Full-automatic driving automobile motion state estimation method based on sampling data neural network observer

Abstract

The invention discloses a full-automatic driving automobile motion state estimation method based on a sampling data neural network observer, the sampling data neural network observer is composed of a continuous state observer and a compensation injector, and the method comprises the steps: firstly, employing a radial basis function neural network to approach the unknown dynamic and modeling uncertainty of a vehicle system; the weight of the neural network is updated online through a newly designed weight updating law; for the sampling and delay problems of sensor data, an integral compensation structure is additionally arranged in a compensation injector to offset sampling information loss and relieve delay influence. According to the method, high-precision continuous estimation of the lateral speed of the vehicle can be realized under the conditions of sampling delay and unknown modeling uncertainty, and the lateral control and safety monitoring performance of the full-automatic driving vehicle is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous vehicle state perception and control, and specifically to a method for estimating the motion state of a fully autonomous vehicle based on a sampled data neural network observer. Background Technology

[0002] With the rapid development of fully autonomous driving technology, accurately acquiring vehicle motion status (especially lateral velocity and other lateral state information) is crucial for ensuring the lateral stability of autonomous vehicles, achieving precise trajectory tracking, and ensuring safety monitoring. Lateral velocity, as a core input parameter for vehicle lateral control strategies (such as yaw moment control and trajectory tracking control), is key to improving the performance of autonomous driving systems and is of great value in promoting the implementation of low-cost, high-reliability autonomous driving solutions.

[0003] In complex driving environments, vehicle motion state estimation systems are inevitably affected by complex factors such as non-periodic sensor sampling, data latency (e.g., GPS signal latency, visual sensor processing latency), and uncertainties in unknown system modeling (e.g., tire nonlinearity, external disturbances). Researchers have focused on different influencing factors and proposed various lateral velocity estimation methods: for cases where the system model is known, Kalman filtering and its improved algorithms (e.g., extended Kalman filtering, unscented Kalman filtering) have been developed; for cases where the dynamic approximation is unknown, estimation methods based on neural networks have been developed, and so on.

[0004] Chinese invention patent CN120003502A discloses a soft measurement method for vehicle centroid sideslip angle considering input and output time delays. Compared with this invention, it differs fundamentally in addressing technical issues such as non-periodic sampling, time-varying delays, and dynamics of strongly nonlinear systems in autonomous driving scenarios. The main limitations of the prior art (CN120003502A) are as follows:

[0005] 1. Different applicable scenarios

[0006] The core objective of the prior art document (CN120003502A) is the vehicle's center of gravity sideslip angle, which is calculated by indirectly estimating the lateral velocity and combining it with the longitudinal vehicle speed. It is suitable for general vehicle state perception scenarios and is not designed for the high-precision lateral control requirements of fully autonomous driving. Its technical solution revolves around the soft measurement of the center of gravity sideslip angle and relies on a linear tire model and an extended state observer (ESO) to handle known types of disturbances.

[0007] The core objective of this invention (sampled data neural network observer) is vehicle lateral velocity, directly adapting to the core requirements of fully autonomous driving such as lateral stability control and trajectory tracking; the technical solution is designed for the non-ideal characteristics and complex dynamic characteristics of sensors in fully autonomous driving scenarios, with high-precision continuous estimation of lateral velocity as the core.

[0008] 2. The mechanisms for handling unknown dynamics and disturbances differ.

[0009] The comparison document relies on ESO to estimate the disturbance, but explicitly requires that the type of disturbance be known and the magnitude unknown. Its disturbance model is based on a preset dynamic equation and cannot compensate for unknown modeling uncertainties such as strong tire nonlinearity and model structural errors.

[0010] This invention approximates the unknown dynamics of a system (including disturbances of unknown type, tire nonlinear residual errors, etc.) online using a radial basis function (RBF) neural network, and designs a real-time weight update law. It does not require preset disturbance types and can adaptively compensate for unknown modeling uncertainties under complex working conditions.

[0011] 3. The adaptability of sensors to non-ideal characteristics varies.

[0012] The comparison document only considers the input / output time delay under a fixed sampling period, and compensates for the time delay through Artstein state transformation and output prediction, but does not address the problem of aperiodic sampling. Its discretization model is based on the assumption of a fixed sampling period, which cannot cope with the sampling interval fluctuation caused by environmental interference of the sensor, nor does it handle the coupling effect of delay and aperiodic sampling.

