressure. This method can be executed by the vehicle's tire pressure regulation system, which can be implemented through hardware and / or software. Figure 1 As shown, the vehicle tire pressure adjustment method provided in this application includes the following steps: S110: Acquire vehicle operation data and road terrain images in front of the vehicle.

[0022] In this embodiment, to ensure the accuracy of the prediction of the probability of getting stuck, an infrastructure can be constructed to provide data for the tire pressure regulation system, thereby ensuring the accuracy of subsequent predictions and tire pressure regulation. For example, this can be deployed based on the vehicle's existing electronic and electrical architecture, selecting multiple sources of sensors, including cameras, LiDAR, inertial navigation modules, GPS (Global Positioning System) modules, IMU (Inertial Measurement Unit) modules, tire pressure sensors from TPMS (Tire Pressure Monitoring System), and CAN (Controller Area Network) bus interfaces, as the infrastructure providing data for the tire pressure regulation system.

[0023] Specifically, cameras can capture images of the road terrain in front of the vehicle, and LiDAR, inertial navigation modules, tire pressure sensors, and other devices can collect vehicle operating data. This operating data can include historical and current operating parameters, such as vehicle speed, wheel speed difference, acceleration, steering wheel angle, engine load, engine torque, drive torque, braking status, drive mode, steering status, current tire pressure of each tire, vehicle load, tilt angle, and vehicle attitude.

[0024] For example, vehicle attitude can be collected through GPS and IMU modules, vehicle speed, drive mode, steering status, engine torque and vehicle load can be collected through CAN bus, and the current tire pressure value of each tire can be collected through tire pressure sensors.

[0025] In this embodiment, to ensure time alignment of multi-source data, an onboard central controller can be used as a unified timestamp publishing node, with all data synchronized based on this time. For example, road terrain images and vehicle operation data can be divided into sampling windows in 5-second units to construct data segments with temporal continuity for model training and prediction; terrain labels and vehicle stuck status are manually recorded during the experiment and bound to specific time periods to form training labels for subsequent supervised learning of the model.

[0026] For model training, considering the potential noise in some data, preprocessing can be performed. For example, road terrain images can be resized and normalized, blurry and abnormally exposed frames can be removed, and signal glitches and sensor jitter can be removed from data collected by the CAN bus and tire pressure sensors. Sliding window filtering can enhance stability, and data from the GPS and IMU modules can be fused to improve trajectory smoothness and vehicle orientation stability. Furthermore, all data can be encapsulated into a structured sample format and stored uniformly on a data platform to support distributed training tasks.

[0027] S120. Based on vehicle operation data and road terrain images, determine the probability information of the vehicle getting stuck.

[0028] After obtaining the vehicle's current operating data and road terrain image, the risk of the vehicle getting stuck can be predicted by combining the vehicle operating data and road terrain image, thus obtaining the probability information of getting stuck.

[0029] The vehicle getting stuck probability information may include the current probability of getting stuck, which represents the risk of the vehicle getting stuck while driving on the road ahead. The higher the current probability of getting stuck, the greater the risk. In addition, the vehicle getting stuck probability information may also include the current confidence level, which is used to assess the stability of this risk prediction.

[0030] In this embodiment, the purpose of predicting the risk of getting stuck using vehicle operation data and road terrain images is as follows: Road terrain images can reflect the road conditions ahead of the vehicle, and the possibility of getting stuck is higher under complex road conditions. Therefore, road terrain images need to be considered in the process of predicting the risk of getting stuck. Furthermore, vehicle operation data can reflect the vehicle's operating status. Conditions such as excessive tire pressure, drive wheel slippage, excessive tilt angle, excessive steering angle, or using an inappropriate drive mode can cause the risk of getting stuck. Therefore, vehicle operation data also needs to be considered in the process of predicting the risk of getting stuck. By combining road terrain images and vehicle operation data, the risk of getting stuck can be comprehensively considered, ensuring the accuracy of the prediction.

[0031] For example, the road condition type ahead of a vehicle can be determined based on a road terrain image, and then the probability of the vehicle getting stuck can be determined based on the road condition type and vehicle operation data. For instance, the road terrain image can be input into a pre-trained classifier to obtain the road condition type, and then the road condition type and vehicle operation data can be input into a pre-trained prediction model to obtain the probability of getting stuck.

[0032] In this embodiment, to achieve automatic vehicle identification and classification under complex terrain conditions, a terrain recognition model (which can be built based on the ConvNeXt architecture) can be deployed for visual perception. This terrain recognition model focuses on identifying the terrain ahead and semantic parsing of ground conditions, and is positioned as a core component of the perception layer of the tire pressure regulation system. Furthermore, a vehicle getting stuck risk prediction model (which can be built based on a temporal convolutional network) can also be deployed. This model focuses on modeling and evaluating potential vehicle getting stuck trends, and is positioned as a decision layer, located between the perception layer and the control layer.

[0033] In one specific implementation, determining the probability of a vehicle getting stuck involves using vehicle operation data and road terrain images, including the following steps: Step 11: Determine the current terrain feature vector based on the road terrain image and the pre-trained terrain recognition model; Step 12: Based on the current terrain feature vector, vehicle operation data, and the pre-trained vehicle getting stuck risk prediction model, determine the probability information of the vehicle getting stuck.

[0034] The input to the terrain recognition model is a road terrain image (a single-frame color image). Before inputting the road terrain image into the terrain recognition model, the road terrain image can be preprocessed, including size unification, normalization, and format conversion.

[0035] For example, the size of the road terrain image can be converted to 640×640, with 3 channels and a data type of float32. The normalization method is to normalize the variance based on the channel mean.

[0036] Specifically, the terrain recognition model can employ a multi-stage convolutional structure, consisting of multiple convolutional stages, each including multiple convolutional blocks for extracting texture and boundary features at different scales. In step 11, the road terrain image can be input into the pre-trained terrain recognition model, where each convolutional stage sequentially performs convolution processing to obtain the current terrain feature vector. This current terrain feature vector can represent the high-level semantic information of the overall terrain distribution in the road terrain image.

[0037] For example, the terrain recognition model can employ four convolutional stages. The first convolutional stage is responsible for receiving road terrain images and extracting basic texture information from them. The second convolutional stage is used to extract semantic enhancement features from the basic texture information. The third convolutional stage is used to continue extracting semantic enhancement features from the output of the second convolutional stage. After the third convolutional stage, a channel compression and average pooling module is connected. After channel compression and average pooling, the output of the third convolutional stage can obtain the current terrain feature vector (a one-dimensional vector with a length of 256).

