Vehicle dynamic parameter estimation method in combination with long short-term memory neural network and unscented Kalman filter

Abstract

The invention discloses a vehicle dynamic parameter estimation method combining a long short-term memory neural network and unscented Kalman filtering, which comprises the following steps of: firstly, establishing an implicit mapping function between measurable parameters of a vehicle-mounted sensor and a longitudinal vehicle speed and a side slip angle based on a seven-degree-of-freedom vehicle nonlinear dynamic model; preliminarily determining input and output vectors of the long and short-term memory neural network (data driving component); secondly, unscented Kalman filtering (a physical model component) is established based on vehicle dynamic characteristics, and reasonability and physical consistency of a process model, a measurement equation and control input are ensured; then, combining the output of the data driving component with the state prediction and updating mechanism of the physical model component, and determining the input and output relationship of each component; and finally, performing staged training (pre-training and end-to-end training), and completing algorithm deployment and testing. According to the method, the neural network and the Kalman filtering are combined, collaborative optimization of the neural network and the Kalman filtering is achieved through end-to-end training, and the method has great significance in improving the estimation precision of the longitudinal vehicle speed and the side slip angle.

Description

Technical Field

[0001] The invention belongs to the field of automobile control, and in particular relates to a vehicle dynamic parameter estimation method combining a long short-term memory neural network with an unscented Kalman filter. Background Art

[0002] The feedback control of the vehicle's active safety system has high accuracy requirements for the vehicle's motion state information. In particular, the two key parameters of longitudinal speed and center of mass sideslip angle have a significant impact on the stability and handling of the vehicle. However, these two parameters are usually not directly measured by sensors equipped on mass-produced vehicles. For this reason, an estimation algorithm based on measurable parameters is generally used in combination with the vehicle's kinematic or dynamic model to estimate these two key parameters in real time.

[0003] Traditional estimation algorithms can be divided into two categories: model-based methods and data-driven methods. Model-based estimation methods usually rely on kinematic models or vehicle dynamic models, and achieve parameter estimation by building a state observer. Kinematic model-based estimation methods use the kinematic relationship between the state parameters to be estimated and the known state parameters to calculate the estimated value of the unknown motion state through integration. However, since the inertial measurement unit signal usually has noise and errors, long-term open-loop integration often leads to error accumulation, resulting in a large deviation between the estimated result and the true value. In contrast, the dynamic model-based estimation method relies on the dynamic model of the vehicle and tire, and is usually combined with Kalman filtering or state observers for state estimation. Although this method shows strong robustness in the face of uncertainty in vehicle and tire parameters, its estimation accuracy is relatively low.

[0004] Unlike model-based estimation methods, data-driven methods analyze sensor measurement data and perform parameter estimation based on the mapping relationship between known state quantities and control quantities and unknown state quantities. Neural networks are currently the most widely used tool in data-driven methods. Data-driven methods can make full use of historical data for training and can provide higher estimation accuracy. However, the main problem with this type of method is that it ignores the explicit constraints of the physical model, resulting in poor robustness and the resulting model often lacks interpretability. Therefore, in practical applications, the reliability and accuracy of pure data-driven methods are still limited.

[0005] In recent years, hybrid methods based on models and data-driven have gradually been proposed to combine the advantages of both. These hybrid methods can be roughly divided into two categories: model-neural network methods and neural network-model methods. In the model-neural network method, the kinematic model provides the neural network with pseudo-measurements containing significant errors, biases and drifts, which improves the estimation accuracy of the neural network. In the neural network-model method, the neural network provides pseudo-measurements for the Kalman filter, which also improves the estimation accuracy. Although some studies have attempted to combine model-based and data-driven estimation algorithms, most of the work simply processes the two in series, failing to effectively consider the impact of the accuracy of the process model, resulting in the inability to reasonably adjust the distrust of the estimation results. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a vehicle dynamic parameter estimation method combining a long short-term memory neural network with an unscented Kalman filter, thereby solving the reliability problem of the vehicle active safety system in the prior art.

