State estimation method and fault detection method for autonomous vehicle

Abstract

The invention relates to the field of vehicle automatic driving, and discloses an automatic driving vehicle state estimation method and a fault detection method, comprising the following steps: S1, firstly establishing a four-wheel independent drive electric vehicle dynamics model and a steer-by-wire system model, the method is used for accurately reflecting dynamic response behaviors of the four-wheel drive electric vehicle under different working conditions. S2, establishing a CAN bus delay fault dynamic model based on the reference model obtained in the step S1; s3, fusing the vehicle dynamic state and the camera measurement information based on an algorithm so as to obtain accurate vehicle state information; and S4, timely and accurate detection of the CAN communication delay fault is realized through the vehicle state information obtained in the step S3 and by using a strategy. According to the method, the dynamic model of the four-wheel independent drive electric vehicle and the visual information of the camera are fused through the unscented Kalman filtering algorithm, and the state estimation precision and robustness are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to a state estimation method and a fault detection method for autonomous vehicles. Background Technology

[0002] In autonomous driving systems, the ability to perceive the environment and recognize its own state directly determines the vehicle's driving safety and level of intelligence. State estimation, as a bridge between perception and control, infers key parameters such as the vehicle's position, attitude, speed, and acceleration in real time by fusing data from multiple sensors, providing accurate and reliable input for path planning, behavior decision-making, and motion control modules.

[0003] However, in practical applications, various sensors are susceptible to external environmental interference, hardware failures, or communication delays, leading to drift, abrupt changes, or interruptions in the collected data, which in turn can cause state estimation errors or even system failures. The introduction of fault detection mechanisms helps to identify potential anomalies early and maintain stable system operation through fault-tolerant strategies or reconstruction algorithms. Therefore, constructing a state estimation method with high accuracy and robustness, and integrating real-time, highly reliable fault detection technology, is a core requirement for achieving safety redundancy design in high-level autonomous driving systems and a key technological pillar for ensuring the stable operation of autonomous vehicles in complex dynamic environments. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an adaptive fault-tolerant control method for autonomous vehicles. This method solves the problem that in real-world applications, various sensors are susceptible to external environmental interference, hardware failures, or communication delays, which can lead to drift, abrupt changes, or interruptions in the collected data, resulting in state estimation errors or even system failures.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a state estimation method and a fault detection method for an autonomous vehicle, comprising the following steps:

[0006] S1. First, establish a dynamic model of the four-wheel independent drive electric vehicle and a steer-by-wire system model to more accurately reflect the dynamic response behavior of the four-wheel drive electric vehicle under different working conditions, and provide a benchmark model for subsequent steps.

[0007] S2. Based on the baseline model obtained in step S1, and considering the impact of CAN bus delay fault on steering control commands, establish a dynamic model of CAN bus delay fault.

[0008] S3. Then, by using the baseline model from step S1 and the CAN bus delay fault dynamic model from step S2, and by fusing vehicle dynamics state and camera measurement information based on the unscented Kalman filter algorithm, accurate vehicle state information is obtained, providing a precise state baseline for subsequent fault detection.

[0009] S4. Using the vehicle status information obtained in step S3, and employing a Bi-LSTM-based fault detection method, timely and accurate detection of CAN communication delay faults is achieved, ultimately realizing the purpose of estimating and detecting vehicle status and vehicle faults during autonomous driving.

[0010] Preferably, in step S1, the different operating conditions include: longitudinal motion operating condition, lateral steering operating condition, yaw dynamic operating condition, environmental adhesion operating condition, load change operating condition, and steer-by-wire operating condition.

[0011] Preferably, in step S1, the dynamic model of the four-wheel independent drive electric vehicle includes: kinematic equations and generalized torque calculation logic;

[0012] The kinematic equations are used to establish the dynamic relationship between the vehicle's center of gravity sideslip angle, yaw rate about the vertical axis, longitudinal velocity of the center of gravity, and front wheel steering angle. The system matrix, input matrix, and disturbance matrix reflect the coupling characteristics of the vehicle's longitudinal, lateral, and yaw directions.