[0038] Furthermore, the fourth convolutional stage is used to construct global perception information from the output of the third convolutional stage. Following the fourth convolutional stage is an upsampling module, which restores the output of the fourth convolutional stage to the original image size to generate a semantic segmentation map aligned with the input road terrain image, serving as the terrain segmentation map. The terrain segmentation map is a two-dimensional tensor with a size of 640×480. Each pixel has an integer value representing the terrain category; for example, 0 represents hard surfaces, 1 represents sand, 2 represents mud, 3 represents grassland, and 4 represents rocky terrain.

[0039] To enhance the identification of terrain boundary areas, a boundary guidance module can be designed into the terrain recognition model. This module, located before the upsampling module, is used to improve the identification of terrain boundary areas. The boundary guidance module consists of a 3×3 convolution and a Sigmoid activation layer, acting before the softmax function output. The current terrain feature vector and terrain segmentation map output by the terrain recognition model correspond to the feature fusion task and terrain visualization, respectively, and they do not share encoding.

[0040] In this embodiment, the terrain recognition model can be trained using supervised learning. The training data is a terrain annotation dataset, which consists of annotated image mask pairs. The mask images are the same size as the original images, and each pixel is labeled with its corresponding terrain category number. The labels can be one-hot encoded for loss calculation. The loss function includes standard cross-entropy loss and boundary-weighted loss, the latter used to improve the recognition accuracy of terrain transition areas. The terrain annotation dataset contains at least 3000 annotated images, covering various terrain combinations, including complex scenes such as road-sand transitions and sand-rock mixtures.

[0041] Furthermore, during the training process of the terrain recognition model, the following enhancement methods can be introduced: image brightness jitter (e.g., 20%), affine transformations (such as rotation) Data augmentation operations are performed during the image reading stage to improve the model's generalization ability under actual operating conditions. These operations include 10-degree blur perturbation (such as using a 3×3 Gaussian blur kernel) and simulated natural occlusion (such as simulating grass, fallen leaves, etc.).

[0042] After obtaining the current terrain feature vector, in step 12, the probability information of the vehicle getting stuck can be determined based on the current terrain feature vector, vehicle operation data, and the pre-trained vehicle getting stuck risk prediction model.

[0043] In this embodiment of the application, considering that the prior art lacks time-series analysis of the vehicle's operating status, resulting in a lag in tire pressure regulation, a vehicle getting stuck risk prediction model can be combined to perform time-series analysis of the operating status in the process of predicting the risk of the vehicle getting stuck, so as to identify the tendency to get stuck in advance and effectively improve the vehicle's passability and stability on unstructured roads.

[0044] In one example, the probability of a vehicle getting stuck is determined based on the current terrain feature vector, vehicle operation data, and a pre-trained vehicle-get-stuck risk prediction model, including the following steps: Step 121: Based on the current terrain feature vector and the current operating parameters in the vehicle operating data, construct the current frame data, and based on the historical terrain feature vector and the historical operating parameters in the vehicle operating data, construct the historical frame data; Step 122: Input the current frame data and historical frame data into the pre-trained vehicle trapping risk prediction model to obtain the vehicle's vehicle trapping probability information.

[0045] Among them, the vehicle getting stuck risk prediction model can adopt the TCN (Temporal Convolutional Network) architecture to make the model have the characteristics of strong temporal dependence, regular computational structure and stable training.

[0046] Specifically, the input to the vehicle getting stuck risk prediction model can be a fixed-duration sequence window, where each frame of data contains multi-dimensional structured features, consisting of operating parameters and terrain feature vectors. For example, if the sampling frequency of the tire pressure regulation system is 5Hz and the time window length is set to 5 seconds, then the input to the vehicle getting stuck risk prediction model can consist of 25 frames of data, which can be generated in real time by the synchronous data acquisition module and continuously fed into the model in a sliding window manner.

[0047] In step 121, the current terrain feature vector and the current operating parameters in the vehicle operating data can be used as the current frame data, and the historical terrain feature vector and the historical operating parameters in the vehicle operating data can be used as the historical frame data. The number of historical frame data can be multiple.

[0048] For example, each frame of data has a total of 49 dimensions and includes the following fields: vehicle speed, wheel speed difference, acceleration, steering wheel angle, engine load, drive torque, braking status, drive mode and steering status, and terrain feature vector. Among them, the terrain feature vector (256 dimensions) can be linearly compressed to 40 dimensions and then concatenated into each frame of data. If the current frame and the 24 historical frames before the current time are input into the vehicle getting stuck risk prediction model, the total dimension of the input to the vehicle getting stuck risk prediction model is 49×25.

[0049] In step 122, the current frame data and historical frame data can be input into a pre-trained vehicle trap risk prediction model. The vehicle trap risk prediction model predicts the vehicle's risk of getting stuck, obtaining vehicle trap probability information. The vehicle trap probability information may include the current vehicle trap probability and the corresponding current confidence level.

[0050] In this embodiment, the vehicle entrapment risk prediction model may include an input transformation layer, a backbone temporal modeling layer, and an output prediction layer. The input transformation layer can convert each frame of data into a uniform dimension (e.g., 64-dimensional) through linear mapping, forming an input tensor (25 frames × 64 dimensions) for subsequent convolutional operations.

[0051] In addition, the backbone temporal modeling layer can adopt a three-layer one-dimensional convolutional structure, which is connected in sequence. Each layer contains dilated convolution, layer normalization and residual connection modules. The kernel size of each layer is 3, and the dilation factors are 1, 2 and 4 respectively. The backbone temporal modeling layer can capture short-term and medium-term behavioral change features, such as continuous deceleration, acceleration and frequent turning.

[0052] Furthermore, the output prediction layer can map the output sequence of the last layer to a single value through a fully connected network. The current probability of getting stuck ranges from [0,1], with a value closer to 1 indicating a higher risk of getting stuck. This probability can serve as an important input indicator for the control layer's policy selection. Additionally, the model also outputs a current confidence level, ranging from [0,1], based on a measure of the consistency between the volatility of the output sequence and the convolutional layer response, used for controlling the decision strength of the policy segment.

