Vehicle control method, electronic device, and vehicle

Abstract

The application relates to a vehicle control method, an electronic device and a vehicle, and relates to the technical field of vehicle control. The method comprises the following steps: acquiring a multi-dimensional feature sequence of a vehicle in a historical period; processing the multi-dimensional feature sequence through a graph construction model, a prediction model and a correction model connected in sequence to obtain a causal graph, a friction trend sequence and correction data of the friction trend sequence output by each model, wherein the causal graph is used to indicate the causal correlation between each dimension feature of the vehicle in the driving process, and the friction trend sequence is used to indicate the change of the tire road adhesion grip force with time in a future period; fusing the causal graph, the friction trend sequence and the correction data to predict a slip probability sequence of the vehicle in the future period; and generating a corresponding vehicle-mounted control instruction based on the slip probability sequence. The application can improve driving safety.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and in particular to a vehicle control method, electronic equipment, and vehicle. Background Technology

[0002] When a vehicle travels on low-adhesion surfaces such as wet and slippery roads, snow and ice, or oily surfaces, the grip between the tires and the road surface decreases significantly, making the wheels prone to slipping and sliding, which greatly reduces the vehicle's driving stability.

[0003] Currently, most vehicle anti-skid stability control systems use feedback control logic, relying on instantaneous state data collected by the vehicle to determine slippage. Intervention can only be initiated after the wheels have shown a tendency to slip, which is a reactive control method. Therefore, the chassis braking, torque control and other controllers cannot complete the preparatory work before slippage occurs, and there is a significant time lag in control intervention, which affects driving safety. Summary of the Invention

[0004] This application provides a vehicle control method, electronic equipment, and vehicle to improve driving safety on low-adhesion road surfaces.

[0005] In a first aspect, this application provides a vehicle control method, the method comprising: Obtain a multidimensional feature sequence of the vehicle over a historical period, wherein the multidimensional feature sequence consists of multiple features related to tire slippage collected at each historical moment during the vehicle's operation. The multidimensional feature sequence is processed by a cascaded graph construction model, a prediction model, and a correction model to obtain the causal graph, friction trend sequence, and correction data of the friction trend sequence output by each model. The causal graph is used to indicate the causal relationship between the features of each dimension of the vehicle during driving, and the friction trend sequence is used to indicate the change of tire road surface adhesion over time in the future. By fusing the causal map, the friction trend sequence, and the correction data, a slip probability sequence of the vehicle in the future time period is predicted. The corresponding vehicle control commands are generated based on the slip probability sequence.

[0006] Optionally, the multidimensional feature sequence is processed by a cascaded graph construction model, a prediction model, and a correction model to obtain the causal graph, friction trend sequence, and correction data of the friction trend sequence output by each model, respectively, including: The multidimensional feature sequence is input into the graph construction model to obtain the causal graph output by the graph construction model. The nodes in the causal graph are used to indicate features related to tire slippage, and the edges in the causal graph represent the causal relationships between the features. The friction trend sequence is generated by adjusting the edge weights and fusing node features of the causal graph using a prediction model combined with the multidimensional feature sequence. The friction trend sequence is corrected by combining the correction model with the current vehicle operating condition data, and the corrected data of the friction trend sequence is output. The corrected data includes the corrected trend sequence and confidence interval of the friction trend sequence.

[0007] Optionally, inputting the multidimensional feature sequence into the graph construction model to obtain the causal graph output by the graph construction model includes: The multidimensional feature sequence is input into the graph construction model, wherein the graph construction model has a built-in multi-head attention module, a causal constraint module, a graph generation module, and an output module; The temporal dependencies between features are discovered through the multi-head attention module. The causal constraint module applies causal constraints based on the temporal dependency relationship and determines the causal correlation strength between features through the soft mask corresponding to the causal constraint. The causal constraint is used to limit only historical features to positively influence future features and prohibit future features from negatively influencing historical features. The soft mask is used to quantize the causal constraint into a continuous weight matrix. The graph generation module filters out real causal associations with a causal association strength greater than a preset strength threshold, and constructs a causal graph based on each feature and the real causal associations. The causal graph is output through the output module.

[0008] Optionally, the friction trend sequence is generated by adjusting the edge weights and fusing node features of the causal graph using a prediction model combined with the multidimensional feature sequence, including: The causal graph and the multidimensional feature sequence are input into the prediction model, wherein the prediction model has a built-in graph attention module, graph reconstruction module, inference module and output module; The graph attention module extracts multidimensional data corresponding to each historical moment in the multidimensional feature sequence in chronological order, and performs node feature fusion based on the multidimensional data. The edge weights of the causal graph are then updated based on the fused node features. After completing the fusion of node features and the update of edge weights for all historical moments in the historical period, a dynamic causal graph is obtained through the graph reconstruction module. The inference module infers the dynamic causal graph according to the time sequence of the future period, generating a friction trend sequence covering the future period; The friction trend sequence is output through the output module.

[0009] Optionally, node feature fusion is performed based on the multidimensional data, and the edge weights of the causal graph are updated based on the fused node features, including: Based on the multidimensional data, feature weighting is performed along the causal path in the causal graph to obtain the current features of each node. The causal path is used to indicate the feature propagation link formed by nodes connected by directed edges in the causal graph. Retrieve the historical features of each node from the previous moment; The current and historical features of the same node are fused, and the edge weights in the causal graph are updated based on the fused node features.

[0010] Optionally, the friction trend sequence is corrected by combining the current vehicle operating condition data with a correction model, and the data output for the friction trend sequence includes: The friction trend sequence and the current vehicle operating condition data are input into the correction model, wherein the correction model has a built-in processing module, correction module, interval generation module and output module; The processing module performs Gaussian distribution processing on the friction trend sequence and the current vehicle operating condition data to obtain the distribution mean and distribution variance of each future time in the friction trend sequence. The distribution mean is used to indicate the residual compensation value, and the distribution variance is used to indicate the prediction confidence. The correction module superimposes the distribution mean corresponding to each future time point onto the friction trend sequence to generate the corrected trend sequence. The interval generation module determines the upper and lower fluctuation range of the friction trend sequence based on the distribution variance at each future time, and generates the confidence interval of the friction trend sequence. The corrected trend sequence and the confidence interval are output through the output module.

[0011] Optionally, fusing the causal map, the friction trend sequence, and the correction data to predict the vehicle's slip probability sequence in the future time period includes: The edge weights and node activation weights in the causal graph are adjusted using the confidence intervals in the correction data to obtain an updated causal graph, wherein the node activation weights are used to indicate the validity of the nodes. An initial friction sequence is generated by combining the multidimensional feature sequence and the updated causal map. The initial friction sequence, the friction trend sequence, and the corrected trend sequence in the correction data are then weighted and fused to obtain a fused friction sequence. The high-risk friction segment in the fused friction sequence is located by using the confidence interval in the correction data and the high-risk friction segment is locally corrected. The high-risk friction segment is used to indicate the time period in which the interval width within the confidence interval is greater than or equal to a set width threshold. By using the edge weights in the updated causal graph as risk weight adjustment factors and combining them with the locally corrected fused friction sequence, the slip probability sequence of the vehicle in the future time period is predicted.

[0012] Optionally, generating corresponding vehicle control commands based on the slip probability sequence includes: Extract the most recent initial slip probability from the slip probability sequence; By combining the slip probabilities of all times in the slip probability sequence, the initial slip probability is smoothed and filtered to obtain the final slip probability of the most recent time. Based on the probability range in which the final slip probability is located, determine the target risk level corresponding to the probability range; Generate corresponding vehicle control commands based on the target risk level.

