A lateral velocity estimation method for a hybrid driving model based on transfer learning

Abstract

The present invention is applicable to the technical field of vehicle parameter estimation, and provides a lateral velocity estimation method of a hybrid drive model based on transfer learning, comprising the following steps: using a kinematic model and a dynamic model to respectively estimate the lateral velocities and, taking them as inputs of an LSTM neural network to form a data-mechanism hybrid drive model, namely, LVENN; using the hybrid drive model to estimate the lateral velocity, the fusion method can significantly reduce the root mean square error and the maximum error of the lateral velocity estimation; by designing a transfer learning algorithm, fine-tuning parameters in a feature extraction layer and retraining in a fitting output layer to increase the transfer learning capability, the present invention can utilize the training data in the source domain to reduce the amount of labeled data in the target domain, and improve the efficiency of model training.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle parameter estimation, and in particular relates to a lateral speed estimation method of a hybrid driving model based on transfer learning. Background Art

[0002] For intelligent vehicles, to achieve precise control of the vehicle, it is necessary to obtain real-time vehicle state parameter information. In the vehicle's stable driving, accurate acquisition of lateral velocity is very important. At present, due to the technical and cost factors of some sensors, large-scale applications cannot be carried out. Therefore, key parameters such as lateral velocity cannot be directly obtained, and the lateral velocity can only be estimated based on relevant estimation algorithms. When using a mechanism model, i.e. a dynamic or kinematic model, to estimate the lateral velocity, drift is likely to occur during vehicle driving, and the estimation error is likely to be large due to inaccurate parameters in the model; on the other hand, the lateral velocity estimation method based on data, i.e. a neural network, is easily limited by the quantity and quality of training data; in addition, in most cases, there is a lack of labeled data in the target domain, and accurate and effective prediction results cannot be obtained. At this time, transfer learning can be used to migrate the model trained in the source domain to the target domain to solve the problem of limited data. For this reason, we propose a lateral velocity estimation method for a hybrid drive model based on transfer learning. Summary of the invention

[0003] The purpose of the present invention is to provide a lateral velocity estimation method of a hybrid driving model based on transfer learning, aiming to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A lateral velocity estimation method of a hybrid driving model based on transfer learning comprises the following steps:

[0006] Step S1, using the mechanism model, i.e., the kinematics and dynamics model, to estimate the lateral velocity of the vehicle, and obtaining the lateral velocity based on the kinematics model and the lateral velocity based on the dynamics model, which are expressed as ;

[0007] Step S2: As the input of LSTM neural network, a data-mechanism hybrid driving model for estimating vehicle lateral velocity is formed, namely LVENN;

[0008] Step S3, LVENN is divided into two parts: an LSTM pre-training model and a transfer learning model, and the LSTM pre-training model and the transfer learning model are divided into two parts: a feature extraction layer and a fitting output layer, so as to achieve parameter fine-tuning in the feature extraction layer and retraining in the fitting output layer.

[0009] Furthermore, the specific steps of step S1 include:

[0010] Step S11, lateral velocity estimation based on kinematic model: First, by integrating equation Get the sideslip angle , and combined with the longitudinal velocity to calculate the lateral velocity, where the longitudinal velocity It is converted from the speed of the non-driven wheel, the lateral and longitudinal acceleration and and the yaw rate Through the IMU output, the integral term is the estimated value at the previous moment;

[0011] Estimating the vehicle's sideslip angle using EKF , and then calculated based on the kinematic model = ;

[0012] Step S12: Estimation of lateral velocity based on dynamic model: Using a three-degree-of-freedom dynamic model, obtain .

[0013] Further, after step S2, using the lateral velocity estimation value The true value of the lateral velocity The mean square error between them is used as the loss function, and then the two different training weight parameters W and b in LVENN are calculated by minimizing the loss function J.

