Apparatus and method for adaptive identification of vehicle models

Abstract

To predict the position of the center of gravity and the azimuth angle of a vehicle with high accuracy even when the vehicle is traveling in a state where the sideslip angle increases due to disturbances. [Solution] An adaptive identification device 50 for a vehicle model includes a neural network (NN) 52 that receives vehicle speed v, steering angle δ, and yaw rate ω as input and outputs the modeling error of a vehicle model 42; an estimate calculation unit 54 that calculates estimates of the vehicle's position coordinates x, y, azimuth angle ψ, and sideslip angle β at each sampling time using δ^ and v^ corrected by the modeling error based on the vehicle model 42; an online learning unit 56 that performs online learning on the NN 52 using the observed values ​​and estimated values ​​of v, δ, ω, x, y, ψ, and β at each sampling time from the current time to a predetermined time ago during a control period so as to minimize an evaluation function including errors of v, ψ, and β; and a model update unit 58 that inputs the observed values ​​of v, δ, and ω at the current time into the trained NN 52 to estimate the modeling error and update the modeling error of the vehicle model 42 to the estimated modeling error.

Description

[Technical Field]

[0001] The present invention relates to a model predictive control used in an automated driving system for a vehicle, and more particularly to an apparatus and method for adaptively identifying a vehicle model. [Background technology]

[0002] Currently, in research into autonomous vehicle driving, technology that combines model-based control and AI is attracting attention. Among these, a control method called model predictive control (MPC), which uses a model to predict future behavior and determine optimal inputs, is attracting attention. In MPC, the quality of the vehicle model that generates the predicted trajectory has a significant impact on control performance.

[0003] One vehicle model commonly used in MPC is the kinematic bicycle model (KBM), which models a vehicle as a bicycle with two axles, front and rear. According to the KBM, the entire vehicle is treated as a single rigid body, and the vehicle's direction of travel is assumed to be controlled by the steering angle of the front wheels, while the rear wheels move in the vehicle's direction of travel. This allows for the focus to be on kinematic elements, ignoring dynamic factors such as tire slip and inertial force. KBMs have the advantage of being simple in structure, resulting in low computational costs and easy adoption in controller design for autonomous driving systems. However, their simplicity also poses the problem of being unable to account for nonlinear vehicle characteristics, resulting in modeling errors between the actual vehicle and the vehicle model.

[0004] To solve this problem, the inventor has proposed a method of expressing the center of gravity position error of a vehicle model as a deviation in the front wheel steering angle, and learning and estimating this deviation using a neural network to accurately express the coordinates of the center of gravity position of the vehicle.Furthermore, in response to the problem of the modeling error of the azimuth angle becoming large under certain conditions such as large steering, acceleration / deceleration, high-speed driving, and road surface conditions, the inventor has proposed a method of expressing the modeling error as a deviation in speed in addition to the deviation in the front wheel steering angle, and estimating this deviation online using a neural network to model the vehicle model in real time and with high accuracy in accordance with changes in the road and driving conditions (for example, see Non-Patent Document 1). [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] S. Matsumoto and M. Saito, “Adaptive Identification Method for Vehicle Driving Model Capable of Driving with Large Acceleration Changes and Steering,” J. Adv. Comput. Intel. Intell. Inform., Vol.27 No.4, pp. 609-615, 2023. Summary of the Invention [Problem to be solved by the invention]

[0006] In the prior example of Non-Patent Document 1, the vehicle's center of gravity position and azimuth angle around the center of gravity, which are among the vehicle behaviors, can be modeled with high accuracy, but there is a problem that the prediction accuracy decreases when driving in a manner in which the sideslip angle increases due to external disturbances. Therefore, an object of the present invention is to make it possible to predict the vehicle's center of gravity position and azimuth angle with high accuracy even when driving in a manner in which the sideslip angle increases due to external disturbances. [Means for solving the problem]