[0013] Secondly, this application provides a vehicle control device, the device comprising: The acquisition module is used to acquire the multidimensional feature sequence of the vehicle in a historical period, wherein the multidimensional feature sequence is multiple features related to tire slippage collected at each historical moment during the vehicle's operation. The processing module is used to process the multidimensional feature sequence by constructing a model, a prediction model and a correction model in a cascaded manner, to obtain the causal map, friction trend sequence and correction data of the friction trend sequence output by each model respectively. The causal map is used to indicate the causal relationship between the features of each dimension of the vehicle during driving, and the friction trend sequence is used to indicate the change of tire road surface adhesion grip over time in the future period. The fusion module is used to fuse the causal map, the friction trend sequence, and the correction data to predict the slip probability sequence of the vehicle in the future time period. The generation module is used to generate corresponding vehicle control commands based on the slip probability sequence.

[0014] Thirdly, this application provides an electronic device, comprising: at least one communication interface; at least one bus connected to the at least one communication interface; at least one processor connected to the at least one bus; and at least one memory connected to the at least one bus.

[0015] Fourthly, this application also provides a vehicle, the vehicle including a memory and a processor, wherein the memory stores executable program code, and the processor is used to call and execute the executable program code to implement the method described in any of the above.

[0016] The technical solutions provided in this application have the following advantages compared with the prior art: This application employs a three-tiered cascaded architecture consisting of a graph construction model, a prediction model, and a correction model. The three models sequentially output causal graphs, friction trend sequences, and correction data. Lossless data transfer throughout the entire process is achieved via a dedicated cascaded transmission link. Then, multi-source information fusion is performed on the three types of differentiated intermediate results. Combining causal correlation, the temporal variation pattern of tire grip, and prediction error compensation, the slip probability sequence of the vehicle in the future time period is deduced, and the temporal prediction results are transformed into onboard anti-skid control commands. Compared to traditional methods that rely solely on the instantaneous state of the vehicle for post-intervention, this application, by retaining complete link information and possessing multi-dimensional fusion prediction capabilities, can identify future slip risks in advance and issue proactive control actions, eliminating the problem of chassis control response lag and improving driving safety on low-adhesion road surfaces. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0020] Figure 1 A schematic diagram of a vehicle control system provided in an embodiment of this application; Figure 2 A flowchart of a vehicle control method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a vehicle control device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] The following disclosure provides numerous different embodiments or examples for implementing various structures of this application. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the scope of this application. Furthermore, reference numerals and / or letters may be repeated in different examples. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0023] The vehicle control method provided in this application can be applied to processing scenarios where the vehicle processor operates independently, as well as to hardware environments where the vehicle and server process collaboratively. In collaborative processing scenarios, such as Figure 1 As shown, server 103 is connected to vehicle 101 via a network. Vehicle 101 sends a multidimensional feature sequence to server 103. Server 103 predicts the vehicle's slip probability sequence in the future based on the multidimensional feature sequence and generates vehicle control commands. Then, the vehicle control commands are fed back to vehicle 101. Database 105 can be set up on the server or independently of the server to provide data storage services for server 103. The above-mentioned network includes, but is not limited to, wide area network, metropolitan area network or local area network.

[0024] Furthermore, embodiments of this application can also be implemented as functional plug-ins, integrated into the vehicle processor, directly predicting the vehicle's slip probability sequence in future time periods based on multi-dimensional feature sequences, and then generating vehicle control commands.

[0025] The following will describe in detail a vehicle control method provided in this application embodiment, taking a processor applied to a vehicle as an example. Figure 2 As shown, the specific steps are as follows: Step 201: Obtain the multidimensional feature sequence of the vehicle in the historical time period, wherein the multidimensional feature sequence is multiple features related to tire slippage collected at each historical moment during the vehicle's driving process; Step 202: The multidimensional feature sequence is processed by the cascaded graph construction model, prediction model and correction model to obtain the causal graph, friction trend sequence and correction data of friction trend sequence output by each model respectively. The causal graph is used to indicate the causal relationship between the features of each dimension of the vehicle during driving, and the friction trend sequence is used to indicate the change of tire road surface adhesion over time in the future period. Step 203: Fuse the causal map, friction trend sequence, and correction data to predict the vehicle's slip probability sequence in future time periods; Step 204: Generate the corresponding vehicle control commands based on the slip probability sequence.

[0026] In step 201, the processor integrates the driving environment data collected by the vehicle sensors. The driving environment data covers five categories: vehicle motion state, driver operation, vehicle inertia, external environment, and road perception. Vehicle motion state data includes wheel speeds, longitudinal and lateral accelerations, and steering wheel angles. Driver operation data includes brake pedal opening and accelerator pedal opening. Vehicle inertia data includes IMU (Inertial Measurement Unit) data. External environment data includes estimated road adhesion coefficients, ambient temperature and humidity, illumination, and rain / snow weather detection signals. Road perception data includes road images captured by cameras and road point cloud data collected by lidar.

[0027] The processor first standardizes all data sampling frequencies to 50Hz, and uses a unified timestamp to achieve time alignment, resolving the issue of asynchronous signals across different types. Then, it performs standardized preprocessing on the driving environment data, removing abnormal data caused by sensor interference, correcting zero-point offsets resulting from long-term operation of inertial sensors, and using a historical sliding window algorithm to fill in temporarily missing data. Simultaneously, it standardizes the coordinate systems and physical units of all sensors. In addition, the processor calculates the first and second-order rates of change of key signals such as wheel speed and acceleration, supplementing information on the trend of vehicle state changes.

[0028] For two types of unstructured perception data, camera images and LiDAR point clouds, the processor performs separate preprocessing for perception data: completing vehicle image distortion calibration, road point cloud ground segmentation processing, extracting the tire-road contact area, cutting effective image blocks related to road friction, and ensuring that visual, radar road perception data and vehicle dynamic sensing data are completely synchronized and matched on the time axis.

[0029] Finally, the processor encapsulates all features into a standard one-dimensional feature vector through the vehicle's unified data bus. Using a sliding time window based on a 3-second duration, it captures driving environment data from the past 2 to 5 seconds. Each second of driving environment data covers five dimensions: vehicle motion state, driver operation, vehicle inertia, external environment, and road perception, thereby generating a continuous multi-dimensional feature sequence.

[0030] In step 202, the processor inputs the multidimensional feature sequence into three concatenated models, which are then processed sequentially through a graph construction model, a prediction model, and a correction model to complete the serial computation. The three models adopt a cascaded direct transmission method, ensuring that all intermediate data processed by the previous model is completely and without omissions to the next model. There is no data deletion or loss throughout the process, preventing the loss of critical intermediate data during multi-level model transmission and ensuring the integrity and reliability of the entire prediction chain.

[0031] First, a multi-dimensional feature sequence is input into the causal graph construction model. This model, relying on multi-head attention and temporal causal constraints, uncovers the sequential driving relationships between different features, ultimately outputting a causal graph. Each node within the causal graph corresponds to a tire slippage-related feature; directed edges within the causal graph represent the actual causal driving relationship between two features.

[0032] Causal graph modeling can explicitly uncover the real physical causal paths during vehicle operation. A causal path is a chain-like interaction in the causal graph, formed by connecting feature nodes and directed causal edges, representing the progressive influence of preceding features on subsequent features. For example, an early change in tire temperature can trigger subsequent wheel speed fluctuations, leading to a decrease in the road surface friction coefficient. Alternatively, a sudden change in steering wheel angle can induce lateral acceleration disturbances in the vehicle body, simultaneously causing a decrease in the road surface friction coefficient.

[0033] Secondly, the causal graph output by the graph construction model, along with the multidimensional feature sequence, is input into the prediction model. The prediction model combines the multidimensional feature sequence from historical time periods to complete the node feature fusion of the causal graph at each time step, while adaptively adjusting the weights of each directed edge within the causal graph, ultimately outputting a continuously changing friction trend sequence.

[0034] The friction trend sequence is a continuous time-series curve. The horizontal axis of the curve represents each future moment in the future time period, and the vertical axis represents the tire's road grip force at each future moment. The entire curve intuitively reflects the continuous change pattern of tire grip force during vehicle operation. For example, the friction trend sequence can show the road grip force per second within the next 1 to 3 seconds.

