Fault-tolerant control method for four-wheel steering vehicle

Abstract

The invention relates to the technical field of automobile control, and discloses a four-wheel steering vehicle fault-tolerant control method, which comprises the following steps: constructing a vehicle two-degree-of-freedom dynamic model; real-time data in the vehicle running process are collected, and a vehicle state estimation value is obtained through an unscented Kalman algorithm; constructing a residual reference model to obtain a residual sequence, and determining a fault sensor by using the residual sequence; and the vehicle state estimation value is used as the input of a reinforcement learning model-free control algorithm, and tire rotation angle control in the vehicle operation process is adjusted in real time. According to the method, the vehicle kinetic equation is used as a network regular constraint term, the accuracy and reliability of residual calculation are remarkably improved, rapid and accurate fault positioning is achieved through matching analysis of the residual sequence and the preset fault vector table, meanwhile, rapid self-adaption and control law optimization of a dynamic environment are achieved through reinforcement learning without a model, and the method is suitable for large-scale popularization and application. And the fault-tolerant performance and the control stability of the system are improved.

Description

Technical Field

[0001] This invention relates to the technical field of automotive control, and more specifically, to a fault-tolerant control method for four-wheel steering vehicles. Background Technology

[0002] With the continuous development of automotive electronics technology and the integration of automotive systems, four-wheel steering based on steer-by-wire technology has become an important component of the vehicle's active chassis control. It eliminates the traditional mechanical steering system; steering signals are transmitted to the electronic control unit (ECU) via sensors. The ECU then analyzes and processes the signals before transmitting them to the steering motor. The steering motor then generates the required steering torque based on the control signals, thereby realizing the driver's steering intention. Four-wheel steer-by-wire vehicles, relying on flexible control strategies for all four wheels, can achieve a smaller turning radius and more flexible and diverse driving directions, enabling free steering and diagonal straight-line driving under extreme conditions, ensuring vehicle stability in extreme environments. However, because the steer-by-wire system introduces electronic components such as sensors, controllers, and steering motors, the probability of system failure also increases. Depending on the location of the failure, it can be categorized as sensor failure, ECU failure, communication bus failure, and actuator failure. In particular, when the sensors used to collect status information in the control loop experience gain variation failures, constant deviation failures, or jamming failures, and cannot correctly collect status information, the vehicle may completely lose its steering ability, easily leading to serious loss of vehicle control.

[0003] Existing vehicle fault-tolerant control technologies mainly focus on constructing basic two-degree-of-freedom vehicle dynamics models, using traditional algorithms for simple fault diagnosis, and designing model-based single fault-tolerant control strategies. However, with the increase in electronic and electrical components in steer-by-wire systems, the risk of failure increases significantly, and diagnostic strategies suffer from slow response and insufficient accuracy, making it difficult to meet functional safety requirements. Existing fault-tolerant control strategies have two main drawbacks: first, they rely on precise vehicle mathematical models, and the control effect significantly decreases when the model is simplified or the parameters change over time (e.g., changes in driving conditions or fault types); second, single control strategies (e.g., model control with fixed parameters) have poor adaptability to different faults (e.g., front wheel steering angle sensor failure, yaw rate sensor failure) and different driving scenarios (e.g., low-speed rural roads, high-speed highways), exhibiting problems such as response lag and insufficient control accuracy, making it difficult to guarantee the stability and safety of the vehicle under fault conditions.

[0004] For example, patent CN120135275A proposes a model-predictive control method that establishes an objective function to obtain the optimal wheel angle and control the steering wheels for steering. Patent CN120065761A proposes a robust performance-preserving, feedforward-coordinated fault-tolerant control that adjusts the control output based on the prediction of ideal input. Patent CN118226759A proposes a fault-tolerant control algorithm for four-wheel independent steer-by-wire actuators in automobiles, which enables the vehicle's stability performance to gradually track the ideal state after actuator failure. These technologies focus more on fault-tolerant control at the algorithmic level.

[0005] For example, patent CN120039306A proposes a fault-tolerant control method for a redundant steer-by-wire system based on dynamic torque distribution. This method calculates the torque curve of the redundant motor through torque relationships, enabling the steering system to switch torques when a fault occurs. Another example is patent CN117184219A, which proposes a fault-tolerant control device and method for the actuator of a steer-by-wire system. This method diagnoses the cause of a fault in the actuator motor and sends electrical parameter compensation data to the drive controller of the actuator motor based on the diagnosis results, thereby achieving fault-tolerant compensation for the actuator motor through the drive controller. These technologies focus more on controlling the vehicle's own motor.

[0006] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention

[0007] In view of the problems in the related technologies, the present invention proposes a fault-tolerant control method for four-wheel steering vehicles to overcome the above-mentioned technical problems existing in the existing related technologies.

[0008] Therefore, the specific technical solution adopted by the present invention is as follows:

[0009] A fault-tolerant control method for a four-wheel steering vehicle includes:

[0010] S1. Combining the lateral motion equilibrium equations and the transverse motion equilibrium equations, as well as the relationship between tire lateral force and tire slip angle, a two-degree-of-freedom dynamic model of the vehicle is constructed.

[0011] S2. Collect real-time data of the vehicle operation process, and use the unscented Kalman algorithm to construct a joint estimator using the real-time data to obtain the vehicle state estimate.

[0012] S3. Construct a residual baseline model, calculate the expected output value of real-time data, and use the pre-acquired actual output value and expected output value to calculate and obtain the residual sequence. Use the residual sequence to determine the faulty sensor.

[0013] S4. Based on the fault sensor, the model-free control algorithm is initialized using the vehicle's two-degree-of-freedom dynamic model. The DQN network algorithm is used to optimize the model-free control algorithm to obtain the reinforcement learning model-free control algorithm. The vehicle state estimate is used as the input of the reinforcement learning model-free control algorithm to adjust the tire steering angle control during the vehicle's operation in real time.