[0013] The generalized torque calculation is based on the driving torque of the left front, right front, left rear, and right rear wheels, the wheel ground contact width, and the equivalent rolling radius. It integrates the resultant torque formed by the lateral forces of each wheel at the center of mass. At the same time, it considers the small steering angle of the front wheels under typical driving conditions and approximately ignores the influence of the lateral force component in the longitudinal direction to simplify the coupling relationship of the model and improve the calculation efficiency and practicality of the model.

[0014] Preferably, in step S1, the steer-by-wire system model fully describes the electromechanical dynamic process through multiple sub-models, including:

[0015] Motor torque model: Based on the product relationship between the motor torque constant and the armature current, the motor output torque is calculated, reflecting the driving effect of the current on the torque;

[0016] Motor voltage balance model: It relates the motor input voltage, back electromotive force, induced voltage generated by armature winding inductance and voltage drop generated by resistance, and fully describes the dynamic characteristics of the motor electrical circuit, supporting motor response analysis;

[0017] Gear and rack transmission model: Combining the effective mass of the rack, damping, external forces on the left and right wheels, and parameters of the motor and reducer, the motor output torque is converted into the linear displacement of the steering rack, reflecting the force and displacement relationship of mechanical transmission;

[0018] Electronic control unit control command generation logic: Taking the error between the desired front wheel steering angle and the current measured steering angle as input, the actual steering command is calculated through the ECU's built-in control algorithm to ensure that the steering angle accurately tracks the desired target, and finally forms a vehicle control input vector containing the actual steering angle and the longitudinal force of each wheel.

[0019] Preferably, in step S1, the dynamic model of the four-wheel independent drive electric vehicle further includes longitudinal control modeling logic: introducing the desired longitudinal acceleration and desired longitudinal velocity, constructing two types of error state variables—acceleration error and velocity error; wherein, the total longitudinal force of the vehicle is obtained by summing the longitudinal forces of each wheel, and the longitudinal force of each wheel is approximately calculated by dividing the product of the driving torque of the corresponding wheel and the cosine of the front wheel steering angle by the equivalent rolling radius of the wheel, providing an error feedback basis for longitudinal speed tracking control.

[0020] Preferably, in step S2, the CAN bus delay fault dynamic model includes:

[0021] Unified modeling is achieved through segmented logic: For constant communication delay, a time delay model with a fixed time range is adopted, delaying the timestamp of the expected steering control command by a fixed amount of time; for random communication delay, a time-varying delay model following a normal distribution is adopted, delaying the timestamp of the expected steering control command by a random amount of time, thereby generating erroneous steering control commands caused by CAN bus delay; at the same time, in the CAN bus delay fault scenario, the erroneous steering control command is substituted into the system control input vector, and combined with the discrete state space model in step S1, a vehicle dynamics model under the fault state is constructed to accurately describe the impact of delay on vehicle dynamic characteristics.

[0022] Preferably, in step S3, the fusion of vehicle dynamics state and camera measurement information based on the unscented Kalman filter algorithm is achieved through the following sub-steps:

[0023] Prediction phase: Generate Sigma points covering the state space. Based on the baseline model in step S1 or the fault model in step S2, calculate the predicted value of each Sigma point at the next time step. Calculate the state mean by weighted summation. Combine the deviation of each Sigma point from the mean and the process noise covariance to calculate the state covariance and complete the state prior estimation.

[0024] Data fusion stage: The Sigma points from the prediction stage are mapped to the measurement space through a measurement function to obtain the Sigma points in the measurement space; the predicted measurement mean is calculated by weighted summation, and the measurement covariance is calculated by combining the deviation of each measurement Sigma point from the mean and the measurement noise; at the same time, the cross covariance between the state Sigma points and the measurement Sigma points is calculated, the Kalman gain is solved, and the weight of measurement information on state correction is quantified.