[0035] Finally, the friction trend sequence, combined with the vehicle's current driving condition data, is input into the final correction model. During actual vehicle operation, road conditions can change suddenly, and sensors may have slight inherent noise, leading to objective biases in the friction trend sequence. This means the sequence cannot perfectly reflect the actual changes in grip on the road surface. The correction model evaluates the overall prediction bias and generates correction data, which objectively measures the degree of deviation and reliability of the predicted friction trend sequence.

[0036] In this embodiment, the three models have a clear division of labor: the graph construction model mines the inherent causal relationship of multi-dimensional feature sequences, the prediction model infers the continuous change trend of tire grip, and the correction model compensates for prediction bias and controls prediction reliability.

[0037] In step 203, the processor fuses the output results of the three models from different dimensions to predict the slip probability of the vehicle at various future times, generating a slip probability sequence for the future time period. The fused slip probability sequence can simultaneously take into account three layers of effective information: first, it relies on the causal graph to make predictions based on the multi-dimensional feature sequences of the vehicle's historical time period, avoiding deviation from the actual driving logic; second, it relies on the continuous friction trend curve to fully preserve the temporal pattern of the dynamic changes in tire grip; and third, it relies on correction data to compensate for prediction errors and improve the reliability of the prediction results. Through the complementary fusion of multi-source information, this application can make the obtained slip probability sequence more accurate from three dimensions: mechanistic causality, temporal continuity, and prediction uncertainty.

[0038] This application adopts a time-series sliding window mechanism, which uses the multidimensional feature sequence of historical observation windows within 2 to 5 seconds to predict the sliding probability sequence in the future prediction window within 1 to 3 seconds. The historical observation window is driven to slide forward as a whole at fixed sampling steps at each interval. The earliest multidimensional data in the historical observation window is removed, and the latest multidimensional data collected by the vehicle is simultaneously included. The sliding probability sequence is continuously updated in real time to ensure that the prediction results are dynamically adapted to the real-time driving status of the vehicle.

[0039] In step 204, the processor comprehensively analyzes the risk change trend of the slip probability sequence, transforms the predicted continuous time-series probability into risk judgment results, and transforms them into standardized control instructions that can be directly executed by the vehicle chassis system.

[0040] The embodiments of this application can predict changes in the risk of vehicle slippage within the next few seconds. Compared with the prior art, which can only passively perform remedial control after slippage occurs, this application can issue control commands in advance to complete anti-slip operations such as pre-charging the braking system and limiting the torque of the drive system in advance, eliminating the response delay of the chassis controller, effectively suppressing the premature occurrence of wheel slippage, and comprehensively improving the vehicle driving stability on low-adhesion road surfaces such as wet, snowy, and oily surfaces.

[0041] This application employs a three-tiered cascaded architecture consisting of a graph construction model, a prediction model, and a correction model. The three models sequentially output causal graphs, friction trend sequences, and correction data. Lossless data transfer throughout the entire process is achieved via a dedicated cascaded transmission link. Then, multi-source information fusion is performed on the three types of differentiated intermediate results. Combining causal correlation, the temporal variation pattern of tire grip, and prediction error compensation, the probability sequence of vehicle slippage in future time periods is deduced, and the temporal prediction results are transformed into onboard anti-skid control commands. Compared to traditional methods that rely solely on the instantaneous state of the vehicle for post-intervention, this application, by retaining complete link information and possessing multi-dimensional fusion prediction capabilities, can identify future slippage risks in advance and issue proactive control actions, eliminating the problem of chassis control response lag and improving driving safety on low-adhesion road surfaces.

[0042] As an optional implementation, in step 202, the multidimensional feature sequence is processed by a cascaded graph construction model, a prediction model, and a correction model to obtain the causal graph, friction trend sequence, and correction data of the friction trend sequence output by each model, respectively: Step S11: Input the multidimensional feature sequence into the graph construction model to obtain the causal graph output by the graph construction model. The nodes in the causal graph are used to indicate features related to tire slippage, and the edges in the causal graph represent the causal relationships between the features. Step S12: By combining the prediction model with multidimensional feature sequences, the edge weights of the causal graph are adjusted and the node features are fused to generate a friction trend sequence; Step S13: Correct the friction trend sequence by combining the correction model with the current vehicle operating condition data, and output the corrected data of the friction trend sequence. The corrected data includes the corrected trend sequence and confidence interval of the friction trend sequence.

[0043] In step S11, the processor inputs the multidimensional feature sequence into the graph construction model. The graph construction model mines the sequential logic of the mutual influence between various features based on time-series data to generate a causal graph. For example, the graph construction model is a Temporal Causal Discovery Transformer (T-CDT).

[0044] Each node in the causal graph corresponds to features related to tire slippage, such as wheel speed, road surface adhesion coefficient, rain and snow images, and road surface water. The lines between nodes represent the objective mutual influence between two sets of features, which can clearly distinguish key data that are prone to slippage from irrelevant interference data, and make explicit the coupling and driving relationship between multi-dimensional features such as vehicle dynamics, road surface conditions, and external environment.

[0045] The connection weights of the causal graph are global fixed weights obtained from feature statistics over the entire historical period. They can only reflect whether there is a correlation between features under all driving conditions, and cannot match the real-time driving status of the vehicle at every moment.

[0046] In step S12, since the causal graph is only a fixed basic framework, its weights are global average weights, which can only adapt to the normal average driving conditions of the vehicle. However, during actual driving, road conditions, driving operations, and the external environment are all changing. For example, on dry roads, the steering wheel angle has a weaker impact on the coefficient of friction; while on icy and slippery roads, the same steering wheel angle has a significantly stronger impact on the coefficient of friction. The fixed global weights cannot adapt to the real-time changing driving environment at every moment. Therefore, the causal graph needs to be updated.

[0047] The processor inputs the causal graph and multidimensional feature sequence into the prediction model. The prediction model, relying on the entire multidimensional feature sequence, uses a time-based, frame-by-frame rolling iteration method to simultaneously complete node feature fusion and edge weight adjustment. Since the multidimensional feature sequence records the actual operating state of the vehicle at each moment, combined with the time-rolling mechanism, the graph information can be continuously updated frame by frame in chronological order. For example, the prediction model can be a Dynamic Graph Neural Network (DGNN).

[0048] Among them, node feature fusion refers to integrating multiple features corresponding to any given moment into the corresponding node by weighting them layer by layer according to the causal path, so that each node can completely retain the comprehensive state of each feature after interaction through the causal link at the current moment.

[0049] Edge weight adjustment refers to modifying the weight value corresponding to each causal connection in each round of time-rolling calculation, based on the multi-dimensional feature sequence at any given moment. The weight represents the strength of the influence between two features, thereby dynamically adapting to the driving environment at that moment.

[0050] By iteratively updating the graph over time, various driving environment changes such as vehicle acceleration, braking, steering, or road surface accumulating water or ice can be reflected in the causal graph in real time. The interaction between various features is dynamically adjusted to accurately restore the real-time trend of tire-ground grip and ultimately generate a friction trend sequence.

[0051] In step S13, the processor inputs the friction trend sequence and the real-time vehicle operating condition data into the correction model for deviation correction processing and outputs the corresponding correction data. For example, the correction model is a Bayesian Gaussian Process Regression (BGPR) model.

[0052] The friction trend sequence output by the prediction model is a predicted curve of road grip force for future vehicle time periods. This curve is generated based on multi-dimensional data from historical periods and can only reconstruct the evolution of friction changes under historical operating conditions. The prediction results have inherent fixed systematic biases. However, the vehicle's current real-time operating conditions, such as current speed, current steering angle, current road surface perception, and current braking status, represent the current vehicle operating environment and determine the error fluctuation range of the friction trend sequence at various future moments. For example, for the same prediction curve, if the vehicle is currently making a sharp turn or the road surface has just iced, the overall prediction bias of this curve will be significantly larger; while driving at a constant speed on a dry road, the bias of the same curve will be very small. Therefore, the correction model uses the current vehicle operating conditions to correct the error of the future grip force prediction results, thereby generating correction data. The correction data includes the corrected trend sequence and confidence intervals.

