Fault-tolerant control method for steer-by-wire of automobile based on body model prediction technology

Abstract

The application relates to the technical field of steer-by-wire fault-tolerant control, in particular to a steer-by-wire fault-tolerant control method for a car fusing body big model prediction technology, which comprises the following steps: acquiring car running state data; determining a working condition abnormality coefficient based on a yaw rate mapping error, a current time-frequency feature, a steering phase difference and a side slip angle change rate; decoupling a hardware degradation baseline by using a Gaussian process regression to extract a mutation factor; predicting a future risk state by using a mutation factor weighted Transformer model, and constructing a gain adjustment factor and a stability risk coefficient based on the same, so that dynamic fault-tolerant control of the steer-by-wire is realized. The application solves the problem that existing fault-tolerant control lacks predictability and slow degradation covers up mutation risk, and improves the prediction accuracy of unknown risks and the control robustness under dynamic disturbance of the car steer-by-wire system.

Description

Technical Field

[0001] This application relates to the field of fault-tolerant control technology for steer-by-wire, specifically to a fault-tolerant control method for automotive steer-by-wire that integrates embodied large model prediction technology. Background Technology

[0002] With the development of automotive intelligence and electrification, steer-by-wire systems have emerged as the core actuators for autonomous driving. They transmit steering commands via electronic signals instead of mechanical connections, offering advantages such as flexible structure, precise control, and ease of integration with intelligent driving systems. However, their highly electronic nature also presents significant safety challenges: the system relies entirely on electronic components, and sensor drift or controller malfunction directly threatens the vehicle's lateral stability and steering ability, easily leading to loss of control. Therefore, ensuring vehicle stability and controllability when the steer-by-wire system experiences partial failures or performance degradation has become a crucial research direction for intelligent vehicles, making steer-by-wire fault-tolerant control technology a focus of attention.

[0003] Existing steer-by-wire fault-tolerant control systems are mostly based on hardware redundancy, fault diagnosis, or model predictive control (MPC). Essentially, they remain within a "perception-control" paradigm, focusing only on passive responses to the current dynamic state and lacking awareness and foresight of environmental risks. While MPC possesses some feedforward predictive capabilities, it relies on a pre-set deterministic dynamic model for short-term trajectory rolling optimization. Faced with nonlinear unknown disturbances and time-varying operating conditions, this model-based calculation is highly susceptible to model mismatch, reducing the accuracy of the steer-by-wire system's prediction of unknown risks and its control robustness under dynamic disturbances. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a fault-tolerant control method for vehicle steer-by-wire that integrates embodied large model prediction technology, employing the following technical solution: One embodiment of this application provides a fault-tolerant control method for vehicle steer-by-wire that integrates embodied large model prediction technology. The method includes the following steps: The system acquires the steering wheel angle, front wheel angle, vehicle center of gravity sideslip angle, vehicle speed, yaw rate, and current in the steering motor circuit during vehicle operation. By analyzing the changes in yaw rate with steering wheel angle, front wheel angle and vehicle speed, the time-domain relative change and high-frequency characteristics of current in the steering motor circuit, an operational anomaly degree is constructed. Combined with the phase difference between steering wheel angle and front wheel angle, and the time-series change of the vehicle's center of gravity side slip angle change rate, the operational anomaly coefficient of the vehicle at each time point is determined. Gaussian process regression model is used to decouple the slow trend of hardware degradation, obtain the benchmark operating condition anomaly coefficient of the vehicle at each time point, and construct the mutation factor at each time point based on the difference between the operating condition anomaly coefficient and the benchmark operating condition anomaly coefficient. By dynamically adjusting the attention calculation weights of samples in the pre-trained Transformer prediction model using mutation factors, the future operating condition anomaly coefficient and future risk state of the vehicle can be predicted. Based on this, a gain adjustment factor and a stability risk coefficient are constructed to perform dynamic fault-tolerant control on the vehicle's steer-by-wire actuator.

[0005] Preferably, the process for constructing the operational anomaly degree is as follows: By fitting the steering wheel angle, front wheel angle, vehicle speed and yaw rate at each time point and in the preset historical time period before that time point, the fitting error of the yaw rate at each time point is obtained. The rate of change of the average current in the steering motor circuit within the preset window before each time point compared to the average current within the preset window before the previous time point is denoted as the time-domain relative change characteristic value of the current in the steering motor circuit at each time point. The result of dividing the weighted sum of the current and its energy proportion at each frequency in the steering motor circuit within the preset window before each time point by the maximum frequency is recorded as the frequency domain characteristic value at each time point. Based on the fitting error, the time-domain relative change feature value, and the frequency-domain feature value, the vehicle's operational anomaly degree at each time point is constructed.

[0006] Preferably, the abnormal operating condition coefficient of the vehicle at each time point is positively correlated with the time-series changes of the abnormal operating degree, the phase difference between the steering wheel angle and the front wheel angle, and the change rate of the vehicle's center of gravity side slip angle.

[0007] Preferably, the process for determining the phase difference between the steering wheel angle and the front wheel angle is as follows: Extract the dominant frequencies of the steering wheel angle and front wheel angle in the frequency domain within the preset window before each time point; If the main frequency of the steering wheel angle and the front wheel angle is the same, then the phase difference at that main frequency is calculated as the phase difference between the steering wheel angle and the front wheel angle at each moment. Otherwise, the phase difference between the steering wheel angle and the front wheel angle at each moment is set to 0.

[0008] Preferably, the quantification process of the time-series change of the vehicle's center of gravity sideslip angle is as follows: The absolute difference between the vehicle's center of gravity sideslip angle at each time point and the adjacent previous time point is divided by the corresponding time interval and recorded as the rate of change of the vehicle's center of gravity sideslip angle at each time point. The ratio of the rate of change of the vehicle's center of gravity sideslip angle at each time point to the average rate of change of the vehicle's center of gravity sideslip angle at all times within the preset window prior to each time point is used as the quantitative result of the temporal change of the rate of change of the vehicle's center of gravity sideslip angle at each time point.