[0014] Furthermore, in step S3, the feature extraction layer uses an LSTM structure to process the time series input, identifies key features for lateral velocity estimation, and fits the key features to estimate the lateral velocity using a shallow neural network in the output layer;

[0015] The specific steps of step S3 include:

[0016] Step S31, source domain training: train LVENN using labeled data from the source domain to create a NetworkSource model;

[0017] Step S32, target domain initialization: use the trained NetworkSource to initialize the LVENN of the target domain;

[0018] Step S33, transfer learning to the target domain: adapt the feature extraction layer from the source domain to the target domain, transfer relevant features for lateral velocity estimation, and then retrain the fitting output layer using the available data in the target domain.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] The lateral velocity estimation method of the hybrid driving model based on transfer learning can significantly reduce the root mean square error and maximum error of lateral velocity estimation; the transfer learning ability can be increased by effectively fine-tuning parameters in the feature extraction layer and retraining parameters in the fitting output layer, that is, the training data in the source domain can be used to reduce the amount of labeled data in the target domain, thereby improving the efficiency of model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a framework diagram of the present invention.

[0022] Figure 2 This is a schematic diagram of the transfer learning principle of the present invention. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0024] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0025] An embodiment of the present invention provides a method for estimating lateral velocity of a hybrid driving model based on transfer learning, comprising the following steps:

[0026] Step S1, using the mechanism model, i.e., the kinematics and dynamics model, to estimate the lateral velocity of the vehicle, and obtaining the lateral velocity based on the kinematics model and the lateral velocity based on the dynamics model, which are expressed as ;

[0027] The specific steps of step S1 include:

[0028] Step S11, lateral velocity estimation based on kinematic model: First, the integral equation Get the sideslip angle , and combined with the longitudinal velocity to calculate the lateral velocity, where the longitudinal velocity Can be converted from the speed of the non-driven wheel, lateral and longitudinal acceleration and and the yaw rate The sideslip angle in the integral term can be output by a low-cost IMU (Inertial Measurement Unit). is the estimated value at the previous moment.

[0029] The state space equation for lateral velocity estimation based on the kinematic model is as follows:

[0030] ;

[0031] In the above formula, X kine is a state variable, is the observed quantity, is the longitudinal acceleration, is the lateral acceleration, is the yaw angular velocity, is the sideslip angle, is the longitudinal velocity, is the state variable at the current moment, is the state variable at the next moment, is the sampling time.

[0032] Considering the nonlinearity of the state-space equation, the EKF (Extended Kalman Filter-Jacobian matrix) is used to estimate the vehicle's sideslip angle. The estimation process can be divided into time update and measurement update. The side slip angle of the vehicle is estimated by EKF. Then, based on the kinematic model, we can calculate = .

[0033] Step S12, estimating the lateral velocity based on the dynamic model;

[0034] The three-degree-of-freedom dynamic model is as follows:

[0035] ;

[0036] ;

[0037] In the above formula, is the driving force applied to each wheel, is the driving resistance (including rolling resistance, wind resistance, slope resistance), is the lateral acceleration, is the longitudinal acceleration, is the yaw angular velocity, is the yaw angular acceleration, is the front wheel turning angle, is the rear wheel turning angle, is the front wheel cornering stiffness, is the rear wheel cornering stiffness, is the lateral velocity, is the longitudinal velocity, is the lateral force of the front wheel, is the rear wheel lateral force, is the longitudinal force, is the lateral force, is the total yaw moment, m is the mass, is the moment of inertia,l f is the front wheelbase, l r is the rear wheelbase, W d For wheelbase.

[0038] The vehicle lateral velocity estimation equation based on the dynamic model is as follows:

[0039] ;

[0040] In the above formula, is the state variable, is the input variable, is the longitudinal velocity, is the yaw angular velocity, is the longitudinal acceleration, is the lateral acceleration, is the longitudinal force, is the lateral force, T d For the sampling time, the lateral velocity estimation based on the dynamic model can also be obtained through the EKF.