[0014] Preferably, the construction of a two-degree-of-freedom vehicle dynamics model by combining the lateral motion equilibrium equations and the transverse motion equilibrium equations, as well as the relationship between tire lateral force and tire slip angle, includes:

[0015] S11. Construct the vehicle's lateral motion equilibrium equations and the vehicle's transverse motion equilibrium equations;

[0016] S12. Based on the vehicle's lateral motion balance equation and the vehicle's transverse motion balance equation, when the vehicle tire is in the linear region, the vehicle tire's lateral force equation and the vehicle tire's slip angle equation are obtained.

[0017] S13. By combining the vehicle's lateral motion equilibrium equations, the vehicle's transverse motion equilibrium equations, the vehicle's tire lateral force equations, and the vehicle's tire slip angle equations, the vehicle's state-space equations are obtained.

[0018] Preferably, the process of collecting real-time data during vehicle operation and using the unscented Kalman algorithm to construct a joint estimator based on the real-time data to obtain vehicle state estimates includes:

[0019] S21. Collect real-time data on vehicle operation, including front wheel steering angle, rear wheel steering angle, center of gravity sideslip angle, yaw rate, and sensor fault data.

[0020] S22. Define the state vector and measurement vector of the unscented Kalman algorithm using the centroid sideslip angle and yaw rate, respectively.

[0021] S23. Discretize the vehicle state-space equations using the Euler discretization method to obtain the state transition model;

[0022] S24. Based on the sideslip angle and yaw rate, the vehicle state estimate is calculated using the unscented Kalman algorithm.

[0023] Preferably, the process of calculating the vehicle state estimate using the unscented Kalman algorithm based on the sideslip angle and yaw rate includes:

[0024] S241, the parameters of the initial centroid sideslip angle and yaw rate, and the scale parameter;

[0025] S242. Calculate the mean weight and covariance weight of the Sigma points based on the scale parameter, and generate several Sigma points based on the current state estimate and covariance.

[0026] S243. Using the state transition model, predict each generated Sigma point to obtain the predicted sampling point and predicted state at the next time step, and use the mean weight to perform a weighted summation of the predicted sampling points to obtain the state prediction value.

[0027] S244. Based on the actual measured values, calculate the measurement prediction covariance and cross-covariance, and use the measurement prediction covariance and cross-covariance to calculate the Kalman gain to obtain the vehicle state estimate.

[0028] Preferably, the step of constructing a residual benchmark model, calculating the expected output value of real-time data, and using the pre-acquired actual output value and expected output value to calculate and obtain a residual sequence, and using the residual sequence to determine the faulty sensor includes:

[0029] S31. Use the residual baseline model to process the data of the real-time vehicle operation process to obtain the desired state value;

[0030] S32. Based on the desired state value, construct the regularization term and error loss of the dynamic equation, use the regularization term and error loss of the dynamic equation to perform a weighted summation to obtain the total loss function, and use the total loss function to optimize the residual benchmark model to obtain the desired output value.

[0031] S34. Calculate the difference between the pre-acquired actual output and the expected output value to obtain residual terms. Aggregate the residual terms to obtain a residual sequence. Construct a fault vector table based on the residual sequence and analyze it in conjunction with the preset residual threshold and residual sequence to determine the faulty sensor.

[0032] Preferably, the process of constructing a regularization term and error loss based on the desired state value, weighting and summing the regularization term and error loss to obtain a total loss function, and then using the total loss function to optimize the residual baseline model to obtain the desired output value includes:

[0033] S321. Substitute the desired state value and the vehicle's basic parameters into the lateral motion balance equation and the transverse motion balance equation to obtain the lateral dynamic constraint residual and the transverse dynamic constraint residual. Then, weight the two residuals to obtain the regularization term of the dynamic equation.

[0034] S322. Subtract the desired state value from the actual data collected by the sensor to obtain the error loss. Sum the error loss with the regularization term of the dynamic equation to obtain the total loss function.

[0035] Preferably, the step of initializing the model-free control algorithm based on the fault sensor using the vehicle's two-degree-of-freedom dynamics model, optimizing the model-free control algorithm using the DQN network algorithm to obtain a reinforcement learning model-free control algorithm, and using the vehicle state estimate as the input to the reinforcement learning model-free control algorithm to adjust the tire steering angle control during vehicle operation in real time includes:

[0036] S41. Initialize the core parameters of the model-free control algorithm, set the input and output dimensions of the model-free control algorithm after initialization, and construct the discrete-time system description of the model-free control algorithm based on the two-degree-of-freedom dynamics model of the vehicle.

[0037] S42. Model the vehicle operating environment as a Markov decision process and define the state space, action space, and reward function.

[0038] S43. Construct a DQN network algorithm with an attention mechanism, and train and optimize the parameters of the algorithm without model control after initialization to obtain optimized parameters;

[0039] S44. Substitute the optimized parameters into the model-free control algorithm after initialization to obtain the reinforcement learning model-free control algorithm.

[0040] Preferably, the initialization of the core parameters of the model-free control algorithm, the setting of the input and output dimensions of the model-free control algorithm after initialization, and the construction of a discrete-time system description of the model-free control algorithm based on the two-degree-of-freedom vehicle dynamics model include:

[0041] S411, the core parameters of the model-free control algorithm are the initial pseudo-partial derivatives and control parameters;

[0042] S412, the input dimensions of the model-free control algorithm are the front wheel angle and the rear wheel angle, and the output dimensions of the model-free control algorithm are the center of gravity lateral angle and the yaw rate.