[0009] Preferably, the construction process of the mutation factor at each time point is as follows: Based on the difference between the abnormal coefficient of the vehicle's operating condition and the abnormal coefficient of the benchmark operating condition at each time point, an abnormal operating condition deviation is constructed. Calculate the exponential decay function value of the difference between the abnormal deviation of the operating condition at each time point and the adjacent previous time point, and denot it as the operating condition sudden change decay value; Based on the abnormal deviation of the operating condition and the abrupt decrease value of the operating condition, a mutation factor is constructed at each time point.

[0010] Preferably, the step of dynamically adjusting the attention calculation weights of samples in the pre-trained Transformer prediction model using mutation factors to predict the future operating condition anomaly coefficient and future risk status of the vehicle includes: Extract the temporal features of the car's operation within a preset window before each time point; The temporal features corresponding to each time step are used as input to the pre-trained Transformer prediction model. The mutation factor at each time step is used as the attention weight of the corresponding temporal feature in the pre-trained Transformer prediction model. The future operating condition anomaly coefficient and future risk state at each time step are output. The future risk state includes the future yaw rate and the future vehicle center of gravity sideslip angle.

[0011] Preferably, the construction of the gain adjustment factor and stability risk coefficient includes: Based on the degree of deviation between the future operating condition anomaly coefficient and the operating condition anomaly coefficient at each time point, a gain adjustment factor is constructed; A stability risk coefficient is constructed based on the difference between the future yaw rate and the maximum permissible yaw rate at each moment, as well as the difference between the future vehicle center of gravity sideslip angle and the maximum permissible sideslip angle.

[0012] Preferably, the dynamic fault-tolerant control of the vehicle's steer-by-wire actuator includes: By using the gain adjustment factor and the stability risk coefficient, the torque of the automotive steer-by-wire actuator is optimized in real time, and the optimized torque is used to perform fault-tolerant control on the automotive steer-by-wire actuator.

[0013] Preferably, the real-time optimization of the torque of the vehicle's steer-by-wire actuator includes: Optimized torque of the automotive steer-by-wire actuator at time i The expression is: In the formula, , This represents the gain adjustment factor and stability risk coefficient at time i; This indicates the driver's initial steering torque requirement; This indicates the preset parameter tuning factor.

[0014] This application has at least the following beneficial effects: This application constructs anomaly coefficients by integrating multi-source heterogeneous data such as dynamic mapping error, motor current time-frequency characteristics, steering input-output phase difference, and dynamic changes in sideslip angle. This solves the problem that existing fault-tolerant control relies only on single or local state perception and is difficult to accurately quantify the degree of damage to system coupling relationships. It achieves a comprehensive and unified quantitative assessment of multi-dimensional anomalies such as hardware degradation, execution failure, and road disturbance, providing a reliable risk quantification basis for accurately decoupling fault evolution trends and formulating forward-looking fault-tolerant control strategies. Furthermore, this application fits the operating condition anomaly benchmark under multidimensional environment and operating parameters using a Gaussian process regression model, successfully decouples the slow-changing baseline of hardware degradation from the comprehensive anomaly, and constructs a mutation factor based on the deviation between the actual anomaly and the benchmark and the degree of evolution. This solves the problem of slow degradation accumulation masking sudden instability signals, and realizes accurate weighting of mutation risk-sensitive samples, laying the foundation for subsequent model reinforcement learning of rapid evolution characteristics and improving the foresight and accuracy of predictive control. Ultimately, this application enhances the learning and prediction of key features of sudden failures through a dynamically weighted Transformer model with mutation factors. Based on the predicted future operating condition anomaly coefficient and risk state, a gain adjustment factor and a stability risk coefficient are constructed, realizing the early perception of future instability trends and adaptive closed-loop adjustment of the execution torque. This solves the problems of intervention lag and over-acting caused by the lack of foresight in traditional fault-tolerant control, and improves the prediction accuracy of the automotive steer-by-wire system for unknown risks and the control robustness under dynamic disturbances. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the steps of a fault-tolerant control method for vehicle steer-by-wire that integrates embodied large model prediction technology, as provided in one embodiment of this application. Figure 2 This is a schematic diagram of the mutation factor extraction process provided in one embodiment of this application. Detailed Implementation

[0016] This application provides an embodiment of a fault-tolerant control method for vehicle steer-by-wire that integrates embodied large model prediction technology. Specifically, the following method for fault-tolerant control of vehicle steer-by-wire that integrates embodied large model prediction technology is provided. Please refer to [link / reference]. Figure 1 The method includes the following steps: Step S1: Obtain the steering wheel angle, front wheel angle, vehicle center of gravity sideslip angle, vehicle speed, yaw rate, and current in the steering motor circuit during vehicle operation.

[0017] (1) Driving input data: A magnetically encoded steering angle sensor is placed at the steering column of the steering wheel to collect the steering wheel angle during the operation of the car. The sampling rate is set to 500Hz.

[0018] (2) Execution status data and control link operation data: A Hall current sensor is arranged in the steering motor drive circuit of the steering actuator to collect current data in the steering motor circuit that reflects the execution status. The sampling rate is set to 5kHz.

[0019] (3) Vehicle dynamics state data: Hall sensor is used to collect front wheel steering angle data at the steering knuckle position; MEMS inertial measurement unit is arranged near the vehicle center of gravity to collect yaw rate and lateral acceleration data; vehicle speed data is collected using ABS speed sensor; for the vehicle center of gravity sideslip angle, based on the yaw rate and lateral acceleration collected by the MEMS inertial measurement unit, the vehicle speed collected by ABS and the front wheel steering angle, combined with the preset vehicle two-degree-of-freedom dynamic model, a state space equation is constructed; the front wheel steering angle is used as the input of the extended Kalman filter (EKF) algorithm, and the yaw rate and lateral acceleration are used as observation corrections. Based on the state space equation, the recursive estimation of state prediction and observation correction is performed, and the estimated vehicle center of gravity sideslip angle data is output in real time; the sampling rate of the above dynamics state data is uniformly set to 500Hz.