[0053] The corrected trend sequence is an optimized curve obtained by superimposing error compensation values ​​on the friction trend sequence. It is used to offset the result deviation caused by sudden changes in road surface and slight noise from sensors, so that the grip prediction value is more in line with the actual driving state of the vehicle.

[0054] The confidence interval represents the reasonable range of fluctuation for the grip value at each future time. The narrower the interval, the more reliable the predicted grip value at that time; the wider the interval, the less reliable the predicted grip value at that time.

[0055] Using the modified trend sequence in combination with the confidence interval can not only improve the accuracy of the grip value, but also intuitively quantify the reliability of the grip prediction value at each moment.

[0056] In this application, the causal graph construction model builds a fixed causal structure based on a global multidimensional feature sequence, clarifying whether there is a causal influence between various features. Then, the causal graph is updated iteratively time-series by the prediction model. For each time step, node feature fusion is performed, and the multidimensional features of a single frame are weighted and aggregated layer by layer along the causal transmission path to the corresponding node, fully preserving the chain coupling effect of multiple features. The edge weights are dynamically updated based on the fused node features, optimizing the globally averaged fixed edge weights into dynamic edge weights adapted to the instantaneous driving environment, thus compensating for the shortcomings of static causal graphs in adapting to real-time road conditions and dynamic changes in driving operations. Finally, combining the dynamically updated causal graph and the vehicle's current operating condition data, the correction model completes error compensation and confidence quantification. A corrected trend sequence is generated through point-by-point residual compensation, and confidence intervals are calculated based on the error variance to quantify the prediction confidence of each future time step, improving the numerical accuracy and referenceability of the future grip prediction.

[0057] As an optional implementation, in step S11, the multidimensional feature sequence is input into the graph construction model to obtain the causal graph output by the graph construction model, including: Step S111: Input the multidimensional feature sequence into the graph construction model, wherein the graph construction model has a built-in multi-head attention module, causal constraint module, graph generation module and output module; Step S112: Discover the temporal dependencies between features using a multi-head attention module; Step S113: Apply causal constraints based on temporal dependencies through the causal constraint module, and determine the causal correlation strength between each feature through the soft mask corresponding to the causal constraints; Step S114: Filter out real causal associations with causal association strength greater than a preset strength threshold through the graph generation module, and construct a causal graph based on each feature and the real causal associations; Step S115: Output the causal graph through the output module.

[0058] In step S111, the processor feeds the multidimensional feature sequence into the graph construction model, which includes a multi-head attention module, a causal constraint module, a graph generation module, and an output module. For example, if the multidimensional feature sequence corresponds to a 2-5 second vehicle history sliding window, a unified encoding operation is first performed on the multidimensional feature sequence within this window to map various features with different dimensions and numerical ranges to the same feature space, eliminating scale differences and obtaining regular and unified encoded features.

[0059] In step S112, the processor inputs the encoded unified features into the multi-head attention module and calculates the numerical linkage pattern between any two sets of features according to the chronological order of the time window, thereby obtaining the temporal dependency relationship. The temporal dependency relationship is the sequential linkage effect of features over time; for example, the feature of road surface water accumulation in the early stage will continuously affect the adhesion coefficient in subsequent time periods. However, the temporal dependency relationship can only statistically analyze the sequential linkage trend of features over time and cannot distinguish whether the linkage is a true causal effect or random data synchronous fluctuation.

[0060] In step S113, the causal constraint module uses a soft mask as the causal constraint and completes the causal constraint based on the temporal dependency relationship. The causal constraint is a set of computational restriction rules that only allow features from earlier moments to affect features from future moments, and do not allow later states to have a reverse effect on the past. The causal constraint can eliminate temporal paradoxes and conforms to the physical logic of the vehicle. For example, the vehicle states exist in a time sequence. Water accumulation on the road will cause friction to decrease. It is impossible for friction to decrease first and then water accumulation to occur. According to this logic, the causal constraint forcibly cuts off the reverse calculation path of future features affecting historical features, which is in line with the real temporal law of vehicle dynamics.

[0061] Each pair of features paired together forms a corresponding soft mask. The number of feature pairs within a window corresponds to the number of independent soft masks. The soft mask is used to determine whether a feature at a given historical moment can effectively influence features at future moments. This soft mask quantifies causal constraints into a continuous weight matrix. Each soft mask stores a value between 0 and 1, representing the degree to which a preceding feature is permissible to have a causal influence on a subsequent feature. A value closer to 1 indicates a strong causal relationship is allowed; a value closer to 0 indicates no real causal interaction and the causal relationship needs to be significantly weakened.

[0062] The standardized score of temporal dependency represents the degree of linkage between two sets of features over time, and is a value in the range of 0 to 1. The soft mask is also a weight value in the range of 0 to 1. The processor multiplies the soft mask value with the temporal dependency one by one. By using the size of the soft mask value, the causal validity of each set of temporal dependencies is quantitatively constrained, and finally the causal association strength between each pair of features is generated.

[0063] A higher soft mask value results in a stronger causal relationship after multiplication, indicating a more significant causal influence of preceding features on subsequent features. Conversely, a lower soft mask value significantly weakens the causal relationship after multiplication, suggesting that the temporal dependencies have almost no real causal effect. For example, in icy and snowy road scenarios, the soft mask values ​​corresponding to the pairing of road adhesion coefficient and wheel speed are relatively large, resulting in a stronger causal relationship after multiplication. Conversely, the soft mask values ​​corresponding to the pairing of ambient light and wheel speed are close to 0, resulting in an extremely low causal relationship after multiplication.

[0064] In step S114, the graph generation module introduces structural sparsity regularization constraints, retaining only causal relationships with a strength exceeding a preset threshold as true causal relationships. This filtering process filters out false causal relationships caused by random noise and instantaneous driving environment fluctuations, reducing invalid data from participating in subsequent graph structure calculations, minimizing computational interference, and improving the overall interpretability of the graph. The processor uses features as graph nodes and the selected true causal relationships as directed edges to generate a causal graph.

[0065] In step S115, the output module outputs the causal graph.

[0066] This application first unifies features through a feature encoding layer to eliminate computational errors caused by inconsistent dimensions; secondly, it quantifies the temporal dependencies of each set of features over time to fully capture the temporal linkage patterns of various features during vehicle operation; thirdly, it uses a soft mask as a concrete causal constraint to perform weighted correction on the temporal dependencies, accurately distinguishing between false causal relationships and true causal relationships, and objectively quantifying the strength of the true causal influence between each pair of features; finally, it combines structural sparsity regularization constraints to eliminate false causal relationships and achieve the accuracy of the constructed causal graph.

[0067] As an optional implementation, in step S12, the friction trend sequence is generated by adjusting the edge weights and fusing node features of the causal graph using a prediction model combined with a multidimensional feature sequence, including: Step S121: Input the causal graph and multidimensional feature sequence into the prediction model, wherein the prediction model has a built-in graph attention module, graph reconstruction module, inference module and output module; Step S122: The graph attention module extracts the multidimensional data corresponding to each historical moment in the multidimensional feature sequence in chronological order, and performs node feature fusion based on the multidimensional data. The edge weights of the causal graph are updated based on the fused node features. Step S123: After completing the fusion of node features and the update of edge weights for all historical moments in the historical period, a dynamic causal graph is obtained through the graph reconstruction module; Step S124: The dynamic causal graph is extrapolated according to the time sequence of future periods through the extrapolation module to generate a friction trend sequence covering future periods.