[0043] Preferably, the DQN network algorithm with attention mechanism is constructed, and the parameters of the initial model-free control algorithm are trained and optimized to obtain the optimized parameters, including:

[0044] S431. Initialize the weight parameters of the attention Q network and the target attention Q network, as well as the experience replay buffer;

[0045] S432. Based on the current state and the attention Q network, the control parameters of the model-free control algorithm are adjusted using preset execution actions;

[0046] S433. Input the adjusted control parameters into the model-free control algorithm and calculate the real-time reward based on the reward function;

[0047] S434. Record the new state, merge the current state, the executed action, the immediate reward, and the new state into an experience quadruple, and put the experience quadruple into the experience buffer.

[0048] S435. When the experience buffer reaches the preset threshold, randomly sample a batch of experience groups and calculate the target Q value;

[0049] S436. Calculate the mean square error between the predicted Q value and the target Q value of the attention Q network, use the mean square error as the loss function, and iteratively update the network parameters through the gradient descent algorithm.

[0050] S437. Synchronize the Q-network parameters that have reached the iteration threshold to the target attention Q-network;

[0051] S438. Repeat steps S432 to S437 until the attention Q network converges and the optimized parameters are obtained.

[0052] Preferably, the DQN network algorithm is used to optimize the model-free control algorithm, resulting in the control law formula for the reinforcement learning model-free control algorithm as follows:

[0053] ;

[0054] In the formula, This represents the control quantity at time k. This represents the control variable from the previous iteration. Let represent the transpose of the estimated c-th type basis function vector at time k. This represents the desired vehicle state at time k+1. This represents the vehicle state at time k. This represents the step size factor under reinforcement learning updates. This represents the regularization factor under reinforcement learning updates.

[0055] The beneficial effects of this invention are as follows: By using vehicle dynamics equations as network regularization constraints, this invention forces the network output to satisfy physical laws, thereby significantly improving the accuracy and reliability of residual calculation. This method can effectively identify various complex fault modes, including sensor jamming, signal gain changes, and zero drift. Through matching analysis of residual sequences with a preset fault vector table, it achieves rapid and accurate location of fault types and fault locations. Simultaneously, by combining reinforcement learning model-free control with experience playback and attention mechanisms, it achieves rapid adaptation to dynamic environments and optimization of control laws, improving the system's fault tolerance and control stability. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart of a fault-tolerant control method for a four-wheel steering vehicle according to an embodiment of the present invention;

[0058] Figure 2 This is a schematic diagram of a two-degree-of-freedom vehicle model in a fault-tolerant control method for a four-wheel steering vehicle according to an embodiment of the present invention;

[0059] Figure 3 This is a block diagram illustrating the principle of a reinforcement learning model-free control algorithm in a fault-tolerant control method for a four-wheel steering vehicle according to an embodiment of the present invention.

[0060] Figure 4 This is a schematic diagram of reinforcement learning in a fault-tolerant control method for a four-wheel steering vehicle according to an embodiment of the present invention. Detailed Implementation

[0061] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0062] According to an embodiment of the present invention, a fault-tolerant control method for a four-wheel steering vehicle is provided.

[0063] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, the fault-tolerant control method for a four-wheel steering vehicle according to an embodiment of the present invention includes:

[0064] S1. Combining the lateral motion equilibrium equations and the transverse motion equilibrium equations, as well as the relationship between tire lateral force and tire slip angle, a two-degree-of-freedom dynamic model of the vehicle is constructed.

[0065] As a preferred embodiment, the construction of a two-degree-of-freedom vehicle dynamics model by combining the lateral motion balance equations and the transverse motion balance equations, as well as the relationship between tire lateral force and tire slip angle, includes the following steps: constructing the vehicle lateral motion balance equations and the vehicle transverse motion balance equations; based on the vehicle lateral motion balance equations and the vehicle transverse motion balance equations, when the vehicle tire is in the linear region, obtaining the vehicle tire lateral force equation and the vehicle tire slip angle equation; and simultaneously solving the vehicle lateral motion balance equation, the vehicle transverse motion balance equation, the vehicle tire lateral force equation, and the vehicle tire slip angle equation to obtain the vehicle state space equation.

[0066] Specifically, such as Figure 2 As shown, when establishing a vehicle dynamics model, a higher degree of freedom leads to increased computational costs and longer computation time. To save computational costs and ensure computational accuracy, a two-degree-of-freedom vehicle dynamics model is established. For ease of study, the following assumptions are made during the modeling process:

[0067] (1) Assume that the pitch and roll angles of the vehicle are 0, and that the vehicle has no vertical motion along the z-axis.

[0068] (2) Assume the vehicle body and suspension are rigid;

[0069] (3) Ignore aerodynamic effects;

[0070] (4) Assume that the left and right front wheels and the left and right rear wheels have the same turning angle and speed during the vehicle's operation.

[0071] The expression for the vehicle's lateral motion equilibrium equation is:

[0072] ;

[0073] In the formula, The derivative representing the centroid sideslip angle. Indicates yaw rate. Indicates the overall vehicle mass, v x Indicates longitudinal vehicle speed, Indicates the lateral force on the front wheel. This indicates the lateral force on the rear wheel.

[0074] The expression for the vehicle's lateral motion equilibrium equation is:

[0075] ;

[0076] In the formula, I represents the moment of inertia of the vehicle about the Z-axis. The derivative of the yaw rate. This indicates the distance from the vehicle's center of gravity to the front axle. This indicates the distance from the vehicle's center of gravity to the rear axle.

[0077] When the vehicle tires are in the linear region, the lateral forces of the front and rear tires can be obtained, and their expressions are as follows:

[0078] ;

[0079] ;

[0080] In the formula, This indicates the lateral force of the vehicle's front tires. This indicates the lateral force of the vehicle's rear tires. This indicates the lateral stiffness of the front tires. This indicates the lateral stiffness of the rear tire. Indicates the front wheel slip angle. This indicates the rear wheel slip angle.