[0020] The two-degree-of-freedom dynamics model of the vehicle and the process of predicting data using the extended Kalman filter algorithm are well-known technologies and will not be elaborated further.

[0021] Step S2: By analyzing the changes in yaw rate with steering wheel angle, front wheel angle and vehicle speed, the time-domain relative change and high-frequency characteristics of the current in the steering motor circuit, an operational anomaly degree is constructed. Combined with the phase difference between the steering wheel angle and the front wheel angle, and the time-series change of the vehicle's center of gravity side slip angle change rate, the operational anomaly coefficient of the vehicle at each moment is determined.

[0022] The stable operation of a vehicle's steer-by-wire system depends on the stable coupling between steering wheel input, actuator response, and vehicle dynamics. Existing fault-tolerant controls largely rely on hardware redundancy or preset models, lacking the ability to predict complex dynamic environments and fault evolution trends. When a vehicle's steer-by-wire system malfunctions, degrades, or encounters environmental disturbances, the aforementioned coupling relationship will be disrupted. Therefore, the core of fault-tolerant control lies in accurately quantifying the system's health status and future stability trends through characteristic changes in multi-source data.

[0023] Specifically, when the vehicle's steer-by-wire system is functioning normally: a stable dynamic mapping is maintained between the steering wheel angle, front wheel angle, and yaw rate; the front wheels' following delay to the steering wheel is fixed and highly synchronized; the motor current is stable; and the sideslip angle change converges. However, when degradation occurs, such as motor demagnetization or bearing wear, the motor requires a larger current to maintain output, leading to an increase in the average current and high-frequency ripple. At the same time, the steering response delay increases, and the phase difference between the steering wheel and the front wheels widens. When road surface adhesion decreases or tires sideslip, the sideslip stiffness drops sharply, and the sideslip angle change rate increases dramatically, accompanied by oscillations.

[0024] Based on the aforementioned coupling failure mechanism, this embodiment analyzes the changes in yaw rate with steering wheel angle, front wheel angle, and vehicle speed, as well as the time-domain relative changes and high-frequency characteristics of the current in the steering motor circuit, to construct an operational anomaly degree. Furthermore, by combining the phase difference between the steering wheel angle and the front wheel angle, and the temporal variation of the vehicle's center of gravity's sideslip angle rate of change, the operational anomaly coefficient of the vehicle at each moment is determined. The specific process is as follows: By fitting the steering wheel angle, front wheel angle, vehicle speed, and yaw rate at each moment and within a preset historical time period, the fitting error of the yaw rate at each moment is obtained. Specifically: In this embodiment, the steering wheel angle, front wheel angle and vehicle speed at each moment and in the preset historical period before are used as independent variables in the multiple linear regression fitting algorithm, and the yaw rate is used as the dependent variable in the multiple linear regression fitting algorithm to obtain the fitting curve. The absolute difference between the yaw rate at each time point and the corresponding fitted value on the fitted curve is taken as the fitting error of the yaw rate at each time point.

[0025] It should be noted that the preset historical time period is set manually. In this embodiment, the preset historical time period is 5 seconds long. This is to construct a statistical sample set that can fully reflect the dynamic state of the vehicle while ensuring real-time calculation. The 5-second time span covers the complete response cycle of the vehicle from steady-state driving to typical dynamic operations (such as single avoidance, lane changing, or lane changing), so that the model obtained by fitting a multivariate function based on the data of this time period can stably capture the coupling relationship between steering wheel input, front wheel response, and yaw output. If the time period is too short, insufficient sample points will make the fitted model susceptible to high-frequency noise interference, and it will be unable to accurately extract the baseline mapping features, thereby increasing the false alarm rate of the fitting error. Conversely, if the time period is too long, it will contain too much outdated historical information, causing the model to lag behind the current instantaneous state changes of the vehicle and reducing the sensitivity to rapidly evolving faults. In addition, the length of the preset historical time period is not limited to 5 seconds. Implementers can flexibly adjust it according to the specific control response time constant of the vehicle and the chip's computing power. This embodiment does not impose any special restrictions.

[0026] The process of fitting data using a multiple linear regression fitting algorithm is a well-known technique and will not be elaborated further.

[0027] Furthermore, in this embodiment, the rate of change of the average current in the steering motor circuit within the preset window before each time point compared to the average current within the preset window before the previous time point is recorded as the time-domain relative change characteristic value of the current in the steering motor circuit at each time point; wherein, in this embodiment, the difference between the average current in the steering motor circuit within the preset window before each time point compared to the average current within the preset window before the previous time point is divided by the average current within the preset window before the previous time point, and this is taken as the rate of change of the average current in the steering motor circuit within the preset window before each time point compared to the average current within the preset window before the previous time point.

[0028] It should be noted that the preset window length is set manually. In this embodiment, the preset window length is set to 1 second. The 1-second time window contains approximately 500 sampling points, which can meet the basic requirements of Fourier transform for frequency resolution, and can also filter out random white noise and instantaneous jitter in the sensor acquisition process through averaging, thereby accurately extracting characteristic parameters reflecting the health status of the system, such as the mean current, frequency centroid, and phase difference. If the window is too short, the frequency domain analysis resolution is insufficient, and it is easily affected by a single noise point, causing the characteristic values ​​to jump drastically. If the window is too long, it will be unable to capture the rapid time-varying characteristics of the vehicle's status in time, resulting in feature lag. The preset window length can also be adjusted according to actual needs, and this embodiment does not impose any special restrictions.