[0013] This invention designs a Sampled Data Neural Network Observer (SDNNO) for non-periodic sampling, and uses a compensation injector to offset the information loss of non-periodic sampling. For delayed scenarios, an integral compensation module is added to the SDNNO to form a Sampled Delayed Data Neural Network Observer (SDDNNO), and coupling constraints between sampling interval and delay are designed to adapt to the coupling characteristics of non-periodic sampling and delay.

[0014] Current lateral velocity estimation methods have significant limitations: Kalman filter-based methods rely on the accuracy of the system model and, being based on discrete system modeling, struggle to adapt to aperiodic sensor sampling and data latency, leading to decreased estimation accuracy. While some neural network methods can approximate unknown dynamics, they often rely on offline training, making them ill-suited for real-time changing driving scenarios. Furthermore, they fail to adequately consider the combined effects of sampling and latency, or their complex structures result in excessively high computational costs, failing to meet the real-time requirements of real-world vehicles. In addition, the uncertainties in modeling unknown vehicle systems and information loss during sampling further exacerbate estimation errors and reduce system stability. Existing technologies have not yet effectively addressed these issues. Therefore, a sampling data neural network observer method for motion state estimation of fully autonomous vehicles is urgently needed, which is of great significance for improving the lateral control and safety monitoring performance of autonomous vehicles. Summary of the Invention

[0015] To address the issue of insufficient lateral velocity estimation accuracy in fully autonomous vehicle motion state estimation caused by non-periodic sensor sampling, data delay, and uncertainties in system modeling (such as tire nonlinearity and external disturbances), this invention proposes a motion state estimation method for fully autonomous vehicles based on a sampled data neural network observer. This method can achieve high-precision continuous estimation of vehicle lateral velocity in complex scenarios, improve the lateral control performance and safety monitoring capabilities of autonomous vehicles, and provide reliable state perception support for low-cost autonomous driving solutions.

[0016] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0017] A method for estimating the motion state of a fully automated vehicle based on a sampled data neural network observer, the method comprising the following steps:

[0018] Step A: Use onboard sensors to acquire the vehicle's longitudinal speed and yaw rate, and simultaneously collect control input information, including steering angle and drive torque, through the vehicle controller local area network bus.

[0019] Step B: Establish a planar motion dynamics model of the vehicle, integrate the Brush tire model to describe lateral forces, and derive the state equations containing unknown modeling uncertainties;

[0020] Step C: For the no-delay sampling scenario, construct a sampling data neural network observer, use a radial basis function neural network to approximate the unknown dynamics of the system, design an online weight update law to achieve adaptive adjustment of neural network parameters, combine a compensation injector to offset the loss of sampling information, and make a preliminary estimate of the lateral velocity based on the collected longitudinal velocity and yaw rate measurements, and then proceed to step E.

[0021] Step D: For sampling scenarios with delay, an integral compensation module is added to the sampling data neural network observer to handle the error accumulation during the delay period, forming a sampling delay data neural network observer. The estimation logic and neural network module of the delay-free scenario are reused. The longitudinal velocity and yaw rate measurements with delay are input, and the lateral velocity after delay compensation is accurately estimated. Then proceed to step E.

[0022] Step E involves analyzing the consistent final boundedness of the state estimation error and the neural network weight error based on Lyapunov function theory, clarifying the constraints of sampling interval and delay, verifying the results through simulation platform and real vehicle experiment, dynamically and iteratively optimizing the observer parameters, and performing high-precision continuous estimation of the vehicle's lateral speed.

[0023] Step B further includes:

[0024] Constructing a planar motion dynamics model for the vehicle:

[0025] (1)

[0026] in, , r and r represent the vehicle's longitudinal velocity, lateral velocity, and yaw rate, respectively. , , These are the vehicle's longitudinal acceleration, lateral acceleration, and... This indicates the longitudinal input torque of the vehicle; These are the lateral forces of the front and rear wheels, respectively; This is the distance from the center of mass to the front axle; For vehicle quality; This refers to the longitudinal air drag coefficient; This refers to the lateral air drag coefficient; For effective longitudinal rotational inertia; This is the distance from the center of mass to the rear axle; The yaw moment of inertia of the vehicle;

[0027] The Brush tire model is integrated to describe lateral forces, where the front and rear wheel lateral forces are calculated based on the Brush tire model, taking into account tire slip characteristics.