[0081] The expression for the front wheel slip angle is:

[0082] ;

[0083] The expression for the rear wheel slip angle is:

[0084] ;

[0085] In the formula, v x Indicates longitudinal vehicle speed, v y Indicates lateral vehicle speed, δ f Indicates the front wheel steering angle, δ r L represents the rear wheel steering angle. f L represents the distance from the vehicle's center of gravity to the front axle. r This indicates the distance from the vehicle's center of gravity to the rear axle.

[0086] Then, by solving all the above equations simultaneously, we get:

[0087] ;

[0088] ;

[0089] Transforming these two equations, we obtain the space state equation, which is expressed as:

[0090] ;

[0091] In the formula, Table state vector, Let A be the rate of change of state, B be the state matrix, U(t) be the input vector, y(t) be the output vector, and C be the output matrix.

[0092] The formula for the state matrix is:

[0093] ;

[0094] The formula for the input matrix is:

[0095] ;

[0096] The formula for the output matrix is:

[0097] .

[0098] S2. Collect real-time data of the vehicle operation process, and use the unscented Kalman algorithm to construct a joint estimator using the real-time data to obtain the vehicle state estimate.

[0099] As a preferred embodiment, the process of collecting real-time data during vehicle operation and constructing a joint estimator using the unscented Kalman algorithm to obtain vehicle state estimates includes the following steps: collecting real-time data during vehicle operation, including front wheel steering angle, rear wheel steering angle, center of gravity sideslip angle, yaw rate, and sensor fault data; defining the state vector and measurement vector for the unscented Kalman algorithm using the center of gravity sideslip angle and yaw rate respectively; discretizing the vehicle state space equation using the Euler discretization method to obtain a state transition model; calculating the vehicle state estimate using the unscented Kalman algorithm based on the center of gravity sideslip angle and yaw rate; initializing the parameters of the center of gravity sideslip angle and yaw rate, and defining scale parameters; calculating the mean weight and covariance weight of Sigma points based on the scale parameters, and generating several Sigma points based on the current state estimate and covariance; using the state transition model, predicting each generated Sigma point to obtain the predicted sampling point and predicted state for the next time step, and using the mean weight to perform a weighted summation of the predicted sampling points to obtain the predicted state value;

[0100] Specifically, such as Figure 3 As shown, the vehicle's front wheel steering angle δ is collected in real time by sensors. ft Rear wheel steering angle δ rt , centroid side slip angle β t yaw rate ω t The system collects sensor fault data P (such as jamming, gain change, constant deviation fault signals) and feeds the collected data back through the sensor module of the perception layer.

[0101] Next, taking the collected vehicle motion state as the core, the state vector and measurement vector of the unscented Kalman algorithm are defined, and its state vector expression is as follows:

[0102] ;

[0103] In the formula, This represents the sideslip angle of the vehicle's center of gravity at time k. Let k represent the yaw rate of the vehicle at time k.

[0104] The expression for its measurement vector is:

[0105] ;

[0106] In the formula, This represents the actual output value of the centroid sideslip angle sensor at time k. This represents the actual output value of the yaw rate sensor at time k.

[0107] Next, the continuous-time state-space equation in the space state equation will be... By performing Euler discretization, we obtain the discrete-time state transition model, whose expression is:

[0108] ;

[0109] In the formula, Let f(⋅) represent the control input at time k, and let f(⋅) represent the nonlinear state transition function. Represents state transition noise

[0110] Since the measurement vector directly corresponds to the sensor output of the state vector, the linear measurement equation is defined as follows:

[0111] ;

[0112] In the formula, H represents the measurement matrix, and s represents the block.

[0113] Based on the vehicle's stationary or low-speed start-up conditions, assuming the initial initial centroid sideslip angle β(0)≈0 and the initial yaw rate ω(0)≈0, the scale parameter is defined as:

[0114] ;

[0115] In the formula, The value range is [0.01, 1]. The range of values ​​for κ is ≥ 0, and the value of β is 2.

[0116] Next, the mean weight and covariance weight of the Sigma points are calculated based on the scale parameter. The expression for the covariance weight is as follows:

[0117] ;

[0118] The expression for mean weight is:

[0119] ;

[0120] For example, substituting n=2, α=0.01, κ=0, and β=2 into the mean weight and covariance weight, we can calculate: ≈0.00005, ~ ≈0.25001, ensuring the total weight is 1.

[0121] For the state estimation and covariance at time k, generate 2n+1=5 Sigma points. :

[0122] ;

[0123] ;

[0124] ;

[0125] In the formula, This represents the result of matrix decomposition. Indicates state estimation, This represents the covariance.

[0126] Next, based on the state transition model, prediction is performed for each Sigma point to obtain the predicted sampling point and predicted state at time k+1, expressed as:

[0127] ;

[0128] By using the mean weight to perform a weighted summation of the predicted sampling points, the predicted state value at time k+1 is obtained, and its summation expression is:

[0129]

[0130] By combining the actual sensor measurement value z(k+1), the predicted state is corrected to obtain the final state estimate:

[0131] Prediction of measurement sampling points:

[0132] ;

[0133] Remeasure the predicted mean:

[0134]

[0135] Next, measure the predicted covariance:

[0136] ;

[0137] Next, calculate the state-measure cross-covariance:

[0138] ;

[0139] The Kalman gain is calculated using the prediction covariance and the state-measurement cross-covariance.

[0140] ;

[0141] Finally, the corrected state estimation formula is obtained:

[0142] ;

[0143] And the state covariance formula:

[0144] ;

[0145] After completing the state estimation at time k+1, the corrected sideslip angle and yaw rate are used as the reference state output for fault diagnosis, and simultaneously... and As the initial input for the unscented Kalman filter estimation at the next time step, a cyclical iterative process of "sampling-prediction-correction" is implemented to ensure the real-time and continuous nature of the state estimation during vehicle operation.