[0007] The present invention adopts the following technical solutions to solve the above technical problems:

[0008] The vehicle dynamic parameter estimation method combining long short-term memory neural network and unscented Kalman filter includes the following steps:

[0009] S1. Based on the nonlinear dynamics model of the seven-degree-of-freedom vehicle, an implicit mapping function between the measurable parameters of the vehicle-mounted sensors and the longitudinal vehicle speed and the sideslip angle of the center of mass is established, and the long short-term memory neural network is preliminarily established as the input and output vectors of the data-driven component;

[0010] S2, establishing an unscented Kalman filter based on vehicle dynamics characteristics as a physical model component;

[0011] S3, based on the joint training requirements of S1 data-driven components and S2 physical model components, the output of the data-driven components is combined with the state prediction and update mechanism of the physical model components to clarify the input-output relationship of each component;

[0012] S4: Based on the input-output relationship of each component in S3, perform phased training, complete algorithm deployment and testing, and achieve high-precision estimation of the vehicle's longitudinal speed and center of mass sideslip angle.

[0013] Step S1 includes the following sub-steps:

[0014] S101. According to Newton's second law, a seven-degree-of-freedom vehicle dynamics model is established, including three degrees of freedom of longitudinal, lateral and yaw motion of the vehicle and four degrees of freedom of wheel rotation;

[0015] S102, based on the requirement of the vehicle dynamics model in S101 for obtaining the lateral longitudinal tire force, establish a nonlinear dynamics equation to analyze the mechanical characteristics of the tire;

[0016] S103. Based on the nonlinear dynamic equations established in S101 and S102, the mapping relationship between the measurable parameters of the vehicle-mounted sensors and the parameters to be estimated is obtained, and the input and output vectors of the long short-term memory neural network are preliminarily determined.

[0017] Step S2 includes the following sub-steps:

[0018] S201. Establish a state prediction process model based on the vehicle dynamics characteristics, including longitudinal vehicle speed, lateral vehicle speed and yaw rate, to ensure that the process model can accurately describe the vehicle's motion behavior;

[0019] S202. Based on the measurable signals of the vehicle-mounted sensors, a measurement equation is designed to clarify the relationship between the vehicle state and the measurable sensor data, and to ensure that the measurement equation can effectively map the sensor data to the actual state of the vehicle;

[0020] S203. Build an unscented Kalman filter based on the process model, measurement equations and control inputs in S201 and S202.

[0021] In step S202, the measurement equation and its control input are:

[0022]

[0023]

[0024] In the formula, , , are the measured values ​​of the vehicle's longitudinal and lateral accelerations and yaw angular velocity, respectively. are the front left, front right, rear left and rear right tire longitudinal forces, are the lateral forces of the front left, front right, rear left and rear right tires respectively, is the front wheel turning angle.

[0025] In step S203, the unscented Kalman filter includes two steps: time update and measurement update. The time update includes two parts: future state prediction and future error covariance prediction; the measurement update includes three parts: calculating the Kalman gain, updating the state estimate, and updating the error covariance.

[0026] Step S3 is specifically as follows:

[0027] Based on the future state prediction and measurement update mechanism of the physical component, the form of combining the output of the data-driven component with the physical component is defined, wherein the input and output of the data-driven component are defined as:

[0028]

[0029] The measurement equations for the physical model components are:

[0030]

[0031] In the formula are pseudo-measured values ​​of longitudinal and lateral vehicle speeds, respectively.

[0032] Step S4 includes the following sub-steps:

[0033] S401. Based on the input-output relationship specified in S3, design and implement a phased training strategy, firstly perform a pre-training phase to ensure that the output distribution of the neural network is within a physically meaningful range;

[0034] S402, based on the neural network pre-trained in S401, end-to-end training is performed in combination with Kalman filtering to optimize the synergy between physical components and data-driven components;

[0035] S403. Based on the neural network obtained through end-to-end training in S402, combined with unscented Kalman filtering, the estimation algorithm is deployed and tested.

[0036] In step S401, the loss function of the pre-training stage is set to:

[0037]

[0038] In the formula is the error between the process noise covariance of the three state quantities output by the data-driven component and the empirical value, is the Gaussian negative log-likelihood loss function, which is in the form of:

[0039]

[0040] In the formula are the process noise covariance empirical values ​​of longitudinal and lateral vehicle speeds and yaw angular velocity, respectively. is the amount of data in a batch during training, A small non-negative constant value to prevent the logarithmic function from being meaningless.