[0028] (2)

[0029] in, These are the lateral stiffness and lateral angle of the i-th tire, respectively;

[0030] (3)

[0031] (4)

[0032] in, This is the coefficient of friction between the tire and the road surface; The vertical load on the tire;

[0033] Based on formulas (1) to (4), define the system state vector, input vector, and unknown dynamic terms, and construct the state-space model:

[0034] (5)

[0035] in: Represents the state vector. Indicates control input ( (It is the steering angle). The dynamic term is known; For unknown dynamic terms that include modeling uncertainties, such as tire nonlinearity error and disturbances;

[0036] Define state variables: Define measurement variables Assuming longitudinal velocity The measurement of the yaw rate r is affected by aperiodic sampling and delay, resulting in the following state equation:

[0037] (6)

[0038] in, This represents the unknown dynamic characteristics of the vehicle; , It is the drag coefficient. It's about vehicle quality. , It is the moment of inertia. , These are parameters related to wheelbase; It is a control input; This represents the vehicle's state vector at time t; Indicates the measurable system output; This indicates the system's measurable output after a non-periodic delay; Indicates the non-periodic sampling time; It is an aperiodic delay;

[0039] Furthermore, in step C, the process of constructing the neural network observer for the sampled data includes the following steps:

[0040] Construct the dynamic equations for the continuous state observer:

[0041] (7)

[0042] in: This is the state estimation vector; For designing the matrix, A−K1D is the Hurwitz matrix; ∈ Here, m represents the activation function matrix of the RBF neural network, and m is the number of neurons. ∈ For neural network weight estimates; K1∈ The observation gain matrix, ∈ As an auxiliary variable, ∈ It is a symmetric positive definite matrix. =D ∈ η∈ To compensate for output errors;

[0043] Construct a compensation injector for the sampling time. In the interval Within, the compensated output error η dynamically satisfies:

[0044] (8)

[0045] Where, K2∈ The compensation gain matrix is ​​used to predict the rate of change of output error within the sampling interval, where i is an integer greater than or equal to 0;

[0046] At each sampling time The output error η is reset and compensated based on the deviation between the measured output and the estimated output.

[0047] (9)

[0048] in: To estimate the output; The sensor output includes the longitudinal velocity and yaw rate obtained from sampling.

[0049] Neural network weights The update satisfies:

[0050] (10)

[0051] Where κ>0 is a design parameter used to ensure the convergence of the weight estimation;

[0052] Auxiliary variables The dynamic equation is:

[0053] (11)

[0054] The sampling interval is non-periodic, and its upper limit is... Satisfying the low gain condition:

[0055] (12)

[0056] In the formula, , , , These are system parameters.

[0057] The activation function for the RBF neural network is:

[0058] (13)

[0059] in, This represents the input vector of the neural network; This represents the center vector of the i-th radial basis function; The width parameter represents the i-th radial basis function;

[0060] Matrix P satisfies the Lyapunov equation, which is used for the stability proof:

[0061] (14)

[0062] In the formula, Q>0 is a positive definite matrix.

[0063] Furthermore, in step D, for sampling scenarios with delays, an integral compensation module is added to the sampling data neural network observer to handle the error accumulation during the delay period. The process of forming a sampling delay data neural network observer includes the following steps:

[0064] Construct the dynamic equations for the continuous state observer:

[0065] (15)

[0066] in: This is the state estimation vector; For designing the matrix; Here, m represents the activation function matrix of the RBF neural network, and m is the number of neurons. These are the weight estimates for the neural network. The observation gain matrix; As an auxiliary variable; It is a symmetric positive definite matrix; ; To compensate for output errors due to delay;

[0067] Construct a compensation injector, and introduce an integral compensation module into the compensation injector to obtain a delay compensation injector, wherein, for the sampling time... and delay In the interval Internal delay compensation for output error It consists of two parts:

[0068] (16)

[0069] In the formula, This represents the rate of change of the time delay error compensation term; This is the compensation gain matrix, used to predict the rate of change of error under delay conditions; and These represent the estimated output and the actual output of the time delay, respectively. Integers greater than or equal to 0; delay The upper limit is ;

[0070] Delay compensation for output error :

[0071] (17)

[0072] in, This represents the time delay error compensation term; Adjust the compensation intensity according to the design parameters.