[0007] According to one aspect of the present invention, there is provided an apparatus for adaptively identifying a vehicle model used in a model-based controlled automated driving system, the vehicle model being a model in which modeling errors in the left-right and front-rear directions of the vehicle are added to a kinematic bicycle model, the apparatus comprising: a neural network that receives vehicle speed, steering angle, and yaw rate as input and outputs the modeling errors; an estimated value calculation unit that calculates estimated values ​​of vehicle position coordinates, azimuth angle, and sideslip angle at each sampling time using the vehicle steering angle and speed corrected by the modeling errors based on the vehicle model; and a control cycle that calculates estimated values ​​of vehicle position coordinates, azimuth angle, and sideslip angle at each sampling time from the current time to a predetermined time before. and a model updating unit that inputs the observed values ​​of the vehicle speed, steering angle, and yaw rate at the current time into the trained neural network to estimate the modeling error, and updates the modeling error of the vehicle model to the estimated modeling error. [Effects of the Invention]

[0008] According to the present invention, the yaw rate of the vehicle is input to a neural network that estimates the modeling error of the vehicle model, and the error in the vehicle's sideslip angle is included in the evaluation function for learning of the neural network. This makes it possible to predict the position of the center of gravity and the azimuth angle of the vehicle with high accuracy even when driving in a situation where the sideslip angle becomes large due to an external disturbance. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a block diagram of an example of a model-based control autonomous driving system. [Figure 2] FIG. 2 is a schematic diagram of a vehicle model employed in the present embodiment. [Figure 3] FIG. 1 is a schematic diagram illustrating an example of the configuration of a neural network that estimates a modeling error. [Figure 4] 4 is a flowchart illustrating adaptive identification of a vehicle model. [Figure 5] 1 is a graph showing a course map used in a simulation and speeds and steering angles input to a vehicle model. [Figure 6] 10 is a graph showing position, azimuth, and sideslip angle estimation errors comparing KBM, prior art, and an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings as appropriate. However, more detailed description than necessary may be omitted. For example, detailed description of well-known matters or redundant description of substantially identical configurations may be omitted. This is to avoid unnecessary redundancy in the following description and to facilitate understanding by those skilled in the art. Note that the inventors provide the accompanying drawings and the following description to enable those skilled in the art to fully understand the present invention, and do not intend for them to limit the subject matter described in the claims.

[0011] <Autonomous Driving System> 1 is a block diagram of an example of a model-based control autonomous driving system. The autonomous driving system 100 is installed in a vehicle such as an automobile that travels on roads, and enables the vehicle to autonomously drive using model predictive control. Specifically, the autonomous driving system 100 includes an operation system 10, a target trajectory generation unit 20, an on-board sensor 30, a model predictive control device 40, and a vehicle model adaptive identification device 50.

[0012] The operation system 10 is a part that controls the accelerator, brake, and steering of the vehicle, which directly affect the running of the vehicle. More specifically, the operation system 10 receives command values ​​for the front wheel steering angle (hereinafter referred to as the steering angle) and speed as control inputs from the model predictive control device 40, and determines the control amounts for the accelerator, brake, and steering accordingly to operate them.

[0013] The target trajectory generating unit 20 is a part that generates a target trajectory for performing model predictive control of vehicle travel. More specifically, the model predictive control device 40 includes map information, a Global Navigation Satellite System (GNSS), a front-mounted camera, LiDAR, etc., and obtains vehicle travel position information from the GNSS, grasps the road shape ahead in the direction of travel from the map information, and further detects the presence or absence of obstacles, etc. from image information ahead of the vehicle, generates a target trajectory that allows the vehicle to travel safely, and provides it to the model predictive control device 40.

[0014] The on-board sensors 30 are part of a group of sensors that observe the behavior of the vehicle. Specifically, the on-board sensors 30 include a speed sensor that detects the speed of the vehicle, a steering angle sensor that detects the angle of the steering shaft, a yaw rate sensor that detects the yawing of the vehicle, a geomagnetic sensor that detects the azimuth angle of the vehicle, and a sideslip angle sensor that detects the sideslip of the vehicle.