[0053] In this embodiment, the vehicle getting stuck risk prediction model can simultaneously output the current vehicle getting stuck probability and the corresponding current confidence level, and transmit both values ​​to the strategy control module, such as the tire pressure regulation model, as one of the inputs to guide the strategy control module to dynamically adjust the tire pressure.

[0054] The training of the vehicle getting stuck risk prediction model can utilize a vehicle getting stuck risk labeled dataset, which can be derived from the off-road test logs and scene reproduction systems of automakers. In the dataset, each sample consists of a 5-second window sequence and a vehicle getting stuck label. Positive samples (getting stuck): within 3 seconds after the sample ends, the vehicle exhibits behaviors such as tire spin, wheel spin, and a continuous zero speed; negative samples (normal): within 10 seconds after the sample ends, the vehicle maintains a stable driving state with no obvious abnormal power output. Samples are automatically generated using a combination of historical vehicle CAN data, terrain recognition results, and manual on-site records, with a positive-to-negative sample ratio of approximately 1:3. Training is performed using a class-balanced sampling method; the loss function can be a standard binary classification cross-entropy. During each iteration of training, 128 sequence samples are sampled from the dataset to form a batch, with 30 training rounds. The Adam optimizer is used, with an initial learning rate set to 0.0001.

[0055] Steps 121-122 above construct current frame data using the current terrain feature vector and current operating parameters, and construct historical frame data using historical terrain feature vectors and historical operating parameters. The current frame data and historical frame data are then input into the vehicle entrapment risk prediction model to obtain vehicle entrapment probability information. This enables temporal analysis of terrain and vehicle operating status, capturing the temporal dependency between terrain and vehicle operating status, and identifying the correlation between them. It also helps to identify entrapment trends in advance, addressing the problem of lagging control in existing technologies and ensuring the accuracy, reliability, and timeliness of vehicle entrapment risk prediction.

[0056] Furthermore, steps 11-12 above determine the current terrain feature vector through road terrain images and terrain recognition models, and determine the probability information of getting stuck through the current terrain feature vector, vehicle operation data, and vehicle getting stuck risk prediction model. This enables accurate perception of the current road conditions and accurate analysis of vehicle getting stuck risk, ensuring accurate prediction of vehicle getting stuck risk.

[0057] S130. Based on the probability of getting stuck and the vehicle operation data, determine the tire pressure adjustment amount for each tire in the vehicle, and adjust the tire pressure according to the tire pressure adjustment amount.

[0058] After obtaining the probability information of getting stuck, in order to achieve dynamic tire pressure adjustment of the vehicle under different terrain and getting stuck risk scenarios, the probability information of getting stuck and the vehicle operation data can be integrated. In addition, the current terrain feature vector can be integrated. The tire pressure adjustment amount of each tire can be determined by the integrated information, which is used to control the tire pressure control unit to perform pressurization or deflation operations in real time.

[0059] In this embodiment of the application, a tire pressure regulation model can be pre-trained. This model can receive the current terrain feature vector output by the terrain recognition model, the vehicle getting stuck probability information output by the vehicle getting stuck risk prediction model, and the current operating parameters in the vehicle operating data. The received information is fused to calculate the tire pressure regulation amount for each tire.

[0060] In one specific implementation, determining the tire pressure adjustment amount for each tire in the vehicle based on the vehicle getting stuck probability information and vehicle operation data includes the following steps: Step 21: Construct the current comprehensive status based on the vehicle getting stuck probability information, the current operating parameters in the vehicle operation data, and the current terrain feature vector; Step 22: Input the current overall state into the strategy network of the pre-trained tire pressure regulation model so that the strategy network can determine the probability distribution of each tire pressure strategy under the current overall state, and determine the target tire pressure strategy based on the probability distribution. The target tire pressure strategy includes the tire pressure adjustment amount of each tire.

[0061] The input to the tire pressure regulation model is the current comprehensive state, which describes the vehicle's comprehensive state information at the current moment, mainly including the probability of getting stuck, current operating parameters, and current terrain feature vector.

[0062] Specifically, the current terrain feature vector (256 dimensions) comes from the terrain recognition model and is used to represent the current ground type, texture, and terrain distribution; the vehicle getting stuck probability information (2 dimensions) comes from the vehicle getting stuck risk prediction model and is used to represent the current likelihood of getting stuck and the reliability of the model's prediction; the current operating parameters can focus on vehicle speed, wheel speed difference, engine torque, drive mode, current tire pressure of each tire, vehicle load, tilt angle, and vehicle attitude. These features are concatenated to form the current comprehensive state, with 270 dimensions. In step 21, the current comprehensive state can be constructed based on the vehicle getting stuck probability information, the current operating parameters in the vehicle operation data, and the current terrain feature vector. Before inputting the current comprehensive state into the tire pressure regulation model, the current comprehensive state can be normalized, for example, by standardizing it using the mean and standard deviation calculated during the training phase.

[0063] In step 22, the current overall state can be input into the tire pressure regulation model, which consists of a policy network and a value network. After the current overall state is input into the tire pressure regulation model, the policy network can determine the probability distribution of each tire pressure strategy under the current overall state through the network parameters (weights and biases) obtained during training. This probability distribution includes the selection probability of each tire pressure strategy under the current overall state, reflecting the quality of each tire pressure strategy. Then, the policy network can select the target tire pressure strategy from among the various tire pressure strategies according to the probability distribution. The tire pressure strategy can include the tire pressure adjustment amount for each tire.

[0064] For example, a policy network can directly select the tire pressure strategy with the highest probability as the target tire pressure strategy, or it can sample and select a tire pressure strategy as the target tire pressure strategy according to the probability distribution.

[0065] In this embodiment of the application, the target tire pressure strategy output by the strategy network may include the tire pressure adjustment amount of each tire at the current moment, a total of 4 floating-point values, in kPa, with a range of [-ΔP, +ΔP], where ΔP is the maximum single pressure adjustment amplitude (configurable, such as 15kPa).

[0066] Specifically, the output of the policy network has a dimension of 4 and an output format of a float32 array, which corresponds to the tire pressure adjustment amount of the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively. Positive values ​​represent restarting, negative values ​​represent deflation, and zero represents maintaining the current tire pressure.

[0067] For example, the policy network can adopt a three-layer fully connected network, with ReLU activation after each layer, and the output is a 4-dimensional action value (tire pressure adjustment amount, i.e. target tire pressure policy). The action output follows a Gaussian distribution, and the mean value is finally taken as the control command.