[0073] Based on the dynamic equations of the continuous state observer and the delay compensation injector, the state estimate and error characteristics are updated in real time within the sampling interval, specifically including the following sub-steps:

[0074] Integrating equations (15), (16), and (17) yields... The continuous evolution, in which The predicted value depends on the state at the previous time step. Neural network output and delay compensation error State estimation error The covariance matrix of the expression monotonically decays over time, and the integral compensation term suppresses error divergence caused by delay; at each sampling time... Reset based on the deviation between the measured delay output and the estimated delay output :

[0075] (18)

[0076] in: For delayed estimation of output; The measured output of the delay sensor includes the delayed longitudinal velocity and the delayed yaw rate;

[0077] Neural network weights The update law is:

[0078] (19)

[0079] in To design parameters, ensure that the weight estimation converges, so that Precisely approximating unknown dynamics ;

[0080] Auxiliary variables The evolution equation is:

[0081] (20)

[0082] upper limit of sampling interval With delay limit satisfy:

[0083] (twenty one)

[0084] in: , It is a positive number; It is a delay-sampling coupling function to ensure the state estimation error. With weight error Consistency is ultimately bounded.

[0085] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0086] First, the fully automated vehicle motion state estimation method based on sampled data neural network observer of the present invention proposes a unified framework composed of sampled data neural network observer (SDNNO) and sampled delay data neural network observer (SDDNNO), which realizes high-precision continuous estimation of lateral velocity, provides reliable state input for vehicle lateral control, and solves the combined effects of sensor non-periodic sampling, data delay and system unknown modeling uncertainty (such as tire nonlinearity and external disturbance) in the motion state estimation of fully automated vehicles.

[0087] Second, the fully automated vehicle motion state estimation method based on a sampled data neural network observer of the present invention achieves real-time approximation of unknown system dynamics and effective suppression of delay errors by designing an online weight update law and integral compensation structure for a radial basis function (RBF) neural network, ensuring the consistent eventual boundedness (UUB) of the estimation error, significantly improving the estimation accuracy in complex scenarios, and meeting the real-time requirements of real vehicles.

[0088] Third, the fully automated vehicle motion state estimation method based on a sampling data neural network observer of the present invention takes into account the non-periodicity of sensor sampling and data delay characteristics in actual driving, constructs an observer structure that is adapted to both delay-free and delay-prone scenarios, and clarifies the constraints of sampling interval and delay through Lyapunov function theory, thereby improving the applicability of the invention in low-cost sensor configuration and complex traffic environments. Attached Figure Description

[0089] Figure 1 This is a flowchart of the motion state estimation method for fully automated vehicles based on a sampled data neural network observer according to the present invention;

[0090] Figure 2 This is a schematic diagram of the physical model of the vehicle of the present invention;

[0091] Figure 3 This is a schematic diagram of the radial basis function (RBF) neural network structure in a specific embodiment of the present invention;

[0092] Figure 4 This is a schematic diagram illustrating the working principle of the sampling data neural network observer for estimating the motion state of a fully automated vehicle constructed in a specific embodiment of the present invention, wherein... Figure 4 (a) represents the working principle diagram of the SDNNO observer when there is no delay. Figure 4 The path (b) in the diagram represents the working principle of the SDDNNO observer when there is a delay. Detailed Implementation

[0093] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0094] This invention discloses a method for estimating the motion state of a fully automated vehicle based on a sampled data neural network observer. The method includes the following steps:

[0095] Step A: Use onboard sensors to acquire the vehicle's longitudinal speed and yaw rate, and simultaneously collect control input information, including steering angle and drive torque, through the vehicle controller local area network bus.

[0096] Step B: Establish a planar motion dynamics model of the vehicle, integrate the Brush tire model to describe lateral forces, and derive the state equations containing unknown modeling uncertainties;

[0097] Step C: For the no-delay sampling scenario, construct a sampling data neural network observer, use a radial basis function neural network to approximate the unknown dynamics of the system, design an online weight update law to achieve adaptive adjustment of neural network parameters, combine a compensation injector to offset the loss of sampling information, and make a preliminary estimate of the lateral velocity based on the collected longitudinal velocity and yaw rate measurements, and then proceed to step E.

[0098] Step D: For sampling scenarios with delay, an integral compensation module is added to the sampling data neural network observer to handle the error accumulation during the delay period, forming a sampling delay data neural network observer. The estimation logic and neural network module of the delay-free scenario are reused. The longitudinal velocity and yaw rate measurements with delay are input, and the lateral velocity after delay compensation is accurately estimated. Then proceed to step E.

[0099] Step E involves analyzing the consistent final boundedness of the state estimation error and the neural network weight error based on Lyapunov function theory, clarifying the constraints of sampling interval and delay, verifying the results through simulation platform and real vehicle experiment, dynamically and iteratively optimizing the observer parameters, and performing high-precision continuous estimation of the vehicle's lateral speed.