[0041] In step S402, the loss function of end-to-end training is set to:

[0042]

[0043] In the formula are the longitudinal and lateral vehicle speeds output by the unscented Kalman filter, respectively.

[0044] A computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, call all or part of the steps of the method.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The present invention first establishes an implicit mapping function between the measurable signal of the vehicle-mounted sensor and the longitudinal vehicle speed and the sideslip angle of the center of mass, and preliminarily clarifies the input and output vectors of the long short-term memory neural network. Then, an unscented Kalman filter for estimating the longitudinal vehicle speed and the sideslip angle of the center of mass is constructed to ensure the rationality and physical consistency of the process model, measurement equations and control inputs. Subsequently, the input and output relationship between the neural network and the Kalman filter is analyzed and finally determined according to the joint training requirements. Finally, based on a hybrid method combining physical models and data-driven, phased training is performed, and model deployment and testing are completed in a joint simulation environment to achieve high-precision estimation of the longitudinal vehicle speed and the sideslip angle of the center of mass.

[0047] 2. Through verification, the hybrid estimation method for longitudinal vehicle speed and sideslip angle of center of mass is reasonable, feasible and effective, and has great significance for improving the estimation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The block diagram of the vehicle dynamic parameter estimation method combining long short-term memory neural network and unscented Kalman filter of the present invention.

[0049] Figure 2 Schematic diagram of a nonlinear vehicle dynamics model in an embodiment of the present invention.

[0050] Figure 3 This is the joint simulation test result of the hybrid estimation method in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0052] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to design and develop a vehicle dynamic parameter estimation method that combines a long short-term memory neural network with an unscented Kalman filter. The long short-term memory neural network is used to capture the temporal relationship between the historical data of the dynamic system, and the unscented Kalman filter with third-order accuracy is used to effectively handle the high nonlinearity of the vehicle and tire dynamics model. In combination with end-to-end training, the neural network learns the accuracy of the Kalman filter, providing an optimal degree of distrust to achieve high-precision estimation of the vehicle dynamic parameters, thereby improving the reliability of the vehicle's active safety system.

[0053] Specifically, a vehicle dynamic parameter estimation method combining a long short-term memory neural network and an unscented Kalman filter includes the following steps:

[0054] S1. Based on the nonlinear dynamics model of the seven-degree-of-freedom vehicle, an implicit mapping function between the measurable parameters of the vehicle-mounted sensors and the longitudinal vehicle speed and the sideslip angle of the center of mass is established, and the long short-term memory neural network is preliminarily established as the input and output vectors of the data-driven component;

[0055] S2, establishing an unscented Kalman filter based on vehicle dynamics characteristics as a physical model component;

[0056] S3, based on the joint training requirements of S1 data-driven components and S2 physical model components, the output of the data-driven components is combined with the state prediction and update mechanism of the physical model components to clarify the input-output relationship of each component;

[0057] S4: Based on the input-output relationship of each component in S3, perform phased training, complete algorithm deployment and testing, and achieve high-precision estimation of the vehicle's longitudinal speed and center of mass sideslip angle.

[0058] Specific embodiments, such as Figures 1 to 3 As shown,

[0059] The present invention provides a vehicle dynamic parameter estimation method combining a long short-term memory neural network and an unscented Kalman filter, comprising the following steps:

[0060] Step S1: Based on the nonlinear dynamics model of the seven-degree-of-freedom vehicle, an implicit mapping function between the measurable control quantity, state quantity parameters and the longitudinal vehicle speed and the sideslip angle of the center of mass is established, and the input and output vectors of the long short-term memory neural network (data-driven component) are preliminarily established; specifically,

[0061] 11) According to Newton's second law, a seven-degree-of-freedom vehicle dynamics model is established, including the three plane motion degrees of freedom of the vehicle longitudinal, lateral and yaw, and the four wheel rotation degrees of freedom, such as Figure 2 As shown;

[0062] 12) Based on the vehicle dynamics model’s need to obtain the lateral longitudinal tire force, a magic formula (i.e., nonlinear dynamic equation) is established to analyze the mechanical properties of the tire;

[0063] 13) Based on the established nonlinear dynamic equations, the mapping relationship between the measurable parameters of the vehicle-mounted sensors and the parameters to be estimated is clarified, and the input and output vectors of the long short-term memory neural network are preliminarily clarified.