[0068] Step S125: Output the friction trend sequence through the output module.

[0069] In step S121, the processor inputs the causal graph and the multidimensional feature sequence into the prediction model. The prediction model integrates a graph attention module, a graph reconstruction module, an inference module, and an output module.

[0070] In step S122, the graph attention module reads the multidimensional feature sequence frame by frame in time-stamp order, and extracts the multidimensional data corresponding to each historical moment in the multidimensional feature sequence. This multidimensional data includes all tire slippage-related data such as wheel speed, pedal opening, IMU data, road perception data, and external environment data. The time-series reading module caches each frame of data separately, strictly following the chronological logic, without disrupting the temporal order, and restoring the real driving logic of vehicle state changes over time.

[0071] Multidimensional data at a single moment can only reflect instantaneous data and cannot reflect the long-term effects of features being transmitted and accumulated along causal links. For example, water accumulation on the road surface will continuously reduce the road surface adhesion coefficient, and the decrease in the road surface adhesion coefficient will further increase the risk of wheel slippage. Data from a single moment cannot capture this chain-like transmission relationship. Therefore, the prediction model relies on the directed causal transmission path built into the causal graph to perform node feature fusion.

[0072] In the causal graph, each node forms a fixed causal transmission path through directed edges, representing the driving transmission relationship between features. At each historical moment, multidimensional data is first transmitted and weighted along each causal path layer by layer. Combined with the node gating module to filter out abnormal changes caused by sensor clutter, the node features of each node after fusion at that historical moment are obtained. The fused node features can capture the superposition effect of multiple features synchronously coupled along the causal link in the current frame, filter feature changes caused by instantaneous sensor noise, and fully carry the comprehensive state of each feature after interaction through the causal link under the combined action of the current road surface, driving operation, and external environment.

[0073] After node fusion is completed, the updated node features undergo two changes. First, the nodes are superimposed with multi-level features that are propagated along the causal path, and are no longer isolated single-moment data. Second, the node features change dynamically in real time with external environmental factors such as road conditions, driver operation, rain, snow, and sunlight, which intuitively reflects the comprehensive influence intensity of the interaction of each set of features through the causal link under the current driving environment.

[0074] The graph attention module reads the updated node features, requantifies the interaction strength between pairs of nodes via causal paths, and replaces the original edge weights of the causal graph with these values, completing the iterative update of the edge weights of the causal graph. The larger the edge weight value, the stronger the causal driving effect of the corresponding two features in the current driving environment.

[0075] For example, on dry roads, the steering wheel angle has a weaker impact on the road adhesion coefficient, and the corresponding edge weight is lower; when a vehicle enters an icy road, the steering wheel angle has a stronger impact on the road adhesion coefficient, and the features after node fusion change significantly. The graph attention module will simultaneously increase the weight of this causal edge, reflecting the strong driving effect of steering action on grip attenuation.

[0076] In step S123, the prediction model repeats the above process, extracts multidimensional data for the next historical moment, completes node feature fusion along the causal path, recalculates and updates edge weights based on updated node features, and iteratively updates the entire causal graph until all features of all historical moments within the window have completed one round of calculation.

[0077] Once the node feature fusion and edge weight dynamic update along the causal path have been completed for all historical moments, a dynamic causal graph is obtained through the graph reconstruction module. This graph is an intermediate product that reflects the real-time causal coupling relationship of each feature within the historical sliding window as the vehicle travel time changes dynamically.

[0078] In step S124, the extrapolation module uses the dynamic causal graph as a basis and extrapolates forward step by step according to the time sequence of future time periods. At each future time, node feature fusion and edge weight update are still performed. At the end of each future time period, the prediction model determines the target node located at the endpoint. This target node summarizes the features transmitted from all other nodes. Then, the adhesion grip force value is determined based on the features of the target node. In this way, after all future time periods are completed, the features of the target node are spliced ​​into a continuous curve in chronological order, and finally a friction trend sequence covering future time periods is generated.

[0079] In step S125, the output module outputs the friction trend sequence.

[0080] In this application, the prediction model first reads multi-dimensional data frame by frame in a historical time period, and iteratively completes node feature fusion and edge weight updates in a time-series loop. It continuously corrects the mutual influence strength between features at different times, allowing the causal graph to dynamically adapt to real-time changes in road surface and driving operations, instead of using fixed global correlation weights. Based on the iteratively optimized causal graph, the changes in road grip force are inferred, which can accurately restore the coupling effect of each feature accumulated over time, filter out instantaneous noise interference from sensors, and obtain the adjusted causal graph. In future time periods, the same method is used to extrapolate and calculate forward time by time to obtain the features of the target node at the endpoint at each future time, thereby inferring the road grip force at each future time. The generated friction trend sequence closely matches the actual road friction changes in future time periods, significantly improving the accuracy of grip force time-series prediction.

[0081] As an optional implementation, node feature fusion is performed based on multidimensional data, and the edge weights of the causal graph are updated based on the fused node features. This includes: weighting features along the causal path in the causal graph based on multidimensional data to obtain the current features of each node; retrieving the historical features of each node at the previous moment; fusing the current features and historical features of the same node, and updating the edge weights in the causal graph based on the fused node features.

[0082] For multidimensional data at any historical moment within the window, the processor follows the causal path within the causal graph, performs feature weighting operations based on the weight of each edge, and follows the inherent causal transmission relationship between features. This weighting process is applied to the features of each node in the causal graph, ultimately yielding the current features corresponding to all nodes. The current features represent the real-time effects of each feature within the instantaneous driving environment of the current historical moment, based on the latest causal relationship logic.

[0083] The processor retrieves the historical features of each node from the previous moment. The historical features record the strength of the causal relationship between the features under the multidimensional data of the previous moment. The historical features can reflect the continuous and lagging effect of the causal relationship of the previous moment on the current feature.

[0084] Vehicle movement is a continuous temporal process. Features such as road conditions and driver actions have a lagging, cumulative effect on tire grip; features from the previous moment do not immediately disappear with graph iteration. The current causal graph only adapts to the immediate driving environment and reflects the causal relationships between features at the current moment, failing to retain the ongoing impact of the previous moment's driving environment. In contrast, the previous moment's causal graph preserves the dynamically adjusted edge weights and causal relationships, representing the actual feature coupling relationships from the previous moment. By utilizing its computational history features, the cumulative effect of the preceding driving environment's continuous propagation along the causal chain can be reconstructed, compensating for the loss of temporal lag effects in a single real-time graph.

[0085] The processor fuses the current and historical features of the same node. The current features capture instantaneous driving environment disturbances such as sudden road changes and pedal operation changes, while the historical features retain the continued effect of the features from the previous moment. After the two are fused, the node features simultaneously contain driving environment information from both the previous and previous time periods, fully restoring the continuous driving environment of the vehicle and avoiding the information bias caused by relying solely on features from a single moment.

[0086] The fused node features simultaneously carry the causal relationships of both the old and new versions of the causal graph. The edge weights updated based on these fused node features can adapt to the latest multidimensional time-series features while retaining the cumulative influence of previous driving states, making the dynamic evolution of the causal graph more consistent with the real physical logic of tire slippage.

[0087] As an optional implementation, in step S13, the friction trend sequence is corrected by combining the correction model with the current vehicle operating condition data, and the corrected data of the friction trend sequence is output, including: Step S131: Input the friction trend sequence and the current vehicle operating condition data into the correction model, wherein the correction model has a built-in processing module, correction module, interval generation module and output module; Step S132: The friction trend sequence and the current vehicle operating condition data are processed by Gaussian distribution through the processing module to obtain the distribution mean and distribution variance of each future time in the friction trend sequence. The distribution mean is used to indicate the residual compensation value, and the distribution variance is used to indicate the prediction confidence. Step S133: The distribution mean corresponding to each future time moment is superimposed onto the friction trend sequence through the correction module to generate the corrected trend sequence; Step S134: The interval generation module determines the upper and lower fluctuation range of the friction trend sequence based on the distribution variance of each future time, and generates the confidence interval of the friction trend sequence. Step S135: Output the corrected trend sequence and confidence interval through the output module.