[0100] Considering existing vehicle motion state estimation technologies, and addressing the issues of insufficient lateral velocity estimation accuracy due to sensor non-periodic sampling, data delay, and uncertainties in system modeling (such as tire nonlinearity and external disturbances) under current technological conditions, as well as the difficulty of balancing real-time performance and adaptability in existing methods, this invention provides a sampling data neural network observer for motion state estimation of fully autonomous vehicles. This observer enables high-precision continuous estimation of vehicle lateral velocity affected by the aforementioned factors, thereby improving the lateral control performance and safety monitoring capabilities of fully autonomous vehicles and fully leveraging the application value of low-cost sensors in autonomous driving.

[0101] As shown in Figure 1, the fully automated vehicle motion state estimation method based on a sampled data neural network observer of the present invention includes:

[0102] S1: Obtain the vehicle's longitudinal speed using low-cost onboard sensors (longitudinal speed sensor, yaw rate sensor). The yaw rate r (including non-periodic sampling interval and data delay characteristics) is used to simultaneously acquire the steering angle via the vehicle controller area network (CAN) bus. Drive torque Control input information, etc.

[0103] S2: Establish a planar motion dynamics model of the vehicle and derive the state equations containing unknown modeling uncertainties.

[0104] based on Figure 2 The vehicle's physical model constructs the vehicle's lateral dynamics equations, where the lateral and yaw dynamics of the vehicle's planar motion satisfy the following:

[0105] (1);

[0106] in, , r and r represent the vehicle's longitudinal velocity, lateral velocity, and yaw rate, respectively. , , These are the vehicle's longitudinal acceleration, lateral acceleration, and... This indicates the longitudinal input torque of the vehicle; These are the lateral forces of the front and rear wheels, respectively; This is the distance from the center of mass to the front axle; For vehicle quality; This refers to the longitudinal air drag coefficient; This refers to the lateral air drag coefficient; For effective longitudinal rotational inertia; This is the distance from the center of mass to the rear axle; The yaw moment of inertia of the vehicle;

[0107] Considering tire slip characteristics, the lateral forces of the front and rear wheels are calculated based on the Brush tire model:

[0108] (2);

[0109] in, These are the lateral stiffness and lateral angle of the i-th tire, respectively;

[0110] (Simplified to a scalar, lateral force dominates)(3);

[0111] (4);

[0112] in, This is the coefficient of friction between the tire and the road surface; The vertical load on the tire;

[0113] The state-space model is established by defining the system state vector, input vector, and unknown dynamic terms according to formulas (1) to (4), and constructing the state-space model:

[0114] (5);

[0115] in:

[0116] State vector Control input Known dynamic terms Composed of the modelable part in formula (1); unknown dynamic term The modeling uncertainties (such as tire nonlinear errors and disturbances) need to be approximated by RBF neural networks.

[0117] Define the state variables and measurement variables separately; where the state variable (lateral velocity estimation) is: (Core states to be estimated: lateral velocity, yaw rate); Measured variables are: (Can be directly measured: longitudinal velocity, yaw rate, including sampling / delay characteristics).

[0118] Assuming longitudinal velocity The measurement of the yaw rate r is affected by aperiodic sampling and delay, resulting in the following state equation:

[0119] (6);

[0120] in, This represents the unknown dynamic characteristics of the vehicle; , It is the drag coefficient. It's about vehicle quality. , It is the moment of inertia. , These are parameters related to wheelbase; It is a control input; This represents the vehicle's state vector at time t; Indicates the measurable system output; This indicates the system's measurable output after a non-periodic delay; Indicates the non-periodic sampling time; It is an aperiodic delay.

[0121] S3: For scenarios with no-delay sampling, a sampled data neural network observer (SDNNO) is constructed, using a radial basis function (RBF) neural network to approximate the unknown dynamics of the system, such as... Figure 3 As shown, an online weight update law is designed ( The neural network parameters are adaptively adjusted, and the loss of sampling information is offset by a compensation injector. Based on the longitudinal speed and yaw rate measurements of the vehicle collected in step S1, the lateral speed is initially estimated by SDNNO.