[0064] Step S2: establishing an unscented Kalman filter (physical model component) based on vehicle dynamics characteristics to ensure the rationality and physical consistency of the process model, measurement equations, and control inputs;

[0065] 21) According to the vehicle dynamics characteristics, establish a state prediction model (process model), including longitudinal speed, lateral speed and yaw rate, to ensure that the process model can accurately describe the vehicle's motion behavior;

[0066] 22) Based on the measurable signals of the on-board sensors, design measurement equations to clarify the relationship between the vehicle state (longitudinal speed, center of mass sideslip angle) and the measurable sensor data (such as vehicle speed sensor, inertial measurement unit, steering wheel angle, etc.), and ensure that the measurement equations can effectively map the sensor data to the actual state of the vehicle;

[0067] 23) Construct an unscented Kalman filter based on the process model, measurement equations and control inputs.

[0068] Step S3: Based on the joint training requirements of step 1) the data-driven component and step 2) the physical model component, the output of the data-driven component is combined with the state prediction and update mechanism of the physical model component to clarify the input-output relationship of each component;

[0069] 31) Based on the state prediction and update mechanism of the physical component in step 2), define the form of combining the output of the data-driven component in step 1) with the physical component, so as to effectively integrate the advantages of the physical model and the data-driven model.

[0070] Step S4: Based on the input-output relationship of each component in step 3), perform phased training (pre-training and end-to-end training), and complete algorithm deployment and testing to achieve high-precision estimation of the vehicle longitudinal speed and center of mass sideslip angle;

[0071] 41) Based on the input-output relationship defined in 3), design and implement a phased training strategy, starting with a pre-training phase to ensure that the neural network output is within a physically meaningful range;

[0072] 42) Based on the pre-trained neural network, end-to-end training is combined with Kalman filtering to optimize the synergy between physical components and data-driven components;

[0073] 43) Based on the neural network obtained through end-to-end training and combined with the unscented Kalman filter, the estimation algorithm is deployed and tested.

[0074] In this embodiment, the specific steps of achieving high-precision estimation of the longitudinal speed and the sideslip angle of the vehicle according to the above method are:

[0075] Step 1: Use the seven-degree-of-freedom nonlinear vehicle dynamics model ( Figure 2 ) Determine the mapping relationship between the measurable state of the vehicle sensor and the state to be measured. According to Newton's second law, the dynamic equation of the vehicle's plane motion can be expressed as:

[0076]

[0077] In the formula , , , and They are vehicle mass, distance from center of mass to front and rear axles, vehicle moment of inertia and wheelbase.

[0078] Furthermore, in order to more accurately describe the vehicle's motion state, the rotational freedom of the four wheels is also considered, and its rotational dynamics equation can be expressed as:

[0079]

[0080] In the formula, is the wheel moment of inertia, is the effective turning radius of the wheel.

[0081] Furthermore, in order to obtain the mapping relationship between tire force and measurable parameters, the tire mechanical characteristics are analyzed through magic formulas. The tire longitudinal force and lateral force can be expressed by a set of formulas:

[0082]

[0083] In the formula, is the vertical offset, is the stiffness factor, is the shape factor, is the peak value, is the vertical load, is the curvature factor, is the independent variable, the longitudinal force , in the lateral force, , is the horizontal offset, and They are the tire side slip angle and longitudinal slip rate, which are calculated as follows:

[0084]

[0085]

[0086] The vertical load can be expressed as:

[0087]

[0088] In the formula, is the acceleration due to gravity, is the vehicle wheelbase, Represents the centroid height.

[0089] Furthermore, based on the above-mentioned seven-degree-of-freedom vehicle dynamics model, wheel rotation dynamics model and tire dynamics model, a mapping relationship between parameters is established, and the input and output vectors of the long short-term memory neural network are:

[0090]

[0091] Step 2, the unscented Kalman filter part, includes the description of the process model, the establishment and prediction of the measurement equation, and the update steps, such as Figure 1 shown.