[0068] Steps 21-22 above can simultaneously integrate vehicle getting stuck probability information, current terrain feature vector, and current operating parameters. Through multi-level feature fusion, the three are kept consistent in the feature representation space, avoiding information fragmentation and forming continuous perception. This helps to improve the consistency and stability of the tire pressure regulation model response and achieve the goal of precise tire pressure regulation.

[0069] In this embodiment of the application, the network parameters in the policy network can be obtained by training the value network. The value network can evaluate the quality of each tire pressure policy through a pre-designed reward function, and then adjust the network parameters in the policy network through a training algorithm (such as policy gradient), so that the probability of the policy network outputting a high-value policy in the same state next time increases and the probability of outputting a low-value policy decreases.

[0070] In some implementations, the network parameters in the policy network are obtained by training the value network in the tire pressure regulation model. The value network is used to determine the policy value of executing the preset tire pressure policy under the preset comprehensive state. The strategic value includes at least one of the following: vehicle pass reward, frequent voltage adjustment penalty, smooth voltage adjustment reward, energy consumption penalty, and rapid response reward.

[0071] Specifically, the value network can calculate the strategy value of executing each preset tire pressure strategy under a preset comprehensive state. The strategy value can include at least one of the following: vehicle passability reward, frequent pressure adjustment penalty, smooth pressure adjustment reward, energy consumption penalty, and rapid response reward.

[0072] Among them, the vehicle pass reward is a positive reward when the vehicle does not get stuck; the frequent pressure adjustment penalty is a penalty when the pressure adjustment frequency is greater than a preset frequency threshold; the smooth pressure adjustment reward is a positive reward when the pressure adjustment amplitude is less than a preset amplitude threshold; the energy consumption penalty is a penalty when the air pump energy consumption is greater than a preset energy consumption threshold during the pressure adjustment process; and the rapid response reward is a positive reward for rapid pressure adjustment when the risk of getting stuck is high.

[0073] In this embodiment, the value network can calculate the strategy value of each preset tire pressure strategy under a preset comprehensive state based on the reward function. After calculating the strategy value, the network parameters in the strategy network are adjusted by comparing the expected value with the strategy value.

[0074] For example, if the value of the preset tire pressure strategy is higher than the expected value, it means that the preset tire pressure strategy produces a positive reward. The network parameters of the strategy network can be adjusted to increase the probability of the strategy network outputting the preset tire pressure strategy under the preset overall state. Conversely, if the value of the preset tire pressure strategy is lower than the expected value, it means that the preset tire pressure strategy produces a negative reward. The network parameters of the strategy network can be adjusted to decrease the probability of the strategy network outputting the preset tire pressure strategy under the preset overall state. In other words, the goal of training is to adjust the network parameters of the strategy network so that the probability distribution output by the strategy network maximizes the cumulative reward in the future.

[0075] After several of the above adjustments, the network parameters of the policy network can learn a kind of "mapping": when a comprehensive state is received, it can output the tire pressure policy that has historically yielded high policy value under that comprehensive state with a high probability.

[0076] In this application's embodiment set, the tire pressure regulation model can be trained using a simulated environment built by the automaker, featuring typical terrain (based on measured data, such as soft sand, slippery mud, rocky slopes, etc.) and a dynamic model. The simulation models model the vehicle's performance under different tire pressures, including adhesion, slippage rate, fuel consumption changes, and passability. By evaluating the vehicle's performance under different strategies, the tire pressure regulation model is continuously optimized. Each training iteration includes a sampling phase, an advantage estimation phase, and a strategy update phase. The training batch size is 64, and the number of iterations is set to 5000. Each iteration samples a complete simulation trajectory (100-200 steps in length). The learned experience (including state, tire pressure strategy, and strategy value) is processed using GAE (Generalized Advantage Estimator).

[0077] Through the above implementation methods, the strategy network can be trained by combining at least one of the following: vehicle passage reward, frequent pressure adjustment penalty, smooth pressure adjustment reward, energy consumption penalty, and rapid response reward. This allows the strategy network to output tire pressure strategies that avoid getting stuck, thereby improving vehicle passage rate, through vehicle passage reward; to avoid outputting frequently adjusted tire pressure strategies, reducing the adjustment frequency and preventing strategy oscillation, through frequent pressure adjustment penalty; to output smooth tire pressure strategies, reducing the adjustment amplitude, and achieving flexible and continuous tire pressure optimization adjustment, thus solving the problem of excessive discrete action adjustment affecting the user's riding experience in existing control strategies, through energy consumption penalty; to output tire pressure strategies with low air pump energy consumption, thereby reducing overall energy consumption; and to adjust tire pressure quickly when the risk of getting stuck is high, through rapid response reward, thus achieving rapid response to the risk of getting stuck.

[0078] It should be noted that in the embodiments of this application, the terrain recognition model, the vehicle getting stuck risk prediction model, and the tire pressure regulation model are not simply used in series, but rather form a collaborative operation mechanism through multi-level feature sharing and dynamic interaction, constructing a closed loop from scene perception to risk judgment, and from risk judgment to action control, which has high interpretability and structural adaptability.

[0079] Specifically, the first layer of fusion occurs at the feature level. The current terrain feature vector output by the terrain recognition model is not only fed into the vehicle entrapment risk prediction model for temporal modeling of vehicle entrapment risk, but also serves as input to the tire pressure regulation model. In the vehicle entrapment risk prediction model, the terrain feature vector is compressed to 40 dimensions and then concatenated into the vehicle's operating parameters in each frame, participating in the vehicle entrapment risk assessment process. In the tire pressure regulation model, the complete terrain feature vector is directly concatenated into the model input as a high-dimensional expression of the current overall state. This multi-path sharing strategy ensures consistency among the three models in the feature representation space, avoids information fragmentation, forms a continuous perception link, and helps improve the consistency and stability of the model response.

[0080] Furthermore, the second layer of integration is reflected in the target tire pressure strategy output stage. The probability information of getting stuck output by the vehicle getting stuck risk prediction model will directly participate in the control decision of the tire pressure regulation model, forming an adaptive strategy driving signal. For example, if the current probability of getting stuck increases and the current confidence level is greater than the set threshold, the tire pressure regulation model will reduce the tire pressure of each tire and widen the action adjustment range; if the current probability of getting stuck is lower than the set threshold and the current confidence level is lower than the set threshold, the tire pressure regulation model will maintain high tire pressure and limit frequent pressure adjustments.