[0122] The working principle of the Sampled Data Neural Network Observer (SDNNO) used in the vehicle lateral velocity estimation method that does not consider sampling delay but takes into account unknown modeling uncertainties is as follows: Figure 4 The path (a) in the diagram is shown. The sampled data neural network observer consists of a continuous-state observer and a compensation injector. The dynamic equation of the continuous-state observer is as follows:

[0123] (7);

[0124] in:

[0125] This is the state estimation vector; To design the matrix, A−K1D must be a Hurwitz matrix (to ensure stability). ∈ is the activation function matrix (Gaussian kernel function, m is the number of neurons) of the RBF neural network. ∈ For neural network weight estimates; K1∈ The observation gain matrix, ∈ As an auxiliary variable, ∈ It is a symmetric positive definite matrix. =D ∈ η∈ To compensate for output errors.

[0126] Construct a compensation injector for the sampling time. In the interval Within, the compensated output error η dynamically satisfies:

[0127] (8);

[0128] Where, K2∈ The compensation gain matrix is ​​used to predict the rate of change of output error within the sampling interval, where i is an integer greater than or equal to 0.

[0129] Sampling Time Error Reset (Measurement Update): At each sampling time ti, the output error η is reset and compensated based on the deviation between the measured output and the estimated output.

[0130] (9);

[0131] in: To estimate the output; The sensor output is the actual measured output (including the longitudinal velocity and yaw rate obtained from sampling).

[0132] Based on the designed sampling data neural network observer (SDNNO) (Equations (7) to (9)), the unknown dynamics are approximated in real time, and the neural network weights are... The update satisfies:

[0133] (10);

[0134] Where κ>0 is a design parameter used to ensure the convergence of the weight estimation.

[0135] because It is a Hurwitz matrix and Since θ is bounded, it is globally bounded. The dynamic equation of the auxiliary variable θ is:

[0136] (11);

[0137] The sampling interval is aperiodic, and its upper limit τM must satisfy the small gain condition. , , , (System parameters)

[0138] (12);

[0139] The activation function for the RBF neural network is:

[0140] (13);

[0141] Matrix P satisfies the Lyapunov equation, used for stability proof (Q>0 is a positive definite matrix):

[0142] (14).

[0143] S4: For sampling scenarios with delays, an integral compensation module is added to the sampling data neural network observer in step S3 to handle the error accumulation during the delay period, forming a Sampling Delay Data Neural Network Observer (SDDNNO). The delay-free scene estimation logic obtained in S3 is reused with the neural network module, and the delayed longitudinal velocity and yaw rate measurements are input. The SDDNNO then achieves accurate estimation of the delayed lateral velocity. The working principle of the Sampling Delay Data Neural Network Observer (SDDNNO) is as follows: Figure 4 As shown in path (b) in the diagram. The Sample Delayed Data Neural Network Observer (SDDNNO) consists of a continuous-state observer dynamic equation and a delay compensation injector. The continuous-state observer dynamic equation it uses is:

[0144] (15);

[0145] in: This is the state estimation vector; For designing the matrix; Here is the activation function matrix of the RBF neural network (m is the number of neurons). These are the weight estimates for the neural network. The observation gain matrix; As an auxiliary variable; It is a symmetric positive definite matrix; ; To compensate for output errors due to delay.

[0146] Construct a compensation injector, and introduce an integral compensation module into the compensation injector to obtain a delay compensation injector, wherein, for the sampling time... and delay (upper limit is) ), in the interval Internal delay compensation for output error It consists of two parts:

[0147] (16)

[0148] In the formula, This represents the rate of change of the time delay error compensation term; This is the compensation gain matrix, used to predict the rate of change of error under delay conditions; and These represent the estimated output and the actual output of the time delay, respectively. Integers greater than or equal to 0; delay The upper limit is ;

[0149] Delay compensation for output error :

[0150] (17)

[0151] in, This represents the time delay error compensation term; Adjust the compensation intensity according to the design parameters.

[0152] Based on the dynamic equations of the continuous state observer and the delay-compensated injector, the state estimate and error characteristics are updated in real time within the sampling interval:

[0153] State estimation: obtained by integrating equations (15), (16), and (17) The continuous evolution, in which The predicted value depends on the state at the previous time step. Neural network output and delay compensation error .

[0154] Implicit constraints on error covariance: due to The Hurwitz matrix represents the state estimation error. The covariance matrix of the equation decays monotonically over time, and the integral compensation term suppresses error divergence caused by delay.

[0155] Measurement update: Sampling time delay error reset at each sampling time. Reset based on the deviation between the measured delay output and the estimated delay output :

[0156] (18);

[0157] in: For delayed estimation of output; The actual output of the delay sensor (including delay longitudinal velocity and delay yaw rate).