[0088] In step S131, the processor inputs the friction trend sequence and the current vehicle operating condition data into the correction model. The correction model has a built-in processing module, correction module, interval generation module and output module.

[0089] In step S132, the processing module models the prediction error throughout the entire process of friction trend prediction as a function that follows a Gaussian distribution, and outputs a complete set of probability distribution results, including the distribution mean and distribution variance corresponding to each future moment in the friction trend sequence.

[0090] Among them, the distribution mean is the residual compensation value, which is used to make up for the prediction deviation between the friction trend sequence and the actual friction data; The variance of the distribution is used to represent the reliability of the prediction. The smaller the variance value, the smaller the fluctuation of the prediction result at that moment, and the higher the reliability of the prediction. The larger the variance value, the more obvious the fluctuation of the prediction result at that moment, and the lower the reliability of the prediction.

[0091] For example, when the vehicle is in a stable driving state, the friction data changes gradually, the corresponding distribution variance is smaller, and the prediction results are more reliable; when the vehicle's driving state changes abruptly, the friction data fluctuates more, the corresponding distribution variance increases, and the reliability of the prediction results decreases accordingly.

[0092] In step S133, the correction module superimposes the distribution mean values ​​corresponding to each future time point onto the predicted values ​​of the friction trend sequence at the corresponding time point. The friction trend sequence itself has a fixed prediction residual. After superimposing the residual compensation values ​​corresponding to the distribution mean values, the inherent prediction bias can be offset point-to-point, correcting the prediction error of the friction trend sequence and ultimately generating a more accurate corrected trend sequence.

[0093] In step S134, the interval generation module uses the friction trend sequence at each future time as the central benchmark, and superimposes and subtracts the distribution variance at the corresponding time to obtain the upper and lower boundary data of the predicted value at that time. The upper boundary data of all future times are concatenated sequentially, and the lower boundary data of all times are concatenated sequentially to form a complete and continuous confidence interval. The confidence interval is used to mark the error fluctuation space of the friction prediction result at each time, intuitively reflecting the uncertainty of the prediction result. The distribution variance directly determines the width of the confidence interval. The smaller the distribution variance, the narrower the confidence interval, the smaller the fluctuation range of the prediction result, and the more stable the result; the larger the distribution variance, the wider the confidence interval, and the stronger the uncertainty of the prediction result.

[0094] For example, when a vehicle is traveling straight at a constant speed, the prediction error is small, the distribution variance is small, and the generated confidence interval is narrower, indicating that the friction prediction result for this segment fluctuates less and the result is reliable. When a vehicle changes lanes or accelerates / decelerates, the driving state fluctuates more, the prediction error increases accordingly, the distribution variance increases, and the confidence interval widens synchronously, indicating that the prediction result for this segment is more uncertain. In step S135, the output module outputs the corrected trend sequence and confidence interval as correction data.

[0095] The modified model simultaneously solves the residual compensation value and prediction confidence for each future moment using a Gaussian distribution. On the one hand, it uses the distribution mean to perform point-to-point compensation on the friction trend sequence to offset inherent prediction bias and improve the accuracy of grip numerical prediction. On the other hand, it relies on the distribution variance to delineate floating intervals and generate confidence intervals, intuitively quantifying the prediction reliability at each moment. This allows the entire prediction chain to simultaneously possess error correction and uncertainty quantification capabilities, which is beneficial for improving the stability of prediction results under complex and variable road conditions.

[0096] As an optional implementation, in step 203, the causal map, friction trend sequence, and correction data are fused to predict the vehicle's slip probability sequence in future time periods, including: Step S21: By correcting the confidence intervals in the data, adjust the edge weights and node activation weights in the causal graph to obtain the updated causal graph, wherein the node activation weights are used to indicate the validity of the nodes. Step S22: Combine the multidimensional feature sequence and the updated causal map to generate an initial friction sequence, and then perform weighted fusion of the initial friction sequence, the friction trend sequence, and the corrected trend sequence in the correction data to obtain a fused friction sequence; Step S23: Locate the high-risk friction segment in the fused friction sequence by using the confidence interval in the corrected data and perform local correction on the high-risk friction segment; Step S24: Use the edge weights in the updated causal graph as risk weight adjustment factors, and combine them with the locally corrected fused friction sequence to predict the vehicle's slip probability sequence in future time periods.

[0097] In step S21, the processor inputs the causal graph, friction trend sequence and correction data into the fusion model, which includes a structural layer, a temporal layer and an output layer.

[0098] The confidence interval in the calibration data can quantify the range of fluctuation of the predicted tire road grip at each future moment. In the structural layer, the fusion model reads the width of the confidence interval corresponding to each future moment, and adjusts the edge weights and node activation weights inside the causal graph in sync with the width of the confidence interval, and finally generates the updated causal graph.

[0099] The node activation weight refers to the effectiveness of each stored feature node within the causal graph. When the confidence interval width at a certain moment is greater than a preset threshold, it indicates that the gripping prediction at that moment has low reliability. The processor reduces the activation weight of the node containing the corresponding feature at that moment, weakening the interference effect of this group of unstable features in subsequent calculations. Conversely, when the confidence interval width at a certain moment is less than or equal to the preset threshold, the processor increases the activation weight of the node containing the corresponding feature at that moment, amplifying the role of stable and effective features in the computation process.

[0100] Edge weights represent the strength of the causal relationship between two sets of features in a causal graph. When the confidence interval width is less than or equal to a preset threshold at a certain moment, the edge weights between the nodes containing the corresponding features are reduced. When the confidence interval width is greater than the preset threshold at a certain moment, the edge weights between the nodes containing the corresponding features are increased simultaneously.

[0101] In step S22, in the time series layer, the fusion model combines the multidimensional feature sequence and the updated causal map to generate the initial friction sequence. At this time, there are three independent curves, namely the initial friction sequence generated by the fusion model, the friction trend sequence output by the prediction model, and the corrected trend sequence output by the correction model.

[0102] The fusion model performs a consistency check mechanism on the three curves obtained to determine the prediction deviation of the three curves at the corresponding time. When the prediction deviation between any two curves is greater than a preset threshold, the result of the current period is directly discarded and the fusion friction sequence of the previous period is used as the final output. This can effectively avoid the abnormal output of a single model from interfering with the judgment of the final risk level, thereby improving the accuracy of vehicle control commands.

[0103] If the prediction deviation between any two curves is less than or equal to a preset threshold, the fusion model assigns specific weights to the three different trend curves. The processor then performs a weighted fusion operation on the three trend curves and their respective weights, outputting a fused friction sequence. This multi-curve weighted fusion method compensates for the prediction deviation problem present in a single road grip trend curve.

[0104] In step S23, in the output layer, the fusion model reads the confidence interval width corresponding to each future moment in the fused friction sequence time-by-time, and defines the time periods where the confidence interval width is greater than or equal to a set width threshold as high-risk friction segments. High-risk friction segments indicate that the grip prediction deviation is larger in these segments. The processor then corrects these high-risk friction segments, making the curve fluctuation confidence interval width smaller than the set width threshold, thereby reducing the fluctuation error of grip prediction.

[0105] In step S24, the fusion model obtains the fusion friction sequence after local compensation correction. Then, the edge weights in the updated causal graph are used as risk weight adjustment factors. The corrected fusion friction sequence is corrected again through the risk weight adjustment factors to amplify the influence weight of strong causal features on slip probability and reduce the interference caused by weak causal features. Finally, the slip probability values ​​of all times are spliced ​​in chronological order to generate a slip probability sequence covering future time periods.