[0092] The process equations are the same as those used in step 1 to describe the longitudinal, lateral, and yaw plane motions of the vehicle. The measurement equations and control inputs are:

[0093]

[0094]

[0095] The time update consists of two parts: predicting the future state and predicting the future error covariance.

[0096]

[0097] The measurement update consists of three parts: calculating the Kalman gain, updating the estimate, and updating the error covariance.

[0098]

[0099] In the formula are the covariance of the state value and the measurement value and the covariance of the measurement value, respectively, and are calculated as follows:

[0100]

[0101] Step 3: Based on the data-driven component in step 1 and the physical model component in step 2, combined with the state prediction and measurement update mechanism of Kalman filtering, define the input-output relationship of the hybrid estimation algorithm.

[0102] In the process of combining the two, the neural network can learn the accuracy of Kalman filtering. In addition to the pseudo-measurement values ​​of longitudinal and lateral vehicle speeds and their measurement noise covariance, the output of the neural network can also include three process noise covariances in the process model, thereby generating adaptability to different working conditions and improving estimation accuracy. That is, the input and output of the long short-term memory neural network are modified to:

[0103]

[0104] The estimated value of the neural network has higher accuracy than that of the Kalman filter. Using its output as the pseudo-measurement value of the Kalman filter can greatly improve its estimation accuracy. That is, the measurement equation of the unscented Kalman filter is modified to:

[0105]

[0106] Step 4: Based on the input-output relationship of each component specified in step 3, perform phased training, and complete algorithm deployment and testing to achieve high-precision estimation of the vehicle's longitudinal speed and center of mass sideslip angle.

[0107] The pre-training stage can ensure that the output distribution of the neural network is within a physically meaningful range, and its loss function is set as:

[0108]

[0109] In the formula is the error between the process noise covariance of the three state quantities output by the data-driven component and the empirical value, is the Gaussian negative log-likelihood loss function, which is in the form of:

[0110]

[0111] Furthermore, the pre-trained neural network and Kalman filter are combined for end-to-end training to optimize the synergy between physical components and data-driven components. The loss function is set as:

[0112]

[0113] The hybrid estimation algorithm is deployed to the Simulink / CarSim co-simulation platform to verify its generalization ability. Figure 3 It is shown that the pure data-driven algorithm shows high estimation accuracy under normal working conditions, but when the motor torque fluctuates greatly, its estimated value jitters more seriously. The estimation algorithm based on the physical model shows strong robustness, but the estimation accuracy is low. In contrast, the robustness and longitudinal vehicle speed and center of mass sideslip angle estimation accuracy of the hybrid estimation algorithm are significantly higher than those of the pure data-driven algorithm and the algorithm based on the physical model.

[0114] It can be seen from the above test results that the present invention can effectively achieve high-precision estimation of longitudinal vehicle speed and center of mass sideslip angle.

[0115] A computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, call all or part of the steps of the method.

[0116] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0117] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A vehicle dynamic parameter estimation method combining long short-term memory neural network and unscented Kalman filter, characterized by: The steps include: S1. Based on the nonlinear dynamics model of the seven-degree-of-freedom vehicle, an implicit mapping function between the measurable parameters of the vehicle-mounted sensors and the longitudinal vehicle speed and the sideslip angle of the center of mass is established, and the long short-term memory neural network is preliminarily established as the input and output vectors of the data-driven component; S2, establishing an unscented Kalman filter based on vehicle dynamics characteristics as a physical model component; S3, based on the joint training requirements of S1 data-driven components and S2 physical model components, the output of the data-driven components is combined with the state prediction and update mechanism of the physical model components to clarify the input-output relationship of each component; S4: Based on the input-output relationship of each component in S3, perform phased training, complete algorithm deployment and testing, and achieve high-precision estimation of the vehicle's longitudinal speed and center of mass sideslip angle.