[0025] Update phase: Combining the measurement vectors acquired by the camera, the state mean and covariance are corrected by Kalman gain to obtain the posterior state estimate, thereby realizing the fusion of vehicle dynamics model and visual measurement information and improving the state estimation accuracy.

[0026] Preferably, in step S3, the acquisition of camera measurement information includes lane boundary fitting and error calculation: by bird's-eye view transformation, threshold segmentation and sliding window quadratic curve fitting, a unified lane curve characterized by three coefficients is obtained; based on the lane curve, the lateral error is calculated with the negative value of the curve constant term, and the heading error is approximately calculated with the negative value of the curve linear term, so as to provide accurate visual measurement input for UKF fusion.

[0027] Preferably, in step S4, the fault detection method based on Bi-LSTM is implemented through time series modeling and residual discrimination, specifically including:

[0028] Residual construction: Define the system residual as the difference between the actual measured front wheel steering angle and the desired front wheel steering angle, and quantify the deviation between the actual execution of the steering control command and the desired target;

[0029] Temporal feature extraction: The input sequence is composed of the residuals at the current time and the adjacent time steps and fed into the Bi-LSTM network. The network captures the historical temporal dependence of the residuals through forward LSTM and captures the future temporal dependence of the residuals through backward LSTM. The hidden states of the two are fused to extract the dynamic evolution features of the residuals in the time dimension.

[0030] Residual prediction and discrimination: The fused time-series features are projected onto the residual space through a fully connected layer to obtain the predicted residual; the mean square error between the actual residual and the predicted residual is calculated, and combined with the dual preset thresholds and the state variables of the previous moment, the existence of delay faults in the CAN bus is determined, so as to achieve accurate fault identification and control of misjudgment / missed judgment.

[0031] Preferably, in step S4, the CAN bus delay fault detection logic specifically includes:

[0032] When the mean square error is greater than the second threshold, it is determined that there is a significant deviation, i.e., a CAN bus delay fault.

[0033] When the mean square error is between the first threshold and the second threshold, and the state variable at the previous moment is greater than the first threshold, it is determined that there is a potential CAN bus delay fault.

[0034] When the mean square error is less than or equal to the first threshold, the CAN bus is considered to be working normally and no delay fault is detected.

[0035] This invention provides an adaptive fault-tolerant control method for autonomous vehicles. It has the following advantages:

[0036] 1. This invention significantly improves the accuracy and robustness of state estimation by fusing the dynamics model of a four-wheel independently driven electric vehicle with camera visual information through the unscented Kalman filter (UKF) algorithm: the vehicle dynamics model provides systematic prior knowledge for state estimation, and the camera can perceive lane boundaries in real time and calculate lateral and heading errors. The fusion of the two effectively avoids the problems of "large data errors and sensor drift" that traditionally rely solely on inertial measurement units (IMUs) or single sensors. At the same time, UKF can efficiently handle the state propagation and observation model of the vehicle's nonlinear dynamic system, which is superior to the traditional linear Kalman filter (KF). Even when the reliability of single-source data decreases, it can still ensure the accuracy of core state estimations such as lateral and heading.

[0037] 2. The fault detection method based on Bi-LSTM in this invention has a much higher ability to identify and adapt to CAN bus delay faults than traditional methods. Compared with traditional threshold methods, statistical methods or static model anomaly detection, Bi-LSTM's bidirectional time-series modeling capability can simultaneously capture the historical and future evolution trends of residuals, overcoming the shortcomings of unidirectional LSTM ("relying only on historical data and lacking dynamic tracking") and static methods ("unable to adapt to complex working conditions"). In response to the uncertainty and randomness of CAN communication delay, Bi-LSTM can learn hidden time-dependent features, accurately identify complex delay fault modes that are easily missed by traditional methods, and the neural network model can adapt to various operating conditions, with stronger anti-interference ability and reduced misjudgment / missed judgment caused by sudden anomalies.