[0146] S3. Construct a residual baseline model, calculate the expected output value of real-time data, and use the pre-acquired actual output value and expected output value to calculate and obtain the residual sequence. Use the residual sequence to determine the faulty sensor.

[0147] As a preferred embodiment, the steps of constructing a residual benchmark model, calculating the expected output value of real-time data, and using the pre-acquired actual output value and the expected output value to calculate a residual sequence, and using the residual sequence to determine the faulty sensor, include the following steps: processing the data of the real-time vehicle operation process using the residual benchmark model to obtain the expected state value; based on the expected state value, constructing a dynamic equation regularization term and an error loss, using the dynamic equation regularization term and the error loss to perform a weighted summation to obtain a total loss function, and using the total loss function to optimize the residual benchmark model to obtain the expected output value; substituting the expected state value and the vehicle's basic parameters into the lateral motion balance equation and the transverse motion balance equation to obtain the lateral dynamic constraint residual and the transverse dynamic constraint residual, and weighting the two residuals to obtain the dynamic equation regularization term; subtracting the expected state value from the actual data collected by the sensor to obtain the error loss, and weighting the error loss with the dynamic equation regularization term to obtain the total loss function. The residual terms are obtained by subtracting the pre-acquired actual output from the expected output value. The residual terms are then aggregated to obtain a residual sequence. A fault vector table is constructed based on the residual sequence, and the faulty sensor is identified by combining the preset residual threshold and the residual sequence.

[0148] Specifically, Physical Information Neural Networks (PINNs) are used to construct the residual baseline model, and real-time data of the vehicle's operation is used as network input to predict the expected output δ of the front wheel steering angle. f0 Expected output δ of rear wheel steering angle r0 The desired outputs of yaw rate ω0 and sideslip angle β0 are used as network outputs. A regularization term from the vehicle dynamics equations is introduced to force the network outputs to satisfy the following expression:

[0149] ;

[0150] ;

[0151] In the formula, This represents the rate of change of momentum of the lateral net force on the vehicle. This represents the rate of change of the yaw moment of the vehicle about the z-axis. This indicates the lateral force of the vehicle's front tires. This indicates the lateral force of the vehicle's rear tires. This represents the yaw moment of the vehicle's front tires about its center of gravity, indicating the lateral force exerted by the front tires. This represents the yaw moment of the vehicle's rear tires about its center of mass.

[0152] Next, using the vehicle's lateral motion equilibrium equations and the vehicle's transverse motion equilibrium equations, the constraints that PINNs must forcibly satisfy are determined, and the output of PINNs is then processed. and vehicle inherent parameters Substituting into the equation, the difference between the left and right sides of the equation is defined as the dynamic constraint residual.

[0153] The expression for the lateral dynamic constraint residual is:

[0154] ;

[0155] The expression for the yaw dynamic constraint residual is:

[0156] ;

[0157] Next, the lateral dynamic constraint residuals and yaw dynamic constraint residuals are incorporated into PINNs for training, and the two residuals are weighted and summed to obtain the regularization term of the dynamic equation, which is expressed as follows:

[0158] ;

[0159] In the formula, Represents the canonical term of the dynamic equation. as well as Both represent non-negative constants.

[0160] The training objective of PINNs is to minimize the total loss function, which consists of two parts: the prediction error loss and the regularization term of the dynamic equation.

[0161] The formula for calculating the deviation between the expected state output by PINNs and the actual state acquired by the sensor is as follows:

[0162] ;

[0163] In the formula, N represents the number of training samples. as well as All of these represent the actual state data collected by the sensors when the vehicle is fault-free. This represents the prediction output of PINNs for the i-th sample.

[0164] Next, the prediction error loss and the regularization term of the dynamic equation are weighted and summed as the final optimization objective of PINNs. The summation formula is as follows:

[0165] ;

[0166] In the formula, λ phy This is represented as the dynamic constraint weight coefficient, and its value range is λ. phy >0, preferably λ phy =1.0, ensuring that the network output strictly satisfies the dynamic equations while meeting the prediction accuracy target.

[0167] Finally, based on the expected state output by PINNs and the actual output of the sensor, the corrected residual sequence is calculated. First, the core residual term is defined:

[0168] r0=|δ f -δ r |、r1=|δ f -δ f0 |、r2=|δ r -δ r0 |, r3=|ω-ω0|, r4=|β-β0|, r5=|δ f -f(ω0,β0)|、r6=|δ r -g(ω0,β0)|.

[0169] The above residual terms are aggregated into a residual sequence, and a sensor fault vector table based on the residual sequence is constructed, as shown in Table 1:

[0170] Table 1 Fault Vector Table

[0171] <![CDATA[r0]]> <![CDATA[r1]]> <![CDATA[r2]]> <![CDATA[r3]]> <![CDATA[r4]]> <![CDATA[r5]]> <![CDATA[r6]]> Faulty sensor 1 1 0 1 1 1 0 Front wheel steering angle sensor 1 0 1 1 1 0 1 Rear wheel steering angle sensor 0 0 0 1 1 1 1 Yaw rate sensor

[0172] For example, let r i(i=0,1,2,...,6) represents the residual. When the residual calculation output is 0, it means the residual is less than the threshold and can be ignored. When the residual calculation output is 1, it means the residual is greater than the threshold. After determining the residual output value, the type of sensor that is malfunctioning can be determined by comparing it with the sensor fault vector table.

[0173] S4. Based on the fault sensor, the model-free control algorithm is initialized using the vehicle's two-degree-of-freedom dynamic model. The DQN network algorithm is used to optimize the model-free control algorithm to obtain the reinforcement learning model-free control algorithm. The vehicle state estimate is used as the input of the reinforcement learning model-free control algorithm to adjust the tire steering angle control during the vehicle's operation in real time.