2. The vehicle dynamic parameter estimation method combining long short-term memory neural network and unscented Kalman filter according to claim 1 is characterized by: Step S1 includes the following sub-steps: S101. According to Newton's second law, a seven-degree-of-freedom vehicle dynamics model is established, including three degrees of freedom of longitudinal, lateral and yaw motion of the vehicle and four degrees of freedom of wheel rotation; S102, based on the requirement of the vehicle dynamics model in S101 for obtaining the lateral longitudinal tire force, establish a nonlinear dynamics equation to analyze the mechanical characteristics of the tire; S103. Based on the nonlinear dynamic equations established in S101 and S102, the mapping relationship between the measurable parameters of the vehicle-mounted sensors and the parameters to be estimated is obtained, and the input and output vectors of the long short-term memory neural network are preliminarily determined.

3. The vehicle dynamic parameter estimation method combining long short-term memory neural network and unscented Kalman filter according to claim 2 is characterized by: Step S2 includes the following sub-steps: S201. Establish a state prediction process model based on the vehicle dynamics characteristics, including longitudinal vehicle speed, lateral vehicle speed and yaw rate, to ensure that the process model can accurately describe the vehicle's motion behavior; S202. Based on the measurable signals of the vehicle-mounted sensors, design measurement equations to clarify the relationship between the vehicle state and the measurable sensor data, and ensure that the observation equations can effectively map the sensor data to the actual state of the vehicle; S203. Build an unscented Kalman filter based on the process model, measurement equations and control inputs in S201 and S202.

4. The vehicle dynamic parameter estimation method combining long short-term memory neural network and unscented Kalman filter according to claim 3 is characterized by: In step S202, the measurement equation and its control input are: In the formula, , , are the measured values ​​of the vehicle's longitudinal and lateral accelerations and yaw angular velocity, respectively. are the front left, front right, rear left and rear right tire longitudinal forces, are the lateral forces of the front left, front right, rear left and rear right tires respectively, is the front wheel turning angle.

5. The vehicle dynamic parameter estimation method combining long short-term memory neural network and unscented Kalman filter according to claim 4 is characterized by: In step S203, the unscented Kalman filter includes two steps: time update and measurement update. The time update includes two parts: future state prediction and future error covariance prediction; the measurement update includes three parts: calculating the Kalman gain, updating the state estimate, and updating the error covariance.

6. The vehicle dynamic parameter estimation method combining long short-term memory neural network and unscented Kalman filter according to claim 5 is characterized by: Step S3 is specifically as follows: Based on the future state prediction and measurement update mechanism of the physical component, the form of combining the output of the data-driven component with the physical component is defined, wherein the input and output of the data-driven component are defined as: The measurement equations for the physical model components are: In the formula are pseudo-measured values ​​of longitudinal and lateral vehicle speeds, respectively.

7. The vehicle dynamic parameter estimation method combining long short-term memory neural network and unscented Kalman filter according to claim 6 is characterized by: Step S4 includes the following sub-steps: S401. Based on the input-output relationship specified in S3, design and implement a phased training strategy, firstly perform a pre-training phase to ensure that the output distribution of the neural network is within a physically meaningful range; S402, based on the neural network pre-trained in S401, end-to-end training is performed in combination with Kalman filtering to optimize the synergy between physical components and data-driven components; S403. Based on the neural network obtained through end-to-end training in S402, combined with unscented Kalman filtering, the estimation algorithm is deployed and tested.

8. The vehicle dynamic parameter estimation method combining long short-term memory neural network and unscented Kalman filter according to claim 7 is characterized by: In step S401, the loss function of the pre-training stage is set to: In the formula is the error between the process noise covariance of the three state quantities output by the data-driven component and the empirical value, is the Gaussian negative log-likelihood loss function, which is in the form of: In the formula are the process noise covariance empirical values ​​of longitudinal and lateral vehicle speed and yaw rate, respectively. is the amount of data in a batch during training, A small non-negative constant value to prevent the logarithmic function from being meaningless.

9. The vehicle dynamic parameter estimation method combining long short-term memory neural network and unscented Kalman filter according to claim 8, characterized in that: In step S402, the loss function of end-to-end training is set to: In the formula are the longitudinal and lateral vehicle speeds output by the unscented Kalman filter, respectively.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, all or part of the steps of the method described in any one of claims 1 to 9 are called.

Images