[0038] 3. This invention improves the overall control performance and safety tolerance of autonomous vehicles through a collaborative design of high-precision state estimation and efficient fault detection: On the one hand, accurate vehicle state estimation provides reliable system feedback for subsequent path tracking control, avoiding control deviations caused by state errors and significantly enhancing the stability and responsiveness of path tracking; on the other hand, timely and accurate CAN bus delay fault detection can quickly warn of communication anomalies, prevent erroneous responses of the steer-by-wire system caused by delayed steering commands, avoid control performance degradation or safety risks, and better meet the safety operation requirements of autonomous driving in complex dynamic environments compared to traditional methods. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention;

[0040] Figure 2 This is a diagram of the vehicle dynamics model in step S1 of the present invention;

[0041] Figure 3 This is a flowchart of the overall fault detection process in step S4 of the present invention. Detailed Implementation

[0042] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Example:

[0044] Please see the appendix Figure 1 -Appendix Figure 3 This invention provides a state estimation method and a fault detection method for autonomous vehicles, comprising the following steps:

[0045] S1. First, establish a dynamic model of the four-wheel independent drive electric vehicle and a steer-by-wire system model to more accurately reflect the dynamic response behavior of the four-wheel drive electric vehicle under different working conditions, and provide a benchmark model for subsequent steps.

[0046] S2. Based on the baseline model obtained in step S1, and considering the impact of CAN bus delay fault on steering control commands, establish a dynamic model of CAN bus delay fault.

[0047] S3. Then, by using the baseline model from step S1 and the CAN bus delay fault dynamic model from step S2, and by fusing vehicle dynamics state and camera measurement information based on the unscented Kalman filter algorithm, accurate vehicle state information is obtained, providing a precise state baseline for subsequent fault detection.

[0048] S4. Using the vehicle status information obtained in step S3, and employing a Bi-LSTM-based fault detection method, timely and accurate detection of CAN communication delay faults is achieved, ultimately realizing the purpose of estimating and detecting vehicle status and vehicle faults during autonomous driving.

[0049] In step S1, the establishment of the dynamic model of the four-wheel independent drive electric vehicle and the steer-by-wire system model comprehensively considers the dynamic characteristics of the vehicle in the longitudinal, lateral and yaw directions in the plane, and ignores the higher-order dynamic effects caused by changes in vehicle attitude (such as pitch and roll).

[0050] The vehicle's kinematic equations are expressed as follows:

[0051]

[0052] Where β represents the vehicle's center of gravity sideslip angle, ω z V represents the yaw rate of the vehicle about its vertical axis. x Let δ be the longitudinal velocity at the vehicle's center of mass. f Let M be the steering angle of the front wheels. The generalized torque M of the vehicle.z The main force originates from the resultant torque formed by the lateral forces of each wheel at the center of mass, which can be approximated as:

[0053]

[0054] In the formula, W represents the ground contact width of the wheel, and T xlf T xrf T xlr T xrr These represent the driving torques of the vehicle's left front, right front, left rear, and right rear wheels, respectively. t Let sinδ be the equivalent rolling radius of the wheel. Considering the front wheel steering angle sinδ under typical driving conditions... f It is usually small and can be approximated as sinδ f ≈δ f Therefore, the component of the lateral force in the longitudinal direction has a small impact and can be ignored, thus simplifying the coupling relationship of the model.

[0055] In longitudinal control modeling, desired longitudinal velocity and longitudinal acceleration are introduced to construct error state variables:

[0056]

[0057] Among them, a xd and v xd These represent the reference longitudinal acceleration and reference longitudinal velocity, respectively, used to describe the desired vehicle motion state. The total longitudinal force F acting on the vehicle... xt It can be given by the following formula:

[0058]

[0059] In the formula, F xlf ,F xrf ,F xlr F xrr This represents the longitudinal component of the tire force.

[0060] To achieve accurate tracking of the target path, a path tracking model based on error state is introduced, defining lateral error and yaw error. Its dynamic model includes:

[0061]

[0062]

[0063] Where e y and For lateral error and yaw error, ω zd Let λ be the desired yaw rate and λ be the road curve.