[0029] Furthermore, in this embodiment, the weighted sum of the current in the steering motor circuit within the preset window before each time point, representing the current at each frequency and its energy percentage in the frequency domain, divided by the maximum frequency, is recorded as the frequency domain characteristic value at each time point. For ease of understanding, a specific calculation expression is given below: Frequency domain eigenvalues ​​at time i The expression is: In the formula, , These represent the j-th frequency value and the energy percentage at the j-th frequency in the energy spectrum obtained by applying the Fourier transform algorithm to the current in the steering motor circuit within the preset window before time i. In this embodiment, the energy percentage is calculated by the ratio of the energy value at the j-th frequency to the maximum value among all energy values ​​at all frequencies. This represents the maximum frequency in the energy spectrum obtained by applying the Fourier transform algorithm to the current in the steering motor circuit within the preset window before time i; This indicates the number of all times within the preset window prior to time i.

[0030] The process of obtaining the energy spectrum using the Fourier transform algorithm is a well-known technique and will not be described in detail here.

[0031] Furthermore, this embodiment constructs the vehicle's operational anomaly degree at each time point based on the fitting error, the time-domain relative change feature value, and the frequency-domain feature value. Specifically: In this embodiment, the abnormality of the car's operation at time i is... The expression is: In the formula, The fitting error of the yaw rate at time i is represented. This represents the maximum value of the yaw rate fitting error at time i and all times within the preset historical time period prior to time i. The time-domain relative change characteristic value of the current in the steering motor circuit at time i represents the value of the change in the current at time i. This represents the frequency domain eigenvalue of the car at time i; This indicates a preset constant greater than 0, used to prevent the denominator from being 0. The value is set manually, in this embodiment. The value of is 0.01. Under the premise of ensuring that the denominator is not 0 and does not excessively affect the calculation result, the implementer can also set it according to the specific situation. This embodiment does not impose special restrictions; max() represents the maximum value function.

[0032] Based on the operational anomaly degree, it can be understood that the operational anomaly degree is used to characterize the overall health level of the automotive steer-by-wire system in terms of mechanical structural integrity, motor drive efficiency, and dynamic coupling stability. The larger the operational anomaly degree, the larger the fitting error, the faster the mean current increases, and the more severe the high-frequency ripple. This directly points to hardware degradation problems such as decreased performance of the automotive steer-by-wire actuator, wear of mechanical components, or reduced energy transfer efficiency. It indicates that the stability margin of the automotive steer-by-wire system is low and requires monitoring attention. Conversely, the smaller the operational anomaly degree, the more closely the operating parameters of the automotive steer-by-wire system fit the expected ideal mapping relationship. The current is stable and there is no high-frequency harmonic interference, reflecting that the automotive steer-by-wire system is in a healthy and efficient operating state, with all components functioning normally and no obvious performance degradation or abnormal wear.

[0033] Furthermore, this embodiment determines the vehicle's operating condition anomaly coefficient at each moment based on the degree of operational anomaly, combined with the phase difference between the steering wheel angle and the front wheel angle, and the temporal change of the vehicle's center of gravity's sideslip angle rate of change. Specifically: In this embodiment, the phase difference between the steering wheel angle and the front wheel angle is calculated, specifically: Extract the dominant frequencies of the steering wheel angle and front wheel angle in the frequency domain within the preset window before each time step; where the dominant frequency of the steering wheel angle is the frequency corresponding to the maximum energy amplitude of the steering wheel angle in the frequency domain, and the dominant frequency of the front wheel angle is the frequency corresponding to the maximum energy amplitude of the front wheel angle in the frequency domain. If the main frequency of the steering wheel angle and the front wheel angle is the same, then the phase difference at that main frequency is calculated as the phase difference between the steering wheel angle and the front wheel angle at each moment. Otherwise, the phase difference between the steering wheel angle and the front wheel angle at each moment is set to 0.

[0034] Furthermore, the temporal variation of the rate of change of the vehicle's center of gravity's sideslip angle is quantified, specifically as follows: In this embodiment, the absolute difference between the vehicle's center of gravity sideslip angle at each time point and the adjacent previous time point is divided by the corresponding time interval and recorded as the rate of change of the vehicle's center of gravity sideslip angle at each time point. The ratio of the rate of change of the vehicle's center of gravity sideslip angle at each time point to the average rate of change of the vehicle's center of gravity sideslip angle at all times within the preset window prior to each time point is used as the quantitative result of the temporal change of the rate of change of the vehicle's center of gravity sideslip angle at each time point.

[0035] Furthermore, this embodiment determines the vehicle's operating condition anomaly coefficient at each moment based on the quantitative results of the time-series changes in the operational anomaly degree, the phase difference between the steering wheel angle and the front wheel angle, and the change rate of the vehicle's center of gravity sideslip angle. Specifically: In this embodiment, the abnormality coefficient of the vehicle at each time point is positively correlated with the quantitative results of the temporal changes of the abnormality degree of operation, the phase difference between the steering wheel angle and the front wheel angle, and the rate of change of the vehicle's center of gravity side slip angle.

[0036] It should be understood that a positive correlation means that the dependent variable increases as the independent variable increases, and the dependent variable decreases as the independent variable decreases. The specific relationship can be additive or multiplicative, etc., and is determined by the actual application. This application does not impose any special restrictions.

[0037] Preferably, as one implementation method, in this embodiment, the abnormal operating condition coefficient of the vehicle at time i is... The expression is: In the formula, This represents the degree of abnormality in the vehicle's operation at time i; This represents the normalized value of the phase difference between the steering wheel angle and the front wheel angle at time i. This represents the quantitative result of the time-series change of the rate of change of the vehicle's center of gravity's slip angle at time i.