[0106] The processor accurately identifies high-risk friction sections with large prediction deviations by using confidence intervals. It first uses a modified trend sequence to complete the initial local numerical compensation, compressing the prediction fluctuation range. Then, it uses the edge weights of the causal graph to carry out a second risk-weighted correction, distinguishing the strength of risks according to the physical causal laws of vehicle slippage. The two corrections are progressively applied, significantly reducing the prediction error of slippage probability in variable low-adhesion road conditions such as ice, snow, and water accumulation, and outputting a slippage probability sequence with continuous time and higher reliability.

[0107] As an optional implementation, in step 204, generating the corresponding vehicle control command based on the slip probability sequence includes: Extract the initial slip probability of the most recent time step from the slip probability sequence; By combining the slip probabilities of all time steps in the slip probability sequence, the initial slip probability is smoothed and filtered to obtain the final slip probability of the most recent time step. Determine the target risk level corresponding to the probability interval based on the probability interval in which the final slip probability lies; Generate corresponding vehicle control commands based on the target risk level.

[0108] The processor extracts the initial slip probability of the most recent moment from the slip probability sequence. For example, the slip probability sequence stores the slip probability of the vehicle in each second from the next 1 to 3 seconds. The processor extracts the value corresponding to the 1st second as the initial slip probability.

[0109] The processor combines the slip probability at all moments within the slip probability sequence and applies a smoothing filter to the initial slip probability to calculate the final slip probability at the most recent moment. Since brief disturbances can cause a sudden spike or drop in the slip probability within a single second, directly using this instantaneous value to determine the risk level would cause the vehicle control program to frequently switch operating modes, resulting in jerking during driving. By combining the slip probability at all moments with the slip probability from several seconds before and after, the processor can calculate the final slip probability and smooth out sudden abnormal fluctuations.

[0110] The processor has multiple pre-defined consecutive probability intervals, each associated with a risk level. The processor compares the final slip probability within this interval to determine the corresponding target risk level. Different target risk levels correspond to different degrees of slippage danger; a higher risk level indicates a greater likelihood of the vehicle slipping.

[0111] The processor retrieves preset mapping relationships, determines the corresponding vehicle control commands based on the target risk level, and sends them out to various electronic control execution units in the chassis. Different vehicle control commands contain differentiated intervention logic, enabling continuous monitoring for low-risk situations and early restriction of drive torque for high-risk situations, thus implementing anti-slip operations.

[0112] Table 1 provides the risk level corresponding to each slip probability, and the on-board control command corresponding to each risk level.

[0113] Table 1

[0114] This application uses full-time slip probability for smoothing filtering and then determines the risk level. It relies on the comprehensive calculation of slip probability in the preceding and following time sequences to eliminate sudden changes in slip probability caused by instantaneous data jumps, avoids frequent switching of control strategies by the vehicle chassis controller, and can also output vehicle control commands according to the actual degree of slip risk. On low-adhesion roads, it can issue pre-anti-slip operations in advance, taking into account both vehicle driving smoothness and active anti-slip safety performance under low-adhesion conditions.

[0115] Based on the same technical concept, this application provides a vehicle control device, such as... Figure 3 As shown, the device includes: The acquisition module 301 is used to acquire the multidimensional feature sequence of the vehicle in a historical period, wherein the multidimensional feature sequence consists of multiple features related to tire slippage collected at each historical moment during the vehicle's operation. The processing module 302 is used to process the multidimensional feature sequence by constructing a model, a prediction model and a correction model in a cascaded manner, to obtain the causal map, friction trend sequence and correction data of the friction trend sequence output by each model respectively. The causal map is used to indicate the causal relationship between the features of each dimension of the vehicle during driving, and the friction trend sequence is used to indicate the change of tire road surface adhesion grip over time in the future period. The fusion module 303 is used to fuse the causal map, friction trend sequence and correction data to predict the vehicle slip probability sequence in the future period. The generation module 304 is used to generate corresponding vehicle control commands based on the slip probability sequence.

[0116] Optionally, the processing module 302 is used for: The multidimensional feature sequence is input into the graph construction model to obtain the causal graph output by the graph construction model. The nodes in the causal graph are used to indicate the features related to tire slip, and the edges in the causal graph represent the causal relationships between the features. By combining a prediction model with multidimensional feature sequences, edge weights are adjusted and node features are fused into the causal graph to generate a friction trend sequence. The friction trend sequence is corrected by combining the correction model with the current vehicle operating condition data, and the corrected data of the friction trend sequence is output. The corrected data includes the corrected trend sequence and confidence interval of the friction trend sequence.

[0117] Optionally, the processing module 302 is specifically used for: Input multidimensional feature sequences into the graph to construct a model; Discover the temporal dependencies between features using a multi-head attention module; Causal constraints are applied based on temporal dependencies, and the strength of causal association between features is determined by the soft mask corresponding to the causal constraints. The system filters out true causal relationships with a causal relationship strength greater than a preset strength threshold, and constructs and outputs a causal graph based on each feature and the true causal relationships.

[0118] Optionally, the processing module 302 is specifically used for: Multidimensional feature sequences are input into the graph construction model, which includes a multi-head attention module, a causal constraint module, a graph generation module, and an output module. Discover the temporal dependencies between features using a multi-head attention module; The causal constraint module applies causal constraints based on temporal dependencies and determines the strength of causal association between features through the soft mask corresponding to the causal constraints. The causal constraints are used to limit the positive influence of historical features on future features and prohibit future features from influencing historical features in the opposite direction. The soft mask is used to quantize the causal constraints into a continuous weight matrix. The graph generation module filters out true causal associations with a causal association strength greater than a preset strength threshold, and constructs a causal graph based on each feature and the true causal associations. The causal graph is output through the output module.

[0119] Optionally, the processing module 302 is specifically used for: The causal graph and multidimensional feature sequence are input into the prediction model, which includes a graph attention module, a graph reconstruction module, an inference module, and an output module. The graph attention module extracts multidimensional data corresponding to each historical moment in the multidimensional feature sequence in chronological order, and performs node feature fusion based on the multidimensional data. The edge weights of the causal graph are then updated based on the fused node features. After completing the fusion of node features and the update of edge weights for all historical moments in the historical period, a dynamic causal graph is obtained through the graph reconstruction module. The dynamic causal graph is extrapolated according to the time sequence of future periods through the extrapolation module, generating a friction trend sequence covering future periods; The friction trend sequence is output through the output module.

[0120] Optionally, the processing module 302 is specifically used for: Based on multidimensional data, feature weighting is performed along the causal path in the causal graph to obtain the current features of each node. The causal path is used to indicate the feature propagation link formed by nodes connected by directed edges in the causal graph. Retrieve the historical features of each node from the previous moment; The current and historical features of the same node are fused, and the edge weights in the causal graph are updated based on the fused node features.

[0121] Optionally, the processing module 302 is specifically used for: The friction trend sequence and current vehicle operating condition data are input into the correction model, which includes a processing module, a correction module, an interval generation module, and an output module. The friction trend sequence and current vehicle operating condition data are processed by Gaussian distribution through the processing module to obtain the distribution mean and distribution variance of each future time in the friction trend sequence. The distribution mean is used to indicate the residual compensation value, and the distribution variance is used to indicate the prediction confidence. The correction module superimposes the distribution mean corresponding to each future time point onto the friction trend sequence to generate a corrected trend sequence. The interval generation module determines the upper and lower fluctuation range of the friction trend sequence based on the distribution variance at each future time point, and generates the confidence interval of the friction trend sequence. The output module outputs the corrected trend sequence and confidence interval.