[0064] The steering system employs a steer-by-wire architecture, integrating motor dynamics and a mechanical steering transmission structure. This system achieves precise control of wheel angle through motor output torque, eliminating the rigid connection of the traditional mechanical steering column and improving control flexibility and adjustability. The motor torque is:

[0065] T m =K t *I

[0066] In the formula, K t Let I be the motor torque constant and I be the armature current. The voltage balance relationship of the motor is expressed as:

[0067]

[0068] In the formula, U represents the input voltage of the motor, K1 is the back electromotive force coefficient, and θ m Let θ be the torsion angle of the motor, and L and R be the inductance and resistance of the armature winding, respectively. This model fully describes the dynamic characteristics of the motor's electrical circuit, which helps in analyzing the motor response and designing corresponding control strategies.

[0069] The gear and rack transmission mechanism converts the motor output torque into the linear displacement of the steering rack, and the relationship is expressed as follows:

[0070]

[0071] In the formula M r and B r For effective mass and damping, and External force on left and right wheels, K md To achieve torsional stiffness, G m For the reduction ratio, r p Let P be the gear pitch. The dynamics of the motor-reducer combination are expressed as:

[0072]

[0073] In the formula J m B is the moment of inertia of the motor. m This is the damping coefficient of the motor.

[0074] In order to achieve the desired front wheel steering angle For accurate tracking, the electronic control unit (ECU) of the steer-by-wire system generates actual steering commands based on the control algorithm. Its expressions include:

[0075]

[0076] Among them, f ECUThis refers to the control algorithm implemented in the electronic control unit (ECU) of the steering-by-wire system. This algorithm uses the desired steering angle... Compared with the current measured steering angle Using the error information as the core, and based on the system dynamics model and control objectives, the actual steering commands to be executed are calculated and generated. Based on this, the vehicle's control input vector can be uniformly represented as:

[0077] Based on the established vehicle dynamics model and the SBW control logic, the entire vehicle control system can be represented as a state-space model in the discrete time domain:

[0078] ξ k+1 =A d ξ k +B d u k +C d

[0079] The state vector is defined as follows:

[0080] In step S2, the consideration of the impact of CAN bus delay faults on steering control commands specifically includes:

[0081] In practical vehicle control systems, issues such as excessive load or message arbitration conflicts on the CAN bus can cause delays in the transmission of control commands, adversely affecting the system's real-time performance and stability. To accurately assess the impact of such communication delay faults on vehicle control performance, two typical time delay models are introduced into the simulation model: time delay and time-varying delay.

[0082] Time-varying delay is used to simulate constant communication latency, commonly seen in deterministic scheduling scenarios with fixed scheduling overhead, or in persistent transmission delays caused by buffer queuing when network load is relatively stable; while time-varying delay is used to characterize random delays caused by network load fluctuations or communication conflicts. Unified modeling includes:

[0083]

[0084] in, This indicates an incorrect steering control command caused by CAN bus delay, where Δ and τ represent fixed time and random time, respectively.

[0085] In the presence of CAN bus communication delays, the electronic control unit (ECU) may receive delayed or erroneous steering commands, leading to erroneous responses and degraded control performance in the steering steer-by-wire system, and even control malfunctions. In the case of a CAN bus delay fault, the system's control input is represented as follows:

[0086]

[0087] When a communication delay fault occurs on the CAN bus, the vehicle's dynamic characteristics will be significantly affected. The vehicle dynamics model under these circumstances can be described as follows:

[0088]

[0089] In step S3, the fusion of vehicle dynamics state and camera measurement information based on the unscented Kalman filter (UKF) algorithm includes:

[0090] Lateral error e y and heading error Accurate estimation is key to achieving efficient path tracking control. This is achieved by using a monocular camera to acquire lane visual information and combining it with a vehicle kinematics model for state fusion estimation.