[0174] As a preferred embodiment, the method of initializing the model-free control algorithm based on the fault sensor and the vehicle's two-degree-of-freedom dynamics model, optimizing the model-free control algorithm using the DQN network algorithm to obtain a reinforcement learning model-free control algorithm, and using the vehicle state estimate as the input of the reinforcement learning model-free control algorithm to adjust the tire steering angle control during vehicle operation in real time includes the following steps: initializing the core parameters of the model-free control algorithm, setting the input and output dimensions of the model-free control algorithm after initialization, and constructing a discrete-time system description of the model-free control algorithm based on the vehicle's two-degree-of-freedom dynamics model; the core parameters of the model-free control algorithm are the initial pseudo-partial derivatives and control parameters; the input dimensions of the model-free control algorithm are the front wheel steering angle and the rear wheel steering angle, and the output dimensions of the model-free control algorithm are the center of gravity yaw angle and the yaw rate. The vehicle operating environment is modeled as a Markov decision process, and the state space, action space, and reward function are defined. A DQN network algorithm with an attention mechanism is constructed, and the parameters of the initialized model-free control algorithm are trained and optimized to obtain optimized parameters. The weight parameters of the attention Q network and the target attention Q network, as well as the experience replay buffer, are initialized. Based on the current state and the attention Q network, the control parameters of the model-free control algorithm are adjusted using preset execution actions. The adjusted control parameters are input into the model-free control algorithm, and the real-time reward is calculated according to the reward function. The new state is recorded, and the current state, execution action, real-time reward, and new state are merged into an experience quadruple, which is then placed into the experience buffer.

[0175] When the experience buffer reaches a preset threshold, a batch of experience groups is randomly sampled, and the target Q-value is calculated. The mean squared error between the predicted Q-value and the target Q-value of the attention Q-network is calculated, and the mean squared error is used as the loss function. The network parameters are iteratively updated using the gradient descent algorithm. The Q-network parameters that have reached the iteration threshold are synchronized to the target attention Q-network. The above steps are repeated until the attention Q-network converges, and the optimized parameters are obtained. The optimized parameters are substituted into the model-free control algorithm after initialization to obtain the reinforcement learning model-free control algorithm.

[0176] Specifically, after determining the exact location of the faulty sensor, the sideslip angle and yaw rate of the vehicle are used as feedback parameters to control the sideslip angle error e. β and yaw rate control error e γ As input to the reinforcement learning model-free control algorithm (RL-MFAC);

[0177] ;

[0178] ;

[0179] In the formula β Expected ω represents the desired sideslip angle of the center of mass output by the two-degree-of-freedom vehicle dynamics model. Expected β represents the desired yaw rate output by the two-degree-of-freedom vehicle dynamics model. Actual ω represents the actual sideslip angle of the center of gravity output in a four-wheel steering vehicle. Actual This represents the actual yaw rate output by a four-wheel steering vehicle.

[0180] The controller outputs the front wheel steering angle and the rear wheel steering angle to jointly control the four-wheel steering vehicle and achieve fault-tolerant control.

[0181] First, the system input and output data are obtained from the vehicle's two-degree-of-freedom dynamics model, including historical and current vehicle state information (center of gravity sideslip angle β, yaw rate ω) and control input (front wheel steering angle δ). f Rear wheel steering angle δ r ) and sensor fault data P.

[0182] Next, the core parameters of the Model-Free Control Algorithm (MFAC) are initialized, including the initial pseudo-partial derivatives φ0, control parameters μ (smoothing factor), ρ (learning rate), η (weight factor), and λ (regularization coefficient). The input-output dimension of MFAC is then set to two inputs and two outputs (the input being the front wheel steering angle δ). f Rear wheel steering angle δ r Output centroid sideslip angle β t yaw rate ω t ).

[0183] Based on the two-degree-of-freedom dynamics model of the vehicle, a discrete-time system description of MFAC is constructed:

[0184] ;

[0185] In the formula, Indicates system output, Represents the control input and transforms the nonlinear system through compact-form dynamic linearization:

[0186] ;

[0187] In the formula, Represents the pseudo-partial derivative matrix. Indicates the increment of the output. This indicates the increment of the input.

[0188] The vehicle operating environment is then modeled as a Markov decision process (MDP) to dynamically optimize the key parameters (μ, ρ, η, λ) of the MFAC.

[0189] First, the state space S is defined as a parameter matrix [μ, ρ, η, λ]. Each state describes the parameter configuration of the current MFAC, reflecting the MFAC's ability to adapt to vehicle faults and changes in operating conditions.

[0190] Where, the smaller μ is, The greater the fluctuation, the more susceptible the MFAC is to fault interference. The larger the μ value, the more... A smooth response indicates that MFAC has strong anti-interference capabilities, but its response will be slightly slower.

[0191] The larger the value of ρ, the more sensitive the response to errors, making it suitable for rapidly changing operating conditions; the smaller the value of ρ, the smoother the response, making it suitable for low-speed operating conditions.

[0192] The larger η is, the faster φ is updated, which can quickly match the system characteristics after a fault; if η is too small, φ is updated slowly.

[0193] The larger λ is, the more conservative the control output; the smaller λ is, the more important the tracking accuracy.

[0194] Define the action space A as a set of parameter adjustment actions, including discrete actions such as increasing or decreasing μ, scaling ρ, and correcting η and λ, to achieve dynamic optimization of the parameters.

[0195] Define the reward function R as:

[0196] ;

[0197] In the formula, E err This represents the vehicle state tracking error (centroid sideslip angle error e). β With yaw rate error e γ (root mean square), S stab This represents the vehicle stability index (the ratio of peak yaw rate to the limit of sideslip angle). A and B both represent weighting coefficients used to balance control accuracy and stability.

[0198] A reinforcement learning agent was built using an attention Q-network, a target attention Q-network, and an experience replay buffer, and then trained.