[0015] The model predictive control device 40 is a control module that predicts future vehicle behavior using a vehicle model 42 from the target trajectory provided by the target trajectory generating unit 20 and various information related to the vehicle behavior observed by the on-board sensors 30, and determines the current optimal steering angle and speed based on this prediction using an optimizer 44. Specifically, the model predictive control device 40 can be configured as a software module executed on a computer such as an on-board ECU (Electric Control Unit).

[0016] The vehicle model adaptive identification device 50 adaptively identifies the vehicle model 42 online using a neural network 52. Details will be described later.

[0017] The autonomous driving system 100 is derived so as to minimize the evaluation function J shown in equation (1). J(δ[k],v[k])=Σ i=1 H {(x^[k+i]-x[k+i]) 2 +(y^[k+i]-y[k+i]) 2 +(ψ^[k+i]-ψ[k+i]) 2}(1) Here, δ is the steering angle of the vehicle, v is the vehicle speed, x is the X coordinate of the vehicle, y is the Y coordinate of the vehicle, ψ is the azimuth angle of the vehicle, and H is the prediction horizon. In addition, brackets indicate the sampling time, values ​​with a hat (^) are estimated values, and values ​​without a hat are observed values.

[0018] The vehicle model 42 predicts the vehicle trajectory, represented by x^[k+i], y^[k+i], and ψ^[k+i] (1≦i≦H). Because this behavior depends on future inputs (δ[k+i],v[k+i]) (1≦i≦H), the input sequence is determined so that the predicted output follows the target trajectory. Of the input sequence (δ[k+i],v[k+i]) (1≦i≦H) obtained in this way, only the vehicle's steering angle and velocity (δ[k],v[k]) at the current time can be actually input and given.

[0019] <Vehicle model> Since the KBM does not take into account vehicle dynamics such as tire deformation, suspension expansion / compression, and body shape, or nonlinear characteristics due to the driving environment such as road gradient and changes in road surface, it is effective in situations where the speed is constant and the amount of steering operation is small, such as on a highway, but in situations where large steering operation or acceleration / deceleration is required, or when driving at high speeds or on slippery roads, modeling errors increase and the vehicle's center of gravity position cannot be accurately represented. Therefore, in this embodiment, a vehicle model 42 is adopted in which parameters for compensating for modeling errors in the left-right direction and the front-to-back direction of the vehicle are added to the KBM.

[0020] 2 is a schematic diagram of a vehicle model used in this embodiment. For simplicity, it is assumed that the center of gravity G of the vehicle is located at the center of the wheelbase in the vehicle model 42. That is, the wheelbase is L, and the distance between the front wheel axle and the center of gravity is l. f , the distance between the rear wheel axis and the center of gravity is l r As, l f =l r Assume that =L / 2.

[0021] α^ is a parameter that represents the modeling error in the left-right direction of the vehicle that occurs when the vehicle turns. The steering angle and the actual direction of travel of the vehicle may not match due to the influence of nonlinear characteristics caused by environmental disturbances. Therefore, in the vehicle model 42, the observed value δ[k] of the steering angle is corrected by the modeling error α^, as shown in equation (2). Corrected rudder angle: δ^[k]=δ[k]+α^[k] (2)

[0022] γ^ is a parameter that represents the modeling error in the longitudinal direction of the vehicle that occurs due to the acceleration / deceleration and speed of the vehicle. As with the steering angle, there is a deviation between the observed speed and the actual speed due to the influence of the tires and road surface. Therefore, in the vehicle model 42, the observed speed value v[k] is corrected by the modeling error γ^ as shown in equation (3). Correction speed: v^[k]=v[k]+γ^[k] (3)

[0023] The calculation formulas for the estimated values ​​of the other parameters in the vehicle model 42 are as follows: The position coordinates of the vehicle center of gravity G are taken as the position coordinates of the vehicle. X coordinate:x^[k+1]=x[k]+v^[k]Δt·cos(ψ[k]+β^[k]) (4) Y coordinate:y^[k+1]=y[k]+v^[k]Δt·sin(ψ[k]+β^[k]) (5) Azimuth: ψ^[k+1]=ψ[k]+v^[k]Δt / l r sinβ^[k] (6) Side slip angle: β^[k]=tan -1 (l r / (l f +l r )tanδ^[k]} (7)