[0038] It should be noted that in this embodiment, the minimum-maximum normalization method is used to normalize the phase difference between the steering wheel angle and the front wheel angle. Specifically, the maximum and minimum values ​​of the phase difference between the steering wheel angle and the front wheel angle at each moment and at all moments within the preset historical time period are used as the maximum and minimum value parameters in the minimum-maximum normalization method. In actual applications, as other implementation methods, implementers may also use other normalization methods according to specific circumstances. This embodiment does not impose any special restrictions on the selection of normalization methods.

[0039] The process of normalizing data using the minimum-maximum normalization method is a well-known technique and will not be elaborated further.

[0040] Based on the operating condition anomaly coefficient, it can be understood that the operating condition anomaly coefficient is used to comprehensively measure the degree of overall operating condition anomaly of a vehicle under specific environmental and operational conditions. It reflects the comprehensive risk of the coupling relationship of the vehicle's steer-by-wire system being disrupted. The larger the operating condition anomaly coefficient, the worse the stability of the vehicle's steer-by-wire system, the longer the response time, and the more drastic the lateral deviation. This usually means that the vehicle's steer-by-wire system has encountered a serious hardware failure, sudden failure, or extreme road instability, which is very likely to cause the vehicle to lose control. In this case, fault-tolerant control needs to intervene immediately for strong intervention. Conversely, the smaller the operating condition anomaly coefficient, the smaller the vehicle's operating parameters are in accordance with physical laws, the response is rapid and the lateral deviation converges, reflecting that the vehicle is in a safe and normal driving state or the environmental disturbances are within a controllable range, and no additional fault-tolerant intervention is required.

[0041] Thus, this embodiment constructs an anomaly coefficient by integrating multi-source heterogeneous data such as dynamic mapping error, motor current time-frequency characteristics, steering input-output phase difference, and dynamic changes in sideslip angle. This solves the problem that existing fault-tolerant control relies only on single or local state perception, making it difficult to accurately quantify the degree of damage to system coupling relationships. It achieves a comprehensive and unified quantitative assessment of multi-dimensional anomalies such as hardware degradation, execution failure, and road disturbance, providing a reliable risk quantification basis for accurately decoupling fault evolution trends and formulating forward-looking fault-tolerant control strategies.

[0042] Step S3: Use a Gaussian process regression model to decouple the slow trend of hardware degradation, obtain the benchmark operating condition anomaly coefficient of the vehicle at each time point, and construct the mutation factor at each time point based on the difference between the operating condition anomaly coefficient and the benchmark operating condition anomaly coefficient.

[0043] While the operating condition anomaly coefficient can effectively quantify the degree of anomaly in the current system, it cannot distinguish the source attribute of the anomaly. In automotive steer-by-wire systems, anomalies are mainly caused by two types of coupling factors: first, hardware degradation (such as motor demagnetization and bearing wear), whose evolution process has significant long-term slow-changing characteristics; and second, environmental disturbances or sudden failures, which manifest as rapid changes in a short period of time. Since hardware degradation is an irreversible physical evolution, fault-tolerant control cannot reverse this trend and can only maintain basic performance through system compensation; while environmental abrupt changes and the risk of rapid instability are the core objectives that fault-tolerant control urgently needs to intervene in and suppress. If these two are not decoupled, the slow-changing degradation will accumulate over a long period of time, leading to a persistently high operating condition anomaly coefficient, which can easily mask the signal of a sudden and rapid increase in anomalies, causing misjudgment or lag in fault-tolerant control. Therefore, it is necessary to remove the slow-changing baseline of hardware degradation from the comprehensive operating condition anomaly coefficient and accurately extract the abrupt component that reflects the actual control requirements.

[0044] Based on the above decoupling requirements, this embodiment utilizes a Gaussian process regression model to decouple the slow-changing trend of hardware degradation, obtains the baseline operating condition anomaly coefficient of the vehicle at each time point, and constructs the mutation factor at each time point based on the difference between the operating condition anomaly coefficient and the baseline operating condition anomaly coefficient. The specific process is as follows: First, sample data is collected at different times, with the sample size set to M. In this embodiment, M is 50,000. Each sample represents data from a car with a healthy steer-by-wire system under different road conditions and operations, including road surface material data, ambient temperature and humidity, rainfall, steering wheel angular velocity, and vehicle speed. The road surface material data can be processed using one-hot encoding, for example, [0,1] represents asphalt and [1,0] represents cement. Then, 2 / 3 of the samples are randomly selected to form the training set. ,in This is the sample data of the i-th sample. It is the first The abnormal coefficients of the working conditions correspond to each sample at a given time. However, due to the randomness of the selection of the training set, the samples in the training set may be discontinuous. Therefore, for the sake of convenience, the abnormal coefficients of the working conditions are... The subscripts in the table represent the sample order, while the calculation method remains unchanged. Finally, the training set is used as the input to the Gaussian process regression (GPR) model, and the operational anomaly coefficients at corresponding time points for each sample in the training set are used as the output of the GPR model. The GPR model is then trained using a Matern 5 / 2 kernel function, and the length scale is automatically optimized through maximum likelihood estimation. The noise variance is set to... The L-BFGS-B optimizer was used for hyperparameter optimization, and the trained Gaussian process regression model was finally obtained. Subsequently, the road surface material data, ambient temperature and humidity, rainfall, steering wheel angular velocity and vehicle speed data at each time point were input into the trained Gaussian process regression model, and the output working condition anomaly coefficient was used as the baseline working condition anomaly coefficient.

[0045] It should be noted that the total sample size M is 50,000 in this embodiment, which is based on the balance between the Gaussian process regression model's ability to learn from small samples and the coverage of data distribution. Although the Gaussian process regression model theoretically has the advantage of handling small samples, in this application scenario, it is necessary to fit a multi-dimensional (road conditions, temperature and humidity, vehicle speed, steering wheel angular velocity, etc.) nonlinear mapping space. Too few samples (such as a few thousand) will cause the model to be unable to cover discrete combinations of working conditions, resulting in the predicted baseline working condition anomaly coefficients oscillating violently or reverting to the global mean under certain working conditions, thus losing the ability to characterize environmental disturbances. 50,000 samples can ensure sufficient point density in the multi-dimensional feature space, enabling the Gaussian process regression model to accurately capture the potential local nonlinear relationship between environmental parameters and working condition anomaly coefficients. At the same time, considering that the computational complexity of the Gaussian process regression model increases cubically with the sample size, 50,000 is also the upper limit of effective data that can be used while ensuring the real-time performance of the model inference (such as meeting the computation cycle requirements of the vehicle controller).