[0158] To approximate the unknown dynamics of the system in real time, neural network weights The update law is consistent with SDNNO:

[0159] (19);

[0160] in To design parameters, ensure that the weight estimation converges, so that Precisely approximating unknown dynamics .

[0161] Auxiliary variables The evolution equation is consistent with that of SDNNO:

[0162] (20);

[0163] S5: Based on Lyapunov function theory, analyze the consistent final boundedness (UUB) of state estimation error and neural network weight error, and clarify the constraints on sampling interval and delay; stability constraints, and upper limit of sampling interval. With delay limit Must meet:

[0164] (twenty one);

[0165] in: ( (positive number) This is a delay-sampling coupling function; this condition guarantees the state estimation error. With weight error Uniformly eventually bounded (UUB).

[0166] The estimated lateral velocity at the current moment is used as the reference input for the state observation at the next moment. The observer parameters are dynamically optimized in cycles S3 and S4 to achieve high-precision continuous estimation of the vehicle's lateral velocity.

[0167] Through the architecture of the embodiment, namely the collaborative design of the Sampled Data Neural Network Observer (SDNNO) and the Sampled Delayed Data Neural Network Observer (SDDNNO), high-precision continuous estimation of the lateral speed of fully automated vehicles is achieved, which is affected by the non-periodic sampling of sensors, data delay, and unknown modeling uncertainties (such as tire nonlinearity and external disturbances). This avoids large deviations in the lateral speed estimate under the influence of the above factors, which is of great significance for improving the lateral control performance and safety monitoring capabilities of automated vehicles.

[0168] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0169] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for estimating the motion state of a fully automated vehicle based on a sampled data neural network observer, characterized in that, The method includes the following steps: Step A: Use onboard sensors to acquire the vehicle's longitudinal speed and yaw rate, and simultaneously collect control input information, including steering angle and drive torque, through the vehicle controller local area network bus. Step B: Establish a planar motion dynamics model of the vehicle, integrate the Brush tire model to describe lateral forces, and derive the state equations containing unknown modeling uncertainties; Step C: For the no-delay sampling scenario, construct a sampling data neural network observer, use a radial basis function neural network to approximate the unknown dynamics of the system, design an online weight update law to achieve adaptive adjustment of neural network parameters, combine a compensation injector to offset the loss of sampling information, and make a preliminary estimate of the lateral velocity based on the collected longitudinal velocity and yaw rate measurements, and then proceed to step E. Step D: For sampling scenarios with delay, an integral compensation module is added to the sampling data neural network observer to handle the error accumulation during the delay period, forming a sampling delay data neural network observer. The estimation logic and neural network module of the delay-free scenario are reused. The longitudinal velocity and yaw rate measurements with delay are input, and the lateral velocity after delay compensation is accurately estimated. Then proceed to step E. Step E involves analyzing the consistent final boundedness of the state estimation error and the neural network weight error based on Lyapunov function theory, clarifying the constraints of sampling interval and delay, verifying the results through simulation platform and real vehicle experiment, dynamically and iteratively optimizing the observer parameters, and performing high-precision continuous estimation of the vehicle's lateral speed.

2. The method for estimating the motion state of a fully automated vehicle based on a sampled data neural network observer according to claim 1, characterized in that, Step B further includes: Constructing a planar motion dynamics model for the vehicle: (1)； in, , r and r represent the vehicle's longitudinal velocity, lateral velocity, and yaw rate, respectively. , , These are the vehicle's longitudinal acceleration, lateral acceleration, and... This indicates the longitudinal input torque of the vehicle; These are the lateral forces of the front and rear wheels, respectively; This is the distance from the center of mass to the front axle; For vehicle quality; This refers to the longitudinal air drag coefficient; This refers to the lateral air drag coefficient; For effective longitudinal rotational inertia; This is the distance from the center of mass to the rear axle; The yaw moment of inertia of the vehicle; Considering tire slip characteristics, the lateral forces of the front and rear wheels are calculated based on the Brush tire model. : (2)； in, These are the lateral stiffness and lateral angle of the i-th tire, respectively; (3)； (4)； in, This is the coefficient of friction between the tire and the road surface; The vertical load on the tire; Based on formulas (1) to (4), define the system state vector, input vector, and unknown dynamic terms, and construct the state-space model: (5)； in: Represents the state vector. Indicates control input ( (It is the steering angle). The dynamic term is known; For unknown dynamic terms that include modeling uncertainties, such as tire nonlinearity error and disturbances; Define state variables Define measurement variables Assuming longitudinal velocity The measurement of the yaw rate r is affected by aperiodic sampling and delay, resulting in the following state equation: (6)； in, This represents the unknown dynamic characteristics of the vehicle; , It is the drag coefficient. It's about vehicle quality. , It is the moment of inertia. , These are parameters related to wheelbase; It is a control input; This represents the vehicle's state vector at time t; Indicates the measurable system output; This indicates the system's measurable output after a non-periodic delay; Indicates the non-periodic sampling time; It is an aperiodic delay.