[0041] Step S2: Use LSTM neural network structure to estimate the lateral speed of the vehicle. As the input of LSTM neural network, a data-mechanism hybrid driving model for estimating vehicle lateral velocity is formed, namely LVENN;

[0042] The specific steps of step S2 include:

[0043] Step S21: and As the input of the LSTM network, LVENN is formed to estimate the lateral velocity of the vehicle;

[0044] The forward propagation of LVENN is as follows:

[0045] ;

[0046] In the above formula, is the operation function of LSTM, It is the activation layer of Tanh Layer. and is the training parameter, , , , For four fully connected layers, the present invention uses the lateral velocity estimation value The true value of the lateral velocity The mean square error between them is used as the loss function. By minimizing the loss function J, the above training weight parameters W and b can be calculated:

[0047] ;

[0048] in, is the total number of training samples.

[0049] Step S3, LVENN is divided into two parts: an LSTM pre-training model and a transfer learning model, and the LSTM pre-training model and the transfer learning model are divided into a feature extraction layer and a fitting output layer. The feature extraction layer uses an LSTM structure to process the time series input, identifies key features for lateral velocity estimation, and the shallow neural network in the fitting output layer fits key features to estimate the lateral velocity.

[0050] The specific steps of step S3 include:

[0051] Step S31, source domain training: train LVENN using labeled data from the source domain to create a NetworkSource model;

[0052] Step S32, target domain initialization: use the trained NetworkSource to initialize the LVENN of the target domain;

[0053] Step S33, transfer learning to the target domain: adapt the feature extraction layer from the source domain to the target domain, transfer relevant features for lateral velocity estimation, and then retrain the fitted output layer using the available data in the target domain, applying different learning rates To improve accuracy.

[0054] The above are only preferred embodiments of the present invention. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention. These should also be regarded as the protection scope of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A lateral velocity estimation method for a hybrid driving model based on transfer learning, characterized in that: The following steps are involved: Step S1, using the mechanism model, i.e., the kinematics and dynamics model, to estimate the lateral velocity of the vehicle, and obtaining the lateral velocity based on the kinematics model and the lateral velocity based on the dynamics model, which are expressed as Step S2: As the input of LSTM neural network, a data-mechanism hybrid driving model for estimating vehicle lateral velocity is formed, namely LVENN; Step S3, dividing LVENN into two parts: an LSTM pre-training model and a transfer learning model; dividing the LSTM pre-training model into two parts: a feature extraction layer and a fitting output layer; and dividing the transfer learning model into two parts: a feature extraction layer and a fitting output layer, so as to achieve parameter fine-tuning in the feature extraction layer and retraining in the fitting output layer; After step S2, the lateral velocity estimate is used The true value of the lateral velocity The mean square error between is used as the loss function, and then two different training weight parameters W and b in LVENN are calculated by minimizing the loss function J; In step S3, the feature extraction layer uses an LSTM structure to process the time series input, identifies key features for lateral velocity estimation, and fits the key features to estimate the lateral velocity through a shallow neural network in the output layer; The specific steps of step S3 include: Step S31, source domain training: train LVENN using labeled data from the source domain to create a NetworkSource model; Step S32, target domain initialization: use the trained NetworkSource to initialize the LVENN of the target domain; Step S33, transfer learning of the target domain: adapting the feature extraction layer from the source domain to the target domain, transferring relevant features for lateral velocity estimation, and then retraining the fitting output layer using available data in the target domain; The specific steps of step S1 include: Step S11, the state space equation for lateral velocity estimation based on the kinematic model is as follows: X kine =[β,V x ] T ,Y kine =[a x ,V x ] T ; In the above formula, X kine is the state variable, Y kine is the observed quantity, a x is the longitudinal acceleration, a y is the lateral acceleration, is the yaw rate, β is the sideslip angle, V x is the longitudinal velocity, is the state variable at the current moment, is the state variable at the next moment, Δt is the sampling time; Considering the nonlinearity of the state-space equations, the EKF is used to estimate the vehicle's sideslip angle. Then, based on the kinematic model, Step S12: Estimation of lateral velocity based on dynamic model: Using a three-degree-of-freedom dynamic model, obtain

Images