[0046] The training process of the Gaussian process regression model is a well-known technique and will not be described in detail here.

[0047] Furthermore, this embodiment constructs a mutation factor at each time point based on the difference between the abnormal coefficient of the operating condition and the abnormal coefficient of the baseline operating condition. Specifically: In this embodiment, an abnormal operating condition deviation is constructed based on the difference between the abnormal operating condition coefficient of the vehicle at each time point and the abnormal operating condition coefficient of the baseline operating condition. Specifically: In this embodiment, the abnormal deviation of the operating condition at time i The expression is: In the formula, , represents the abnormal coefficient of the vehicle's operating condition and the abnormal coefficient of the baseline operating condition at time t, respectively; max() represents the maximum value function.

[0048] Furthermore, the exponential decay function value of the difference between the abnormal deviation of the operating condition at each time point and its adjacent previous time point is calculated and denoted as the operating condition abrupt change decay value. For ease of understanding, its specific calculation formula is as follows: Operating condition abrupt change at time i The expression is: In the formula, , These represent the abnormal deviations in operating conditions at time i and time i-1, respectively. represents the preset adjustment factor; exp[ ] represents the exponential function with the natural constant as the base.

[0049] Furthermore, this embodiment constructs a mutation factor at each time point based on the abnormal deviation of the operating condition and the abrupt decay value of the operating condition, specifically: In this embodiment, the mutation factor at time i The expression is: In the formula, This indicates an abnormal deviation in the operating conditions at time i; This represents the attenuation value of the sudden change in operating conditions at time i; represents the preset boundary factor; max() represents the maximum value function.

[0050] It should be noted that a boundary factor is preset in this embodiment. The value of τ is set to 0.1 to prevent the calculated mutation factor from being too small or even zero under certain extreme normal operating conditions, which would cause the corresponding sample to be completely ignored during model training. Setting τ to 0.1 is equivalent to reserving a minimum "participation qualification" for each sample, ensuring that the model can maintain continuous learning of basic data features even when the system is completely stable, preventing the model from "forgetting" or becoming biased. Of course, the value of τ is not fixed and can be adjusted according to the size and noise level of the training data. For example, for datasets with high noise, τ can be appropriately increased to 0.2 to retain more samples participating in gradient updates, while for scenarios with extremely high data quality, it can be reduced to 0.05 to focus more on high-weight samples.

[0051] Preferably, the schematic diagram of the mutation factor extraction process provided in this embodiment is as follows: Figure 2 As shown.

[0052] Based on the mutation factor, it can be understood that the mutation factor is a dimensionless weighting parameter used to characterize the importance of data samples in revealing the risks of sudden failures and rapid instability in automotive steer-by-wire systems. The mutation factor reflects the degree of deviation of the abnormal operating condition from the benchmark and the severity of its evolution over time. The larger the mutation factor, the more the actual abnormality exceeds the benchmark and the more drastic the changes are between adjacent time points. This directly points to the sudden deterioration or sudden failure of the hardware, which is a sensitive signal that fault-tolerant control needs to pay attention to. Therefore, the corresponding samples should occupy a very high weight in training to enhance the predictive sensitivity of the model. Conversely, the smaller the mutation factor, the more the abnormality mainly comes from slow degradation or environmental disturbances, or the automotive steer-by-wire system is in a convergent and stable state. Such samples have a low contribution to predicting the risk of sudden changes, so they are given a lower weight in training to avoid interfering with the model's learning of key mutation features.

[0053] Thus, this embodiment has successfully decoupled the slow-changing baseline of hardware degradation from the comprehensive anomaly by fitting the operating condition anomaly benchmark under multi-dimensional environment and operating parameters through a Gaussian process regression model. Based on the deviation between the actual anomaly and the benchmark and the degree of evolution, a mutation factor was constructed, which solved the problem of slow degradation accumulation masking sudden instability signals. It also achieved accurate weighting of samples sensitive to mutation risk, laying the foundation for subsequent model reinforcement learning of rapid evolution characteristics and improving the foresight and accuracy of predictive control.

[0054] Step S4: Dynamically adjust the attention calculation weights of samples in the pre-trained Transformer prediction model using mutation factors to predict the future operating condition anomaly coefficient and future risk state of the vehicle. Based on this, construct the gain adjustment factor and stability risk coefficient to perform dynamic fault-tolerant control on the vehicle's steer-by-wire actuator.

[0055] Since the original data has a high sampling frequency, directly using it for training would lead to a surge in computational complexity and information redundancy. Therefore, the temporal features within a preset window before the corresponding time of each continuous running sample are extracted. The temporal features include the phase difference between the steering wheel angle and the front wheel angle corresponding to each continuous running sample, the high-frequency characteristic value of the current in the steering motor circuit, the temporal change quantification result of the side slip angle change rate of the vehicle's center of gravity, and the ambient temperature, humidity and rainfall at the corresponding time of each continuous running sample. The above temporal features constitute the feature vector of each continuous running sample.

[0056] Subsequently, a predetermined number of consecutive running samples were divided into a training set and a test set in a 7:3 ratio. The feature vectors of all consecutive running samples in the training set were used as the input to the Transformer model, and the abnormality coefficient, yaw rate, and sideslip angle of the vehicle's center of gravity at the corresponding time point of the consecutive running samples were used as the output. The encoding dimension was set to 128, the number of multi-head attention heads was set to 8, the number of encoding layers was set to 4, and the learning rate was set to... Finally, a trained Transformer model is obtained.