[0081] It should also be noted that the probability and confidence level of getting stuck output by the vehicle getting stuck risk prediction model are not used as ordinary concatenation inputs to the tire pressure regulation model. Instead, they are used to control the activation paths of key layers in the policy network through a dynamic gating mechanism. For example, the confidence level is used to adjust the channel weight distribution of the hidden layer, while the probability of getting stuck participates in the variance adjustment of the action distribution, achieving a dynamic balance between policy flexibility and stability.

[0082] In this embodiment, to further improve the prediction accuracy of the three models, a weakly coupled joint training mechanism can be introduced during the policy network training phase. For example, the terrain recognition model and the vehicle-trapping risk prediction model are pre-trained models, but the last fine-tunable layer is retained. During the reinforcement learning training of the tire pressure regulation model, some gradients are introduced into the output layers of the terrain recognition model and the vehicle-trapping risk prediction model through backpropagation to optimize their response accuracy under the tire pressure regulation model. A policy effect feedback term can be added to the loss function, enabling collaborative optimization among the terrain recognition model, the vehicle-trapping risk prediction model, and the tire pressure regulation model.

[0083] In some implementations, the training steps for the terrain recognition model, the vehicle getting stuck risk prediction model, and the tire pressure regulation model include: Step 31: Train a terrain recognition model based on a terrain annotation dataset, and train a vehicle getting stuck risk prediction model based on a vehicle getting stuck risk annotation dataset; Step 32: Freeze the terrain recognition model and the vehicle getting stuck risk prediction model, and train the tire pressure regulation model based on the terrain feature vector output by the terrain recognition model and the vehicle getting stuck probability output by the vehicle getting stuck risk prediction model. Step 33: Thaw part of the terrain recognition model and vehicle entrapment risk prediction model, and adjust the parameters in the thawing structure according to the strategy network in the tire pressure regulation model.

[0084] In step 31, the terrain recognition model and the vehicle getting stuck risk prediction model can be trained using the terrain annotation dataset and the vehicle getting stuck risk annotation dataset, respectively, to obtain stable terrain recognition models and vehicle getting stuck risk prediction models. This stage ensures that the perception module has the ability to accurately represent the environment and vehicle state, and all output structures of the terrain recognition model and the vehicle getting stuck risk prediction model are fixed at this stage.

[0085] Furthermore, in step 32, the terrain recognition model and the vehicle-trapping risk prediction model can be frozen, ensuring that they are in a frozen state and do not participate in gradient updates. The tire pressure regulation model receives the terrain feature vector output by the terrain recognition model and the vehicle-trapping risk prediction model output the vehicle-trapping probability. Simultaneously, it interacts with the vehicle simulation environment to obtain vehicle operation data and reward feedback, optimizing the network parameters in the policy network and thus training the policy network.

[0086] After the policy network converges, in step 33, a portion of the structure of the terrain recognition model and the vehicle getting stuck risk prediction model can be unfrozen. For example, the last layer or the embedding layer of the terrain recognition model and the vehicle getting stuck risk prediction model can be unfrozen to allow the terrain recognition model and the vehicle getting stuck risk prediction model to receive feedback generated by the tire pressure regulation model and fine-tune the parameters in the unfrozen structure through the joint loss function.

[0087] For example, in the initialization phase, the terrain recognition model and the vehicle-trapping risk prediction model can be pre-trained and the training results saved. A state cache module can be built so that the tire pressure regulation model can call the output of the terrain recognition model and the vehicle-trapping risk prediction model. In the policy training phase, the tire pressure regulation model starts training and calls the output of the terrain recognition model and the vehicle-trapping risk prediction model to generate each round of trajectory sampling. After every 1000 rounds of policy training, the state-reward record is saved once for later feedback analysis of the perception model (terrain recognition model and vehicle-trapping risk prediction model). In the joint fine-tuning phase, states with low cumulative value are filtered, a policy feedback sample set is built, the output layer of the perception model is fine-tuned, and the policy performance changes are observed. During the fine-tuning process, the output interface format is kept consistent to avoid interrupting the interface connection.

[0088] Regarding step 33 above, in one example, the parameters in the defrosting structure are adjusted according to the strategy network in the tire pressure regulation model, including the following steps: Step 331: Based on the error signal fed back by the strategy network, determine the corresponding terrain mapping vector and the probability of getting stuck. The error signal is fed back by the strategy network when the cumulative value is lower than the preset value. The cumulative value is the sum of the strategy values ​​of each tire pressure strategy under the terrain feature vector and the probability of getting stuck information. Step 332: Based on the terrain mapping vector, terrain feature vector, vehicle stuck mapping probability, and vehicle stuck probability, determine the loss value, and adjust the parameters in the unfreezing structure according to the loss value.

[0089] In step 331, the policy network can record the cumulative value generated by the corresponding tire pressure policy under each state input (including terrain feature vector and vehicle getting stuck probability information). The cumulative value is the sum of the policy values ​​of all tire pressure policies under the state input (including terrain feature vector and vehicle getting stuck probability information).

[0090] For state inputs (including terrain feature vectors and vehicle-stuck probability information) with a cumulative value lower than a preset value, the policy network can feed back an error signal to indicate that the state input has failed to accurately reflect environmental changes. Specifically, after receiving the error signal from the policy network, the error signal can be mapped back to the form of terrain feature vectors and vehicle-stuck probability to obtain a terrain mapping vector and a vehicle-stuck probability mapping.

[0091] Furthermore, in step 332, the difference between the terrain mapping vector and the terrain feature vector, and the difference between the vehicle getting stuck mapping probability and the vehicle getting stuck probability can be calculated. A loss value is obtained through a weighted loss function, and this loss value is fed back to the unfrozen structure in the terrain recognition model and the vehicle getting stuck risk prediction model, such as the output layer of the terrain recognition model and the penultimate layer of the vehicle getting stuck risk prediction model, to optimize the parameters in the unfrozen structure. It should be noted that the parameters of the tire pressure regulation model remain unchanged during the fine-tuning process.