[0091] The image processing section extracts lane boundaries using methods such as bird's-eye view transformation, thresholding, and sliding window quadratic curve fitting. The fitted curves include:

[0092] y(x)=A1+A2x+A3x 2

[0093] Where A1, A2, and A3 are the unified lane curve parameters obtained by averaging the coefficients of the fitted equations for the left and right lane boundaries, and x represents the lateral position coordinates in the bird's-eye view (BEV) coordinate system. Based on the fitted lane model, the lateral error e y and heading error The calculation can be performed using the following formula:

[0094]

[0095] in and These are the lateral and heading errors calculated based on camera data.

[0096] The above variables are treated as measurement vectors and used in the subsequent state estimation process. Based on the UKF method, its prediction phase can be described as follows:

[0097]

[0098] in, Let represent the state vector of the i-th sampling point at time k-1. During the prediction phase, the estimates of the state mean and covariance can be calculated in the following manner:

[0099]

[0100]

[0101] Among them, Q k-1 Let w represent the process noise covariance matrix. i with w c These are the weighting coefficients used for calculating the state mean and covariance, respectively.

[0102] During the data fusion phase, the Sigma points generated in the prediction phase need to be mapped to the measurement space using a measurement function. The mapping process includes:

[0103]

[0104] Among them, h k (·) represents the measurement function. Let be the representation of the i-th Sigma point in the measurement space, including lateral and heading error components. Based on this, the predicted measurement value and its covariance can be calculated in the following manner:

[0105]

[0106] In the formula R 1,k Represents measurement noise. Cross-covariance matrix P ξΓ The calculation formula includes:

[0107]

[0108] Kalman gain K k for:

[0109]

[0110] Combine camera measurement information to construct measurement vectors Based on this observation, the update process for the state mean and covariance includes:

[0111]

[0112] In step S4, the Bi-LSTM-based fault detection method includes:

[0113] To identify control faults caused by CAN communication delays, a residual modeling and discrimination method based on a bidirectional long short-term memory (Bi-LSTM) network is introduced. The Bi-LSTM network possesses powerful time-series modeling capabilities by simultaneously capturing the dynamic dependencies between the forward and backward directions in the time series, making it particularly suitable for identifying the changing trends of control signal residuals.

[0114] The residual of the current system is defined as the difference between the actual measurement and the predicted measurement, and is specifically expressed as follows:

[0115]

[0116] At each time step, model k takes the residual sequence of the current time step and its adjacent time steps as input to capture the dynamic evolution characteristics of the residuals in the time dimension, thereby achieving efficient identification of potential failure modes of the system.

[0117] A k ={Z k-i Z k-i+1 ,…,Z k-1 Z k+1 Z k+2 ,…,Z k+n}

[0118] Input sequence A k The sequence is fed into a Bi-LSTM network for processing. The network encodes the sequence using both forward LSTM and backward LSTM to extract the corresponding hidden states. and To capture the dependencies of time series in both directions.

[0119] For both forward and backward LSTMs, the input Z at a given time step k... k In this case, the state update process can be expressed as recursive equations of the following forms:

[0120]

[0121] in, This represents the internal state of the LSTM unit, used to store long-term dependency information. The final output of the Bi-LSTM is obtained by concatenating the forward hidden state and the backward hidden state, i.e.:

[0122]

[0123] The encoded features are projected onto the residual space through a fully connected layer, specifically including:

[0124]

[0125] in, Let η represent the prediction residual output by the Bi-LSTM decoder, where η is the weight matrix of the decoder, and b dec This is the bias term for the decoder. The mean squared error (MSE) between the predicted residual and the actual residual can be calculated using the following formula:

[0126]

[0127] For the predicted residual ζ k Discriminant analysis is performed to determine if a delay fault exists on the CAN bus. The discriminant logic includes:

[0128] When ζ k If the value is greater than μ2, it indicates that there is a significant deviation, and the system is likely experiencing a CAN bus delay fault.

[0129] When μ1 < ζ k If μ2 ≤ μ2, then the state variable ξ from the previous time step is further considered. k-1 When ξ k-1 If the value is greater than μ1, then it is also considered that there may be a CAN bus delay.

[0130] When ζ k If the value is less than μ1, it is determined that the current CAN bus is working normally and no delay has been detected.