[0057] It should be noted that the preset number of data points in this embodiment is 50,000, which is an empirical lower limit determined based on the requirements of deep learning models for data richness and coverage. The automotive steer-by-wire system faces extremely complex combinations of operating conditions (permutations of different vehicle speeds, road surfaces, temperatures, and operating habits). 50,000 sample data points can cover most typical driving scenarios and edge conditions with a high probability, ensuring that the Transformer model can learn a complete mapping relationship between the environment, operation, and risk state, and avoiding overfitting or poor generalization due to insufficient data. Of course, this number is only a baseline value, and it is entirely possible and generally recommended to expand it based on the accumulation of actual test mileage. This embodiment does not impose any special restrictions.

[0058] The training process of the Transformer model is a well-known technique and will not be described in detail here.

[0059] Furthermore, the temporal features of the vehicle's operation within a preset window before each time point are extracted; the temporal features corresponding to each time point are used as input to a pre-trained Transformer prediction model, wherein the mutation factor at each time point is used as the attention weight of the corresponding temporal feature in the pre-trained Transformer prediction model, and the output operating condition anomaly coefficient, yaw rate, and sideslip angle of the vehicle's center of gravity are used as the future operating condition anomaly coefficient and future risk state at each time point, wherein the future risk state includes the future yaw rate and the future sideslip angle of the vehicle's center of gravity.

[0060] Furthermore, based on the future operating condition anomaly coefficient and future risk status, a gain adjustment factor and a stability risk coefficient are constructed, specifically: In this embodiment, a gain adjustment factor is constructed based on the degree of deviation between the future operating condition anomaly coefficient and the operating condition anomaly coefficient at each time point, specifically as follows: Gain adjustment factor at time i The expression is: In the formula, , Let represent the future operating condition anomaly coefficient and the operating condition anomaly coefficient at time i, respectively.

[0061] Furthermore, based on the difference between the future yaw rate and the maximum permissible yaw rate at each moment, and the difference between the future vehicle center of gravity sideslip angle and the maximum permissible sideslip angle, a stability risk coefficient is constructed, specifically: Stability risk coefficient at time i The expression is: In the formula, , These represent the future yaw rate and the future sideslip angle of the vehicle's center of gravity at the given time, respectively. , These represent the maximum permissible yaw rate and the maximum permissible sideslip angle of a car's steer-by-wire system, respectively.

[0062] It should be noted that in this embodiment... , The values ​​are 0.5 and 0.5 respectively. In practical applications, as other implementation methods, implementers can also set them according to specific circumstances. This embodiment does not impose any special restrictions.

[0063] Based on the gain adjustment factor and stability risk coefficient, we can understand that a smaller gain adjustment factor indicates a more conservative control strategy, which reduces the gain to avoid excessive action of the vehicle's steer-by-wire actuator under risky conditions. The stability risk coefficient is used to quantify the probability of the vehicle becoming unstable in the future. A larger stability risk coefficient means that the yaw rate and sideslip angle are closer to the physical limits, and the risk of instability is extremely high. The combination of the two, from the two dimensions of "the magnitude of control action" and "the safety of vehicle status", together constitutes the decision basis for the fault-tolerant control strategy, ensuring that the vehicle can smoothly reduce its response sensitivity when facing potential risks, while retaining sufficient handling margin when on the verge of losing control.

[0064] Furthermore, this embodiment optimizes the torque of the vehicle's steer-by-wire actuator in real time by utilizing a gain adjustment factor and a stability risk coefficient, and uses the optimized torque to perform fault-tolerant control on the vehicle's steer-by-wire actuator.

[0065] It should be noted that in this embodiment, the optimized torque of the vehicle's steer-by-wire actuator at time i is... The expression is: In the formula, , This represents the gain adjustment factor and stability risk coefficient at time i; This indicates the driver's initial steering torque requirement; This indicates the preset parameter tuning factor.

[0066] In this embodiment, the preset parameter adjustment factor is set to 1.3 to ensure that the coefficient terms are within the acceptable range when constructing the final torque control law. To ensure the value remains consistently positive, and to avoid the dangerous situation of the control gain reversing (i.e., becoming negative) due to fluctuations in the stability risk coefficient or calculation errors, a value of 1.3 was chosen as the numerical basis, based on the stability risk coefficient. The theoretical extreme value is 1, so α needs to be greater than 1. Leaving a redundancy margin of 0.3 can effectively resist numerical disturbances and sensor measurement noise during the calculation process, and ensure the smoothness and directional correctness of the control command. Of course, the value of α can be fine-tuned according to the specific safety margin requirements of the car. For example, for cars with higher safety requirements, it can be increased to 1.5 to obtain a larger safety factor, while for scenarios with higher responsiveness requirements, it can be reduced to 1.1 to improve control sensitivity. This embodiment does not impose any special restrictions.

[0067] Thus, this embodiment enhances the learning and prediction of key characteristics of sudden failures through a dynamically weighted Transformer model with mutation factors. Based on the predicted future operating condition anomaly coefficient and risk state, it constructs gain adjustment factors and stability risk coefficients, realizing the early perception of future instability trends and adaptive closed-loop adjustment of execution torque. This solves the problems of intervention lag and over-acting caused by the lack of foresight in traditional fault-tolerant control, and improves the prediction accuracy of the vehicle's steer-by-wire system for unknown risks and the control robustness under dynamic disturbances.