[0199] Among them, the attention Q network introduces an attention mechanism into the traditional Q network, weights the importance of different parameters in the state space, and outputs the expected cumulative reward of the current state-action pair.

[0200] The target attention Q-network is a copy of the Q-network used to compute the target Q-value to improve training stability.

[0201] The role of the experience replay buffer is to store the experience quadruple (s, a, r, s′) of the agent's interaction with the environment, where s is the current state, a is the action performed, r is the reward obtained, and s′ is the new state after performing the action.

[0202] The training process is as follows Figure 4 As shown:

[0203] Step 1: Initialize the weight parameters of the attention Q network and the target attention Q network, as well as the experience replay buffer (set the capacity to 10,000 experiences).

[0204] Step 2: Based on the current state s and the attention Q-network, the agent selects action a using an ε-greedy strategy (ε initially set to 0.9, linearly decaying to 0.1 with training iterations). This action is executed to adjust the MFAC parameters μ, ρ, η, and λ to ensure the controller parameters effectively improve the overall system control performance. Under low-speed conditions, μλ is decreased and ρη is increased; under high-speed conditions, μλ is increased and ρη is decreased.

[0205] Step 3: Input the adjusted parameters into MFAC, apply them to the steer-by-wire chassis four-wheel steering vehicle model, and calculate the vehicle state tracking error E. err and stability index S stab And calculate the immediate reward r based on the reward function;

[0206] Step 4: Record the new state s′ and store the experience (s,a,r,s′) into the experience replay buffer;

[0207] Step 5: When the amount of data in the buffer reaches a preset threshold (e.g., 1000 records), randomly sample a batch of data (batch size 64) and calculate the target Q value.

[0208] ;

[0209] In the formula, y target This represents the target value, and γ represents the discount factor (taken as 0.9). To find the maximum value among all possible next actions, The output of the target attention Q-network;

[0210] Step 6: Using the mean squared error between the predicted Q-value and the target Q-value of the attention Q-network as the loss function, update the network parameters using the gradient descent algorithm:

[0211] ;

[0212] Step 7: Every 100 iterations, synchronize the parameters of the attention Q network to the target attention Q network;

[0213] Step 8: Repeat steps 2-7 until the attention Q-network converges, and obtain the optimized parameters μ. t ρ t η t , λ t The parameter update formula is as follows:

[0214] ;

[0215] ;

[0216] ;

[0217] ;

[0218] In the formula, α p The learning rate, γ, represents the importance weights. t This represents the discount factor raised to the power of t. This represents the state transition probability.

[0219] Finally, the optimized parameters are substituted into MFAC to form a reinforcement learning-based model-free adaptive control algorithm (RL-MFAC), whose control law formula is:

[0220] ;

[0221] In the formula, This represents the control variable from the previous iteration. Let represent the transpose of the c-th type basis function vector estimated at time t. Indicates the desired vehicle status. Indicates vehicle status. express, express.

[0222] in, Updated in real time using the following formula:

[0223] :

[0224] In the formula, This represents the change in the system output at time k. This represents the change in input at the previous moment. This represents the transpose of the change in system input at the previous moment.

[0225] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0226] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault-tolerant control method for a four-wheel steering vehicle, characterized in that, include: S1. Combining the lateral motion equilibrium equations and the transverse motion equilibrium equations, as well as the relationship between tire lateral force and tire slip angle, a two-degree-of-freedom dynamic model of the vehicle is constructed. S2. Collect real-time data of the vehicle operation process, and use the unscented Kalman algorithm to construct a joint estimator using the real-time data to obtain the vehicle state estimate. S3. Construct a residual baseline model, calculate the expected output value of real-time data, and use the pre-acquired actual output value and expected output value to calculate and obtain the residual sequence. Use the residual sequence to determine the faulty sensor. S4. Based on the fault sensor, the model-free control algorithm is initialized using the vehicle's two-degree-of-freedom dynamic model. The DQN network algorithm is used to optimize the model-free control algorithm to obtain the reinforcement learning model-free control algorithm. The vehicle state estimate is used as the input of the reinforcement learning model-free control algorithm to adjust the tire steering angle control during the vehicle's operation in real time.

2. The fault-tolerant control method for a four-wheel steering vehicle according to claim 1, characterized in that, The construction of a two-degree-of-freedom vehicle dynamics model, combining the lateral motion equilibrium equations and the transverse motion equilibrium equations, as well as the relationship between tire lateral force and tire slip angle, includes: S11. Construct the vehicle's lateral motion equilibrium equations and the vehicle's transverse motion equilibrium equations; S12. Based on the vehicle's lateral motion balance equation and the vehicle's transverse motion balance equation, when the vehicle tire is in the linear region, the vehicle tire's lateral force equation and the vehicle tire's slip angle equation are obtained. S13. By combining the vehicle's lateral motion equilibrium equations, the vehicle's transverse motion equilibrium equations, the vehicle's tire lateral force equations, and the vehicle's tire slip angle equations, the vehicle's state-space equations are obtained.

3. The fault-tolerant control method for a four-wheel steering vehicle according to claim 1, characterized in that, The process of collecting real-time data during vehicle operation and using the unscented Kalman algorithm to construct a joint estimator based on the real-time data to obtain vehicle state estimates includes: S21. Collect real-time data on vehicle operation, including front wheel steering angle, rear wheel steering angle, center of gravity sideslip angle, yaw rate, and sensor fault data. S22. Define the state vector and measurement vector of the unscented Kalman algorithm using the centroid sideslip angle and yaw rate, respectively. S23. Discretize the vehicle state-space equations using the Euler discretization method to obtain the state transition model; S24. Based on the sideslip angle and yaw rate, the vehicle state estimate is calculated using the unscented Kalman algorithm.