[0131] Where μ1 and μ2 are preset residual thresholds, satisfying μ2 = 2μ1.

[0132] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A state estimation method and a failure detection method for an autonomous vehicle, characterized by, The method comprises the following steps: S1, first establish a four-wheel independent drive electric vehicle dynamics model and a steer-by-wire system model, which are used to more accurately reflect the dynamic response behavior of the four-wheel drive electric vehicle under different working conditions, and provide a benchmark model for subsequent steps; S2, then based on the benchmark model obtained in step S1, and considering the influence of CAN bus delay fault on the steering control command, a CAN bus delay fault dynamic model is established; S3, then through the benchmark model of step S1 and the CAN bus delay fault dynamic model of step S2, and based on the unscented Kalman filtering algorithm, the vehicle dynamics state and the camera measurement information are fused to obtain accurate vehicle state information, which provides accurate state benchmark for subsequent fault detection; S4, through the vehicle state information obtained in step S3, and using the fault detection method based on Bi-LSTM, the CAN communication delay fault is detected in time and accurately, and finally the purpose of estimating and detecting the vehicle state and vehicle fault in the automatic driving process is realized.

2. The state estimation method and fault detection method for an autonomous vehicle according to claim 1, characterized by, In step S1, the different working conditions include: longitudinal motion condition, lateral steering condition, yaw dynamic condition, environmental adhesion condition, load change condition and steer-by-wire condition.

3. The state estimation method and fault detection method for an autonomous vehicle according to claim 1, characterized by, In step S1, the four-wheel independent drive electric vehicle dynamics model comprises: kinematics equation and generalized moment calculation logic; The kinematics equation is used to establish the dynamic correlation between the vehicle mass center side slip angle, the yaw angular velocity around the vertical axis, the mass center longitudinal velocity and the front wheel steering angle, and through the system matrix, the input matrix and the disturbance matrix, the coupling characteristics of the vehicle in the longitudinal, lateral and yaw directions are embodied; The generalized moment calculation is based on the driving torque of the left front, right front, left rear and right rear wheels, the wheel ground width and the equivalent rolling radius, integrates the resultant moment of the lateral force of each wheel at the mass center, and considering the characteristics of small front wheel steering angle under typical driving conditions, approximately ignores the component influence of lateral force in the longitudinal direction, to simplify the coupling relationship of the model, improve the calculation efficiency and practicability of the model.

4. The state estimation method and fault detection method for an autonomous vehicle according to claim 1, characterized by, In step S1, the steer-by-wire system model completely describes the electromechanical dynamic process through multiple sub-models, including: Motor torque model: based on the product relationship of motor torque constant and armature current, the motor output torque is calculated, which reflects the driving effect of current on torque; Motor voltage balance model: the motor input voltage, back electromotive force, induced voltage generated by armature winding inductance and voltage drop generated by resistance are related to each other, which completely describes the dynamic characteristics of motor electric circuit and supports motor response analysis; Gear rack transmission model: combining the effective mass of rack, damping, left and right wheel external force and motor-reducer parameters, the motor output torque is converted into linear displacement of steering rack, which embodies the force and displacement relationship of mechanical transmission; Electronic control unit control instruction generation logic: taking the error between the expected front wheel steering angle and the current measured steering angle as input, the actual steering command is calculated through the built-in control algorithm of ECU, which ensures accurate tracking of the expected target of steering angle, and finally forms a vehicle control input vector containing actual steering angle and longitudinal force of each wheel.

5. The method of state estimation and fault detection for an autonomous vehicle according to claim 1, wherein In step S1, the four-wheel independent drive electric vehicle dynamics model further comprises longitudinal control modeling logic: introducing desired longitudinal acceleration and desired longitudinal speed, constructing two types of error state variables: acceleration error and speed error; wherein the total longitudinal force of the vehicle is obtained by summing the longitudinal forces of each wheel, and the longitudinal force of each wheel is calculated by multiplying the driving torque of the corresponding wheel by the cosine of the front wheel steering angle, and dividing by the approximate calculation of the equivalent rolling radius of the wheel, to provide error feedback basis for longitudinal speed tracking control.