[0092] The weighted loss function can be: Loss = Original Perception Loss + Policy Feedback Loss × Policy Dominant Weight. The policy feedback loss can be determined based on the difference between the terrain mapping vector and the terrain feature vector, and the difference between the vehicle getting stuck mapping probability and the vehicle getting stuck probability; the policy dominant weight can be increased from 0.1 to 0.5 to control the fine-tuning range of the terrain recognition model and the vehicle getting stuck risk prediction model.

[0093] In steps 331-332 above, the strategy network can feed back error signals based on the cumulative value, and then adjust the unfrozen structure in the terrain recognition model and the vehicle entrapment risk prediction model based on the error signals. This can further improve the consistency between perception and strategy control, thereby improving the accuracy of tire pressure regulation.

[0094] Steps 31-33 above balance the controllability of independent modeling with the synergy of overall system performance. Unlike the traditional end-to-end training mode, it has higher flexibility and system stability, and is more suitable for multi-model combinations in complex perception-decision-control chains, further improving the adaptability of perception, decision-making and control, and improving the accuracy of tire pressure regulation.

[0095] It should be noted that, in order to achieve the operability of collaborative optimization among multiple models, the interface format between the multiple models needs to be standardized and fixed. For example, the terrain feature vector output by the terrain recognition model is determined at the beginning of training and is not allowed to change dynamically; the input structure (49-dimensional × 25 frames) of the vehicle trap risk prediction model consists of vehicle operation data and compressed terrain feature vector, and remains consistent throughout all training stages; the input structure of the tire pressure regulation model is a stitched 270-dimensional state vector, which consists of the output of the perception model and vehicle operation data. After the perception model is fine-tuned, the inference data cache is regenerated to maintain interface alignment.

[0096] In this embodiment of the application, after obtaining the tire pressure adjustment amount of each tire, an adjustment command can be generated based on the tire pressure adjustment amount, and the adjustment command can be sent to the tire pressure adjustment hardware interface. The adjustment command can be executed by the electronically controlled air pressure module to complete the tire pressure adjustment of each tire.

[0097] To help users understand the risk of getting stuck more intuitively, after obtaining the probability information of the vehicle getting stuck, a risk warning message and terrain segmentation map can be displayed so that users can understand the current road conditions and the risk of getting stuck in a timely manner.

[0098] In some implementations, in addition to determining the current terrain feature vector, the method further includes: determining a terrain segmentation map based on the road terrain image and a pre-trained terrain recognition model, wherein the terrain segmentation map is used to describe the terrain category corresponding to each pixel in the road terrain image; After determining the probability information of a vehicle getting stuck based on vehicle operation data and road terrain images, the system also includes: in response to the current probability of getting stuck being higher than a preset probability threshold, displaying a vehicle getting stuck risk warning message and terrain segmentation map through the vehicle interface.

[0099] Specifically, after inputting the road terrain image into the terrain recognition model, the model can simultaneously output the current terrain feature vector and the terrain segmentation map. For example, after upsampling, the output of the fourth convolutional stage in the terrain recognition model can be restored to the size of the road terrain image, generating a semantic segmentation map aligned with the input. This semantic segmentation map is the terrain segmentation map, which includes the terrain category corresponding to each pixel in the road terrain image. The values ​​of each pixel in the terrain segmentation map are integers, representing the terrain category. For example, the terrain categories are defined as follows: 0: hard road surface; 1: sandy land; 2: muddy land; 3: grassland; 4: rocky terrain.

[0100] Furthermore, after obtaining the vehicle getting stuck probability information, it can be determined whether the current vehicle getting stuck probability in the vehicle getting stuck probability information is higher than the preset probability threshold. If so, it means that the vehicle is at high risk of getting stuck. At this time, it is necessary to remind the user in time, generate a vehicle getting stuck risk warning message, and display the vehicle getting stuck risk warning message on the vehicle interface.

[0101] While displaying the risk warning information of getting stuck through the vehicle's infotainment system, the system can also display the terrain segmentation map output by the terrain recognition model, so that users can more intuitively understand the root cause of the risk of getting stuck.

[0102] The above implementation method can display vehicle getting stuck risk warning information and terrain segmentation map through the vehicle interface when the current probability of getting stuck is higher than a preset probability threshold. This allows users to be aware of the risk of getting stuck and the current road conditions in a timely manner, thus providing timely reminders to users. If subsequent tire pressure adjustment fails to eliminate the risk of getting stuck, it facilitates timely manual intervention by the user, thereby improving the riding experience and driving safety.

[0103] In this embodiment of the application, in order to further improve safety in extreme scenarios, a risk redundancy channel can also be introduced. This channel can trigger static logic judgment when the current confidence level is lower than the preset confidence level threshold, so as to determine whether the vehicle is in a preset stuck condition, and ensure that the subsequent pressure regulation strategy will not be misoperated due to prediction errors.

[0104] In one specific implementation, before adjusting the tire pressure according to the tire pressure adjustment amount, the following steps are also included: Step 41: In response to the current confidence level in the vehicle getting stuck probability information being lower than the preset confidence threshold, determine whether the vehicle is in the preset vehicle getting stuck condition based on the vehicle operation data; Step 42: In response to the vehicle being in a preset stuck condition, perform the step of adjusting the tire pressure according to the tire pressure adjustment amount.

[0105] In step 41, it can be determined whether the current confidence level in the vehicle getting stuck probability information is lower than the preset confidence threshold. If so, it means that the prediction reliability of the vehicle getting stuck risk is low. At this time, redundant judgment can be made by combining vehicle operation data to determine whether the vehicle is in the preset vehicle getting stuck condition.

[0106] Specifically, the vehicle's speed, gradient, and road surface adhesion coefficient can be used to determine whether the vehicle is in a preset stuck condition. A preset stuck condition is defined as follows: the vehicle speed is below a preset speed threshold, the vehicle gradient is above a preset gradient threshold, and the road surface adhesion coefficient is below a preset adhesion coefficient threshold.

[0107] If the vehicle speed is lower than the preset speed threshold, the vehicle's driving gradient is greater than the preset gradient threshold, and the road surface adhesion coefficient is lower than the preset adhesion coefficient threshold, it indicates that the vehicle is at extremely high risk of getting stuck. It can be determined that the vehicle is in the preset stuck condition. Further, in step 42, the step of adjusting the tire pressure according to the tire pressure adjustment amount can be executed. Otherwise, the step of adjusting the tire pressure according to the tire pressure adjustment amount can be cancelled.