4. The fault-tolerant control method for a four-wheel steering vehicle according to claim 3, characterized in that, The vehicle state estimate calculated using the unscented Kalman algorithm based on the sideslip angle and yaw rate includes: S241, the parameters of the initial centroid sideslip angle and yaw rate, and the scale parameter; S242. Calculate the mean weight and covariance weight of the Sigma points based on the scale parameter, and generate several Sigma points based on the current state estimate and covariance. S243. Using the state transition model, predict each generated Sigma point to obtain the predicted sampling point and predicted state at the next time step, and use the mean weight to perform a weighted summation of the predicted sampling points to obtain the state prediction value. S244. Based on the actual measured values, calculate the measurement prediction covariance and cross-covariance, and use the measurement prediction covariance and cross-covariance to calculate the Kalman gain to obtain the vehicle state estimate.

5. The fault-tolerant control method for a four-wheel steering vehicle according to claim 1, characterized in that, The process of constructing a residual baseline model, calculating the expected output value of real-time data, and using the pre-acquired actual output value and expected output value to calculate and obtain a residual sequence, and using the residual sequence to determine the faulty sensor, includes: S31. Use the residual baseline model to process the data of the real-time vehicle operation process to obtain the desired state value; S32. Based on the desired state value, construct the regularization term and error loss of the dynamic equation, use the regularization term and error loss of the dynamic equation to perform a weighted summation to obtain the total loss function, and use the total loss function to optimize the residual benchmark model to obtain the desired output value. S34. Calculate the difference between the pre-acquired actual output and the expected output value to obtain residual terms. Aggregate the residual terms to obtain a residual sequence. Construct a fault vector table based on the residual sequence and analyze it in conjunction with the preset residual threshold and residual sequence to determine the faulty sensor.

6. The fault-tolerant control method for a four-wheel steering vehicle according to claim 5, characterized in that, Based on the desired state value, a regularization term and error loss are constructed from the dynamic equation. The total loss function is obtained by weighted summation of the regularization term and error loss, and then optimized using the total loss function to obtain the desired output value, including: S321. Substitute the desired state value and the vehicle's basic parameters into the lateral motion balance equation and the transverse motion balance equation to obtain the lateral dynamic constraint residual and the transverse dynamic constraint residual. Then, weight the two residuals to obtain the regularization term of the dynamic equation. S322. Subtract the desired state value from the actual data collected by the sensor to obtain the error loss. Sum the error loss with the regularization term of the dynamic equation to obtain the total loss function.

7. The fault-tolerant control method for a four-wheel steering vehicle according to claim 1, characterized in that, The method involves initializing a model-free control algorithm based on a fault sensor and a two-degree-of-freedom vehicle dynamics model, optimizing the model-free control algorithm using a DQN network algorithm to obtain a reinforcement learning model-free control algorithm, and using the vehicle state estimate as the input to the reinforcement learning model-free control algorithm to adjust the tire steering angle control during real-time vehicle operation. This includes: S41. Initialize the core parameters of the model-free control algorithm, set the input and output dimensions of the model-free control algorithm after initialization, and construct the discrete-time system description of the model-free control algorithm based on the two-degree-of-freedom dynamics model of the vehicle. S42. Model the vehicle operating environment as a Markov decision process and define the state space, action space, and reward function. S43. Construct a DQN network algorithm with an attention mechanism, and train and optimize the parameters of the algorithm without model control after initialization to obtain optimized parameters; S44. Substitute the optimized parameters into the model-free control algorithm after initialization to obtain the reinforcement learning model-free control algorithm.

8. The fault-tolerant control method for a four-wheel steering vehicle according to claim 7, characterized in that, The core parameters of the model-free control algorithm are initialized, and the input and output dimensions of the model-free control algorithm are set after initialization. Based on the two-degree-of-freedom dynamics model of the vehicle, the discrete-time system description of the model-free control algorithm is constructed, including: S411, the core parameters of the model-free control algorithm are the initial pseudo-partial derivatives and control parameters; S412, the input dimensions of the model-free control algorithm are the front wheel angle and the rear wheel angle, and the output dimensions of the model-free control algorithm are the center of gravity lateral angle and the yaw rate.

9. A fault-tolerant control method for a four-wheel steering vehicle according to claim 8, characterized in that, The DQN network algorithm with attention mechanism is constructed, and the parameters of the initial model-free control algorithm are trained and optimized to obtain the optimized parameters, including: S431. Initialize the weight parameters of the attention Q network and the target attention Q network, as well as the experience replay buffer; S432. Based on the current state and the attention Q network, the control parameters of the model-free control algorithm are adjusted using preset execution actions; S433. Input the adjusted control parameters into the model-free control algorithm and calculate the real-time reward based on the reward function; S434. Record the new state, merge the current state, the executed action, the immediate reward, and the new state into an experience quadruple, and put the experience quadruple into the experience buffer. S435. When the experience buffer reaches the preset threshold, randomly sample a batch of experience groups and calculate the target Q value; S436. Calculate the mean square error between the predicted Q value and the target Q value of the attention Q network, use the mean square error as the loss function, and iteratively update the network parameters through the gradient descent algorithm. S437. Synchronize the Q-network parameters that have reached the iteration threshold to the target attention Q-network; S438. Repeat steps S432 to S437 until the attention Q network converges and the optimized parameters are obtained.

10. A fault-tolerant control method for a four-wheel steering vehicle according to claim 7, characterized in that, The DQN network algorithm is used to optimize the model-free control algorithm, resulting in the control law formula for the reinforcement learning model-free control algorithm: ； In the formula, This represents the control quantity at time k. This represents the control variable from the previous iteration. Indicates time k The next estimated first c Transpose of base function vectors express k+1 Expected vehicle condition at all times express k Vehicle status at any given time. This represents the step size factor under reinforcement learning updates. This represents the regularization factor under reinforcement learning updates.

Images