6. The method of state estimation and fault detection for an autonomous vehicle according to claim 1, wherein In step S2, the CAN bus delay fault dynamic model comprises: Fault classification modeling is achieved: for constant communication delay, a fixed time range timing delay model is used to delay the timestamp of the desired steering control command by a fixed time; for random communication delay, a time-varying delay model subject to normal distribution is used to delay the timestamp of the desired steering control command by a random time, to generate an error steering control command caused by CAN bus delay; at the same time, in the CAN bus delay fault scene, the error steering control command is substituted into the system control input vector, and the vehicle dynamics model under fault state is constructed by combining the discrete state space model of step S1, to accurately describe the influence of delay on the dynamic characteristics of the vehicle.

7. The method of state estimation and fault detection for an autonomous vehicle according to claim 1, wherein In step S3, the vehicle dynamics state and camera measurement information are fused based on the unscented Kalman filter algorithm, which is achieved through the following sub-steps: Prediction stage: generate Sigma points covering the state space, calculate the predicted value of each Sigma point at the next time based on the reference model of step S1 or the fault model of step S2; calculate the state mean value by weighted summation, and calculate the state covariance by combining the deviation of each Sigma point and the mean value and the process noise covariance, to complete the prior state estimation; Data fusion stage: map the Sigma points of the prediction stage to the measurement space through the measurement function to obtain the Sigma points in the measurement space; calculate the predicted measurement mean value by weighted summation, and calculate the measurement covariance by combining the deviation of each measurement Sigma point and the mean value and the measurement noise; at the same time, calculate the cross-covariance of the state Sigma point and the measurement Sigma point, solve the Kalman gain, and quantify the weight of the measurement information for state correction; Update stage: combine the measurement vector obtained by the camera to correct the state mean value and covariance through the Kalman gain, to obtain the posterior state estimation value, realize the fusion of the vehicle dynamics model and the visual measurement information, and improve the state estimation accuracy.

8. The state estimation method and fault detection method for an automated vehicle according to claim 7, characterized by, In step S3, the acquisition of camera measurement information includes lane boundary fitting and error calculation: through bird's eye transformation, threshold segmentation and sliding window quadratic curve fitting, a unified lane curve represented by three coefficients is obtained; based on the lane curve, the lateral error is calculated by the negative value of the constant term of the curve, and the heading error is approximately calculated by the negative value of the first term of the curve, to provide accurate visual measurement input for UKF fusion.

9. The method of state estimation and fault detection for an autonomous vehicle according to claim 1, wherein In step S4, the fault detection method based on Bi-LSTM is realized through time series modeling and residual discrimination, specifically including: Residual construction: define the system residual as the difference between the actual measured front wheel steering angle and the expected front wheel steering angle, quantify the deviation between the actual execution of the steering control instruction and the expected target; Temporal feature extraction: input sequence composed of current time and adjacent time step residuals into Bi-LSTM network; network captures historical temporal dependence of residuals through forward LSTM, captures future temporal dependence of residuals through backward LSTM, fuses hidden states of both to extract dynamic evolution features of residuals in time dimension; Residual prediction and discrimination: project the fused temporal features into residual space through fully connected layer to obtain predicted residuals; calculate the mean square error between actual residuals and predicted residuals, combine with double preset thresholds and state variables at previous time to discriminate whether there is delay fault of CAN bus, realize accurate identification and misjudgment / omission control of faults.

10. The state estimation method and fault detection method for an automated vehicle according to claim 9, wherein In step S4, the discrimination logic of the CAN bus delay fault is specifically: When the mean square error is greater than the second threshold, it is determined that there is a significant deviation, i.e. the CAN bus delay fault; When the mean square error is between the first threshold and the second threshold, and the state variable at the previous time is greater than the first threshold, it is determined that there is a potential CAN bus delay fault; When the mean square error is less than or equal to the first threshold, it is determined that the CAN bus is working normally and no delay fault is detected.

Images