[0108] Steps 41-42 above, when the current confidence level is lower than the preset confidence threshold, can accurately identify whether the vehicle is about to get stuck by introducing redundant judgments. This avoids subsequent tire pressure adjustment from being misoperated due to risk prediction errors, thus providing a safety net for tire pressure adjustment, further ensuring the reliability of tire pressure adjustment, and improving the accuracy of tire pressure adjustment in extreme scenarios.

[0109] The vehicle tire pressure adjustment method provided in this application acquires vehicle operating data and road terrain images in front of the vehicle, and then predicts the risk of the vehicle getting stuck based on the vehicle operating data and road terrain images to obtain the vehicle's probability of getting stuck. By combining the probability of getting stuck with the vehicle operating data, the tire pressure adjustment amount for each tire in the vehicle is determined, and tire pressure is adjusted accordingly. This allows for timely tire pressure adjustment before the vehicle gets stuck. This method can combine vehicle operating data and road terrain images to predict the risk of the vehicle getting stuck in advance. The accuracy of the risk prediction is ensured through environmental perception and operating data. Furthermore, the tire pressure adjustment amount is determined in a timely manner based on the probability of getting stuck and the vehicle operating data, ensuring the accuracy of tire pressure adjustment. This achieves the goal of timely intervention before the vehicle gets stuck, preventing the vehicle from getting stuck in complex road conditions such as sand and mud due to mismatched tire pressure, improving the vehicle's ability to get out of trouble, and enhancing the vehicle's passability in various complex road conditions.

[0110] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0111] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0112] Based on the same inventive concept, and corresponding to any of the above embodiments, this application also provides a vehicle tire pressure regulating device. Figure 2 This is a schematic diagram of a vehicle tire pressure regulating device provided in an embodiment of this application, with reference to... Figure 2 The vehicle tire pressure regulation device includes an acquisition module 210, a vehicle getting stuck prediction module 220, and a strategy adjustment module 230, wherein: The acquisition module 210 is used to acquire vehicle operation data and road terrain images in front of the vehicle. The vehicle getting stuck prediction module 220 is used to determine the probability information of the vehicle getting stuck based on the vehicle operation data and the road terrain image; The strategy adjustment module 230 is used to determine the tire pressure adjustment amount of each tire in the vehicle based on the vehicle getting stuck probability information and the vehicle operation data, and to adjust the tire pressure according to the tire pressure adjustment amount.

[0113] Based on the above implementation, optionally, the vehicle-trapping prediction module 220 includes a terrain recognition unit and a vehicle-trapping prediction unit, wherein: The terrain recognition unit is used to determine the current terrain feature vector based on the road terrain image and the pre-trained terrain recognition model. The vehicle getting stuck prediction unit is used to determine the probability information of the vehicle getting stuck based on the current terrain feature vector, the vehicle operation data, and a pre-trained vehicle getting stuck risk prediction model.

[0114] Based on the above implementation method, optionally, the vehicle getting stuck prediction unit is specifically used for: Based on the current terrain feature vector and the current operating parameters in the vehicle operating data, the current frame data is constructed, and based on the historical terrain feature vector and the historical operating parameters in the vehicle operating data, historical frame data is constructed. The current frame data and the historical frame data are input into a pre-trained vehicle getting stuck risk prediction model to obtain the vehicle's probability of getting stuck.

[0115] Based on the above implementation method, optionally, the strategy adjustment module 230 is specifically used for: The current overall status is constructed based on the vehicle getting stuck probability information, the current operating parameters in the vehicle operation data, and the current terrain feature vector; The current overall state is input into the policy network of a pre-trained tire pressure regulation model, so that the policy network determines the probability distribution of each tire pressure policy under the current overall state, and determines the target tire pressure policy based on the probability distribution, wherein the target tire pressure policy includes the tire pressure adjustment amount of each tire.

[0116] Based on the above implementation, optionally, the network parameters in the strategy network are obtained by training the value network in the tire pressure regulation model, and the value network is used to determine the strategy value of executing the preset tire pressure strategy under the preset comprehensive state. The strategy value includes at least one of the following: vehicle pass reward, frequent voltage adjustment penalty, smooth voltage adjustment reward, energy consumption penalty, and rapid response reward.

[0117] Based on the above embodiments, optionally, the terrain recognition unit is further configured to determine a terrain segmentation map based on the road terrain image and a pre-trained terrain recognition model, wherein the terrain segmentation map is used to describe the terrain category corresponding to each pixel in the road terrain image; the device further includes a visualization module, which is configured to display a vehicle-mounted vehicle risk warning information and the terrain segmentation map through the vehicle interface in response to the current vehicle-mounted vehicle probability information being higher than a preset probability threshold.

[0118] Based on the above implementation method, optionally, the strategy adjustment module 230 is further configured to: In response to the current confidence level in the vehicle getting stuck probability information being lower than a preset confidence threshold, the system determines whether the vehicle is in a preset vehicle getting stuck condition based on the vehicle operation data. In response to the vehicle being in a preset stuck condition, the step of adjusting the tire pressure according to the tire pressure adjustment amount is performed; The preset vehicle-trapping condition is defined as follows: the vehicle speed is lower than a preset speed threshold, the vehicle's driving gradient is greater than a preset gradient threshold, and the road surface adhesion coefficient is lower than a preset adhesion coefficient threshold.

[0119] Optionally, based on the above embodiments, the device further includes a model training module, the model training module being used for: A terrain recognition model was trained based on a terrain-annotated dataset, and a vehicle-trapping risk prediction model was trained based on a vehicle-trapping risk-annotated dataset. Freeze the terrain recognition model and the vehicle getting stuck risk prediction model, and train the tire pressure regulation model based on the terrain feature vector output by the terrain recognition model and the vehicle getting stuck probability output by the vehicle getting stuck risk prediction model. Unfreeze a portion of the terrain recognition model and the vehicle getting stuck risk prediction model, and adjust the parameters in the unfrozen structure according to the strategy network in the tire pressure regulation model.

[0120] Optionally, based on the above embodiments, the model training module is further configured to: Based on the error signal fed back by the strategy network, the corresponding terrain mapping vector and vehicle getting stuck mapping probability are determined. The error signal is fed back by the strategy network when the cumulative value is lower than a preset value. The cumulative value is the sum of the strategy values ​​of each tire pressure strategy under the terrain feature vector and the vehicle getting stuck probability information. Based on the terrain mapping vector, the terrain feature vector, the vehicle getting stuck mapping probability, and the vehicle getting stuck probability, a loss value is determined, and the parameters in the unfreezing structure are adjusted according to the loss value.