Claims

1. A fault-tolerant control method for steer-by-wire in automobiles that integrates embodied large model prediction technology, characterized in that, The method includes the following steps: The system acquires the steering wheel angle, front wheel angle, vehicle center of gravity sideslip angle, vehicle speed, yaw rate, and current in the steering motor circuit during vehicle operation. By analyzing the changes in yaw rate with steering wheel angle, front wheel angle and vehicle speed, the time-domain relative change and high-frequency characteristics of current in the steering motor circuit, an operational anomaly degree is constructed. Combined with the phase difference between steering wheel angle and front wheel angle, and the time-series change of the vehicle's center of gravity side slip angle change rate, the operational anomaly coefficient of the vehicle at each time point is determined. Gaussian process regression model is used to decouple the slow trend of hardware degradation, obtain the benchmark operating condition anomaly coefficient of the vehicle at each time point, and construct the mutation factor at each time point based on the difference between the operating condition anomaly coefficient and the benchmark operating condition anomaly coefficient. By dynamically adjusting the attention calculation weights of samples in the pre-trained Transformer prediction model using mutation factors, the future operating condition anomaly coefficient and future risk state of the vehicle can be predicted. Based on this, a gain adjustment factor and a stability risk coefficient are constructed to perform dynamic fault-tolerant control on the vehicle's steer-by-wire actuator. The process of constructing the operational anomaly degree is as follows: By fitting the steering wheel angle, front wheel angle, vehicle speed and yaw rate at each time point and in the preset historical time period before that time point, the fitting error of the yaw rate at each time point is obtained. The rate of change of the average current in the steering motor circuit within the preset window before each time point compared to the average current within the preset window before the previous time point is denoted as the time-domain relative change characteristic value of the current in the steering motor circuit at each time point. The result of dividing the weighted sum of the current and its energy proportion at each frequency in the steering motor circuit within the preset window before each time point by the maximum frequency is recorded as the frequency domain characteristic value at each time point. Based on the fitting error, the time-domain relative change feature value, and the frequency-domain feature value, the vehicle's operational anomaly degree at each time point is constructed. The abnormal operating condition coefficients of the vehicle at each time point are positively correlated with the time-series changes of the abnormal operating degree, the phase difference between the steering wheel angle and the front wheel angle, and the rate of change of the vehicle's center of gravity side slip angle. The process for determining the phase difference between the steering wheel angle and the front wheel angle is as follows: Extract the dominant frequencies of the steering wheel angle and front wheel angle in the frequency domain within the preset window before each time point; If the main frequency of the steering wheel angle and the front wheel angle is the same, then the phase difference at that main frequency is calculated as the phase difference between the steering wheel angle and the front wheel angle at each moment. Otherwise, the phase difference between the steering wheel angle and the front wheel angle at each moment is set to 0.

2. The vehicle steer-by-wire fault-tolerant control method integrating embodied large model prediction technology as described in claim 1, characterized in that, The quantification process of the time-series variation of the rate of change of the vehicle's center of gravity slip angle is as follows: The absolute difference between the vehicle's center of gravity sideslip angle at each time point and the adjacent previous time point is divided by the corresponding time interval and recorded as the rate of change of the vehicle's center of gravity sideslip angle at each time point. The ratio of the rate of change of the vehicle's center of gravity sideslip angle at each time point to the average rate of change of the vehicle's center of gravity sideslip angle at all times within the preset window prior to each time point is used as the quantitative result of the temporal change of the rate of change of the vehicle's center of gravity sideslip angle at each time point.

3. The vehicle steer-by-wire fault-tolerant control method integrating embodied large model prediction technology as described in claim 1, characterized in that, The construction process of the mutation factors at each time point is as follows: Based on the difference between the abnormal coefficient of the vehicle's operating condition and the abnormal coefficient of the benchmark operating condition at each time point, an abnormal operating condition deviation is constructed. Calculate the exponential decay function value of the difference between the abnormal deviation of the operating condition at each time point and the adjacent previous time point, and denot it as the operating condition sudden change decay value; Based on the abnormal deviation of the operating condition and the abrupt decrease value of the operating condition, a mutation factor is constructed at each time point.

4. The vehicle steer-by-wire fault-tolerant control method integrating embodied large model prediction technology as described in claim 1, characterized in that, The method of dynamically adjusting the attention calculation weights of samples in the pre-trained Transformer prediction model using mutation factors to predict the future operating condition anomaly coefficient and future risk status of the vehicle includes: Extract the temporal features of the car's operation within a preset window before each time point; The temporal features corresponding to each time step are used as input to the pre-trained Transformer prediction model. The mutation factor at each time step is used as the attention weight of the corresponding temporal feature in the pre-trained Transformer prediction model. The future operating condition anomaly coefficient and future risk state at each time step are output. The future risk state includes the future yaw rate and the future vehicle center of gravity sideslip angle.

5. The steer-by-wire fault-tolerant control method for automobiles that integrates embodied large model prediction technology as described in claim 4, characterized in that, The constructed gain adjustment factor and stability risk coefficient include: Based on the degree of deviation between the future operating condition anomaly coefficient and the operating condition anomaly coefficient at each time point, a gain adjustment factor is constructed; A stability risk coefficient is constructed based on the difference between the future yaw rate and the maximum permissible yaw rate at each moment, as well as the difference between the future vehicle center of gravity sideslip angle and the maximum permissible sideslip angle.

6. The vehicle steer-by-wire fault-tolerant control method integrating embodied large model prediction technology as described in claim 1, characterized in that, The dynamic fault-tolerant control of the vehicle's steer-by-wire actuator includes: By using the gain adjustment factor and the stability risk coefficient, the torque of the automotive steer-by-wire actuator is optimized in real time, and the optimized torque is used to perform fault-tolerant control on the automotive steer-by-wire actuator.

7. The steer-by-wire fault-tolerant control method for automobiles that integrates embodied large model prediction technology as described in claim 6, characterized in that, The real-time optimization of the torque of the automotive steer-by-wire actuator includes: Optimized torque of the automotive steer-by-wire actuator at time i The expression is: In the formula, , This represents the gain adjustment factor and stability risk coefficient at time i; This indicates the driver's initial steering torque requirement; This indicates the preset parameter tuning factor.

Images