[0122] Optionally, the fusion module 303 is used for: By correcting the confidence intervals in the data, the edge weights and node activation weights in the causal graph are adjusted to obtain the updated causal graph, where the node activation weights are used to indicate the validity of the nodes. An initial friction sequence is generated by combining the multidimensional feature sequence and the updated causal map. The initial friction sequence, the friction trend sequence, and the corrected trend sequence in the correction data are then weighted and fused to obtain the fused friction sequence. High-risk friction segments in the fused friction sequence are located by correcting the confidence intervals in the data and local correction is performed on the high-risk friction segments. The high-risk friction segments are used to indicate the time period in which the interval width within the confidence interval is greater than or equal to a set width threshold. By using the edge weights in the updated causal graph as risk weight adjustment factors and combining them with the locally corrected fused friction sequence, the slip probability sequence of the vehicle in future time periods is predicted.

[0123] Optionally, the generation module 304 is used for: Extract the initial slip probability of the most recent time step from the slip probability sequence; By combining the slip probabilities of all time steps in the slip probability sequence, the initial slip probability is smoothed and filtered to obtain the final slip probability of the most recent time step. Determine the target risk level corresponding to the probability interval based on the probability interval in which the final slip probability lies; Generate corresponding vehicle control commands based on the target risk level. For example... Figure 4 As shown, this application provides an electronic device including a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404.

[0124] Memory 403 is used to store computer programs.

[0125] In one embodiment of this application, the processor 401, when executing the program stored in the memory 403, implements the vehicle control method provided in any of the foregoing method embodiments.

[0126] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the vehicle control method provided in any of the foregoing method embodiments.

[0127] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0128] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0129] It should be understood that the terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “described” as used herein may also include the plural forms. The terms “comprising,” “including,” “containing,” and “having” are inclusive and therefore indicate the presence of the stated features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or combinations thereof. The method steps, processes, and operations described herein are not construed as requiring them to be performed in a particular order described or illustrated unless the order of performance is explicitly indicated. It should also be understood that additional or alternative steps may be used.

[0130] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A vehicle control method, characterized in that, The method includes: Obtain a multidimensional feature sequence of the vehicle over a historical period, wherein the multidimensional feature sequence consists of multiple features related to tire slippage collected at each historical moment during the vehicle's operation. The multidimensional feature sequence is processed by a cascaded graph construction model, a prediction model, and a correction model to obtain the causal graph, friction trend sequence, and correction data of the friction trend sequence output by each model. The causal graph is used to indicate the causal relationship between the features of each dimension of the vehicle during driving, and the friction trend sequence is used to indicate the change of tire road surface adhesion over time in the future. By fusing the causal map, the friction trend sequence, and the correction data, a slip probability sequence of the vehicle in the future time period is predicted. The corresponding vehicle control commands are generated based on the slip probability sequence.

2. The method according to claim 1, characterized in that, The multidimensional feature sequence is processed by a cascaded graph construction model, prediction model, and correction model to obtain the causal graph, friction trend sequence, and correction data of the friction trend sequence output by each model, respectively: The multidimensional feature sequence is input into the graph construction model to obtain the causal graph output by the graph construction model. The nodes in the causal graph are used to indicate features related to tire slippage, and the edges in the causal graph represent the causal relationships between the features. The friction trend sequence is generated by adjusting the edge weights and fusing node features of the causal graph using a prediction model combined with the multidimensional feature sequence. The friction trend sequence is corrected by combining the correction model with the current vehicle operating condition data, and the corrected data of the friction trend sequence is output. The corrected data includes the corrected trend sequence and confidence interval of the friction trend sequence.

3. The method according to claim 2, characterized in that, Inputting the multidimensional feature sequence into the graph construction model yields the following causal graphs output by the graph construction model: The multidimensional feature sequence is input into the graph construction model, wherein the graph construction model has a built-in multi-head attention module, a causal constraint module, a graph generation module, and an output module; The temporal dependencies between features are discovered through the multi-head attention module. The causal constraint module applies causal constraints based on the temporal dependency relationship and determines the causal correlation strength between features through the soft mask corresponding to the causal constraint. The causal constraint is used to limit only historical features to positively influence future features and prohibit future features from negatively influencing historical features. The soft mask is used to quantize the causal constraint into a continuous weight matrix. The graph generation module filters out real causal associations with a causal association strength greater than a preset strength threshold, and constructs a causal graph based on each feature and the real causal associations. The causal graph is output through the output module.

4. The method according to claim 2, characterized in that, The friction trend sequence is generated by adjusting the edge weights and fusing node features of the causal graph using a prediction model combined with the multidimensional feature sequence, including: The causal graph and the multidimensional feature sequence are input into the prediction model, wherein the prediction model has a built-in graph attention module, graph reconstruction module, inference module and output module; The graph attention module extracts multidimensional data corresponding to each historical moment in the multidimensional feature sequence in chronological order, and performs node feature fusion based on the multidimensional data. The edge weights of the causal graph are then updated based on the fused node features. After completing the fusion of node features and the update of edge weights for all historical moments in the historical period, a dynamic causal graph is obtained through the graph reconstruction module. The inference module infers the dynamic causal graph according to the time sequence of the future period, generating a friction trend sequence covering the future period; The friction trend sequence is output through the output module.

5. The method according to claim 4, characterized in that, The node features are fused based on the multidimensional data, and the edge weights of the causal graph are updated based on the fused node features, including: Based on the multidimensional data, feature weighting is performed along the causal path in the causal graph to obtain the current features of each node. The causal path is used to indicate the feature propagation link formed by nodes connected by directed edges in the causal graph. Retrieve the historical features of each node from the previous moment; The current and historical features of the same node are fused, and the edge weights in the causal graph are updated based on the fused node features.

6. The method according to claim 2, characterized in that, The friction trend sequence is corrected by combining the correction model with the current vehicle operating condition data, and the output data of the friction trend sequence includes: The friction trend sequence and the current vehicle operating condition data are input into the correction model, wherein the correction model has a built-in processing module, correction module, interval generation module and output module; The processing module performs Gaussian distribution processing on the friction trend sequence and the current vehicle operating condition data to obtain the distribution mean and distribution variance of each future time in the friction trend sequence. The distribution mean is used to indicate the residual compensation value, and the distribution variance is used to indicate the prediction confidence. The correction module superimposes the distribution mean corresponding to each future time point onto the friction trend sequence to generate the corrected trend sequence. The interval generation module determines the upper and lower fluctuation range of the friction trend sequence based on the distribution variance at each future time, and generates the confidence interval of the friction trend sequence. The corrected trend sequence and the confidence interval are output through the output module.

7. The method according to claim 1, characterized in that, By fusing the causal map, the friction trend sequence, and the correction data, the predicted slip probability sequence of the vehicle in the future time period includes: The edge weights and node activation weights in the causal graph are adjusted using the confidence intervals in the correction data to obtain an updated causal graph, wherein the node activation weights are used to indicate the validity of the nodes. An initial friction sequence is generated by combining the multidimensional feature sequence and the updated causal map. The initial friction sequence, the friction trend sequence, and the corrected trend sequence in the correction data are then weighted and fused to obtain a fused friction sequence. The high-risk friction segment in the fused friction sequence is located by using the confidence interval in the correction data and the high-risk friction segment is locally corrected. The high-risk friction segment is used to indicate the time period in which the interval width within the confidence interval is greater than or equal to a set width threshold. By using the edge weights in the updated causal graph as risk weight adjustment factors and combining them with the locally corrected fused friction sequence, the slip probability sequence of the vehicle in the future time period is predicted.

8. The method according to claim 1, characterized in that, Generating corresponding vehicle control commands based on the slip probability sequence includes: Extract the most recent initial slip probability from the slip probability sequence; By combining the slip probabilities of all times in the slip probability sequence, the initial slip probability is smoothed and filtered to obtain the final slip probability of the most recent time. Based on the probability range in which the final slip probability is located, determine the target risk level corresponding to the probability range; Generate corresponding vehicle control commands based on the target risk level.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method described in any one of claims 1-8.

10. A vehicle, characterized in that, The vehicle includes a memory and a processor, wherein the memory stores executable program code, and the processor is used to call and execute the executable program code to implement the method as described in any one of claims 1 to 8.

Images