[0121] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0122] The apparatus described above is used to implement the corresponding vehicle tire pressure adjustment method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0123] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the vehicle tire pressure regulation method described in any of the above embodiments.

[0124] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a vehicle, the vehicle comprising: 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 tire pressure regulation method described in any of the above embodiments.

[0125] Figure 3 Figure 3 is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 3 illustrates a specific hardware structure of the electronic device provided in this embodiment. The device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0126] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0127] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0128] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0129] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0130] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0131] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0132] The electronic devices described above are used to implement the corresponding vehicle tire pressure regulation methods in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0133] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the vehicle tire pressure regulation method as described in any of the above embodiments.

[0134] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0135] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the vehicle tire pressure adjustment method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0136] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0137] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0138] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0139] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A method for adjusting vehicle tire pressure, characterized in that, include: Acquire vehicle operation data and road terrain images in front of the vehicle; Based on the vehicle operation data and the road terrain image, determine the probability information of the vehicle getting stuck; Based on the vehicle getting stuck probability information and the vehicle operation data, the tire pressure adjustment amount of each tire in the vehicle is determined, and the tire pressure is adjusted according to the tire pressure adjustment amount.

2. The vehicle tire pressure adjustment method according to claim 1, characterized in that, Based on the vehicle operation data and the road terrain image, the probability information of the vehicle getting stuck is determined, including: Based on the road terrain image and the pre-trained terrain recognition model, determine the current terrain feature vector; Based on the current terrain feature vector, the vehicle operation data, and the pre-trained vehicle getting stuck risk prediction model, the probability information of the vehicle getting stuck is determined.

3. The vehicle tire pressure adjustment method according to claim 2, characterized in that, Based on the current terrain feature vector, the vehicle operation data, and the pre-trained vehicle-trapping risk prediction model, the probability information of the vehicle getting stuck is determined, including: Based on the current terrain feature vector and the current operating parameters in the vehicle operating data, the current frame data is constructed, and based on the historical terrain feature vector and the historical operating parameters in the vehicle operating data, historical frame data is constructed. The current frame data and the historical frame data are input into a pre-trained vehicle getting stuck risk prediction model to obtain the vehicle's probability of getting stuck.

4. The vehicle tire pressure adjustment method according to claim 2, characterized in that, Based on the vehicle getting stuck probability information and the vehicle operation data, determine the tire pressure adjustment amount for each tire in the vehicle, including: The current overall status is constructed based on the vehicle getting stuck probability information, the current operating parameters in the vehicle operation data, and the current terrain feature vector; The current overall state is input into the policy network of a pre-trained tire pressure regulation model, so that the policy network determines the probability distribution of each tire pressure policy under the current overall state, and determines the target tire pressure policy based on the probability distribution, wherein the target tire pressure policy includes the tire pressure adjustment amount of each tire.

5. The vehicle tire pressure adjustment method according to claim 4, characterized in that, The network parameters in the strategy network are obtained by training the value network in the tire pressure regulation model. The value network is used to determine the strategy value of executing the preset tire pressure strategy under the preset comprehensive state. The strategy value includes at least one of the following: vehicle pass reward, frequent voltage adjustment penalty, smooth voltage adjustment reward, energy consumption penalty, and rapid response reward.

6. The vehicle tire pressure adjustment method according to claim 2, characterized in that, In addition to determining the current terrain feature vector, the following are also included: Based on the road terrain image and the pre-trained terrain recognition model, a terrain segmentation map is determined, wherein the terrain segmentation map is used to describe the terrain category corresponding to each pixel in the road terrain image; After determining the probability information of the vehicle getting stuck based on the vehicle operation data and the road terrain image, the method further includes: If the current probability of getting stuck in the vehicle is higher than a preset probability threshold in the vehicle getting stuck probability information, the vehicle getting stuck risk warning information and the terrain segmentation map are displayed on the vehicle interface.

7. The vehicle tire pressure adjustment method according to claim 1, characterized in that, Before adjusting the tire pressure according to the stated tire pressure adjustment amount, the method further includes: In response to the current confidence level in the vehicle getting stuck probability information being lower than a preset confidence threshold, the system determines whether the vehicle is in a preset vehicle getting stuck condition based on the vehicle operation data. In response to the vehicle being in a preset stuck condition, the step of adjusting the tire pressure according to the tire pressure adjustment amount is performed; The preset vehicle-trapping condition is defined as follows: the vehicle speed is lower than a preset speed threshold, the vehicle's driving gradient is greater than a preset gradient threshold, and the road surface adhesion coefficient is lower than a preset adhesion coefficient threshold.

8. The vehicle tire pressure adjustment method according to claim 4, characterized in that, The training steps for the terrain recognition model, the vehicle getting stuck risk prediction model, and the tire pressure regulation model include: A terrain recognition model was trained based on a terrain-annotated dataset, and a vehicle-trapping risk prediction model was trained based on a vehicle-trapping risk-annotated dataset. Freeze the terrain recognition model and the vehicle getting stuck risk prediction model, and train the tire pressure regulation model based on the terrain feature vector output by the terrain recognition model and the vehicle getting stuck probability output by the vehicle getting stuck risk prediction model. Unfreeze a portion of the terrain recognition model and the vehicle getting stuck risk prediction model, and adjust the parameters in the unfrozen structure according to the strategy network in the tire pressure regulation model.

9. The vehicle tire pressure adjustment method according to claim 8, characterized in that, The parameters in the defrosting structure are adjusted according to the strategy network in the tire pressure regulation model, including: Based on the error signal fed back by the strategy network, the corresponding terrain mapping vector and vehicle getting stuck mapping probability are determined. The error signal is fed back by the strategy network when the cumulative value is lower than a preset value. The cumulative value is the sum of the strategy values ​​of each tire pressure strategy under the terrain feature vector and the vehicle getting stuck probability information. Based on the terrain mapping vector, the terrain feature vector, the vehicle getting stuck mapping probability, and the vehicle getting stuck probability, a loss value is determined, and the parameters in the unfreezing structure are adjusted according to the loss value.

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 tire pressure regulation method as described in any one of claims 1 to 9.

Images