3. The method for estimating the motion state of a fully automated vehicle based on a sampled data neural network observer according to claim 1, characterized in that, In step C, the process of constructing the neural network observer for the sampled data includes the following steps: Construct the dynamic equations for the continuous state observer: (7)； in: This is the state estimation vector; For designing the matrix, A−K1D is the Hurwitz matrix; ∈ Here, m represents the activation function matrix of the RBF neural network, and m is the number of neurons. ∈ For neural network weight estimates; K1∈ The observation gain matrix, ∈ As an auxiliary variable, ∈ It is a symmetric positive definite matrix. =D ∈ η∈ To compensate for output errors; Construct a compensation injector for the sampling time. In the interval Within, the compensated output error η dynamically satisfies: (8)； Where, K2∈ The compensation gain matrix is ​​used to predict the rate of change of output error within the sampling interval; i is an integer greater than or equal to 0. At each sampling time The output error η is reset and compensated based on the deviation between the measured output and the estimated output. (9)； in: To estimate the output; The sensor output includes the longitudinal velocity and yaw rate obtained from sampling. Neural network weights The update satisfies: (10)； Where κ>0 is a design parameter used to ensure the convergence of the weight estimation; Auxiliary variables The dynamic equation is: (11)； The aperiodic sampling interval between adjacent sampling times is denoted as Its upper limit is It satisfies the small gain condition: (12)； In the formula, , , , For system parameters; The activation function for the RBF neural network is: (13)； in, This represents the input vector of the neural network; This represents the center vector of the i-th radial basis function; The width parameter represents the i-th radial basis function; Matrix P satisfies the Lyapunov equation, which is used for the stability proof: (14)； In the formula, Q>0 is a positive definite matrix.

4. The method for estimating the motion state of a fully automated vehicle based on a sampled data neural network observer according to claim 1, characterized in that, In step D, for sampling scenarios with delays, an integral compensation module is added to the sampling data neural network observer to handle the error accumulation during the delay period. The process of forming a sampling delay data neural network observer includes the following steps: Construct the dynamic equations for the continuous state observer: (15)； in: This is the state estimation vector; For designing the matrix; Here, m represents the activation function matrix of the RBF neural network, and m is the number of neurons. These are the weight estimates for the neural network. The observation gain matrix; As an auxiliary variable; It is a symmetric positive definite matrix; ; To compensate for output errors due to delay; Construct a compensation injector, and introduce an integral compensation module into the compensation injector to obtain a delay compensation injector, wherein, for the sampling time... and delay In the interval Internal delay compensation for output error It consists of two parts: (16)； In the formula, This represents the rate of change of the time delay error compensation term; This is the compensation gain matrix, used to predict the rate of change of error under delay conditions; and These represent the estimated output and the actual output of the time delay, respectively. Integers greater than or equal to 0; delay The upper limit is ; Delay compensation for output error : (17)； in, This represents the time delay error compensation term; Adjust the compensation intensity according to the design parameters.

5. Based on the dynamic equations of the continuous state observer and the delay compensation injector, the state estimate and error characteristics are updated in real time within the sampling interval, specifically including the following sub-steps: Integrating equations (15), (16), and (17) yields... The continuous evolution, in which The predicted value depends on the state at the previous time step. Neural network output and delay compensation error State estimation error The covariance matrix of the expression monotonically decays over time, and the integral compensation term suppresses error divergence caused by delay; at each sampling time... Reset based on the deviation between the measured delay output and the estimated delay output : (18)； in: For delayed estimation of output; The measured output of the delay sensor includes the delayed longitudinal velocity and the delayed yaw rate; Neural network weights The update law is: (19)； in To design parameters, ensure that the weight estimation converges, so that Precisely approximating unknown dynamics ; Auxiliary variables The evolution equation is: (20)； upper limit of sampling interval With delay limit satisfy: (21)； in: , It is a positive number; It is a delay-sampling coupling function to ensure the state estimation error. With weight error Consistency is ultimately bounded.

Images