[0024] The estimated value of the azimuth angle in equation (6) can also be calculated from the difference between the observed value of the azimuth angle of the vehicle one time before and the position coordinate of the vehicle from one time before, as shown in the following equation (6'). TIFF2026042417000002.tif18128

[0025] <Adaptive identification of vehicle models> The modeling errors α^ and γ^ in the vehicle model 42 change instantaneously due to environmental disturbances and vehicle dynamics. Therefore, the vehicle model adaptive identification device 50 models the nonlinear system with high accuracy and estimates the modeling errors α^ and γ^ using a hierarchical neural network 52 with excellent learning, generalization, and approximation capabilities.

[0026] Figure 3 is a schematic diagram showing an example of the configuration of a neural network that estimates modeling errors. The neural network 52 is a simple three-input, two-output neural network consisting of three layers: an input layer, a hidden layer, and an output layer. The number of nodes in the hidden layer is, for example, four. The input layer and hidden layer, and the hidden layer and output layer are fully connected. The vehicle speed v[k], steering angle δ[k], and yaw rate ω[k] are input to the three nodes in the input layer, and modeling errors α^[k] and γ^[k] are output from the two nodes in the output layer.

[0027] In the prior art, three items, namely, velocity v[k], steering angle δ[k], and acceleration a[k], are input to the neural network, but in this embodiment, yaw rate ω[k] is input instead of acceleration a[k]. This is based on the idea that by inputting the yaw rate, which is correlated with the acceleration in the lateral direction of the vehicle, to the neural network 52, the estimation accuracy of modeling error, particularly in the lateral direction of the vehicle, is improved.

[0028] The input value of the intermediate layer is s, the output value of the intermediate layer is h, the threshold is c, and the weighting coefficients of each link between the input layer and the intermediate layer and between the intermediate layer and the output layer are w1 and w2, respectively. The input / output relationship of neural network 52 is expressed by the following equations (8) to (12). A sigmoid function is used to determine the output of the intermediate layer. I=(v[k],δ[k],ω[k]) T (8) s i [k]=Σ j=1 3 I j [k]·w1 ji [k]+c j [k] (9) h i [k]=1 / {1+exp(-(s j [k]))} (10) α^=Σ i=1 4 w2 i1 [k]·h i [k] (11) γ^=Σ i=1 4 w2 i2 [k]·h i [k] (12)

[0029] The neural network 52 uses various observed values ​​of vehicle behavior from the current time k up to a predetermined time ago represented by the window width W and estimated values ​​calculated based on the vehicle model 42 to learn and update the weighting factors w1 and w2 so as to minimize the evaluation function E shown in equation (13), which includes errors in the vehicle position coordinates (x, y), azimuth angle ψ, and sideslip angle β. The Levenberg-Marquardt method can be used as the learning function for the neural network 52. TIFF2026042417000003.tif13152

[0030] In the prior art, the evaluation function does not include an error in the sideslip angle β, but in this embodiment, the error in the sideslip angle β is included. This is based on the idea that by including the error in the sideslip angle β, which is correlated with the acceleration in the lateral direction of the vehicle, in the evaluation function of the neural network 52, in addition to inputting the yaw rate to the neural network 52, the estimation accuracy of the modeling error, particularly in the lateral direction of the vehicle, is improved.

[0031] 1 , the vehicle model adaptive identification device 50 includes a neural network 52, an estimated value calculation unit 54, an online learning unit 56, and a model update unit 58. Specifically, the vehicle model adaptive identification device 50 can be configured as a software module executed on a computer such as an on-board ECU, similar to the model predictive control device 40.

[0032] Based on the vehicle model 42, the estimated value calculation unit 54 uses the vehicle steering angle δ^ and velocity v^ corrected by the modeling errors α^ and γ^ according to equations (2) and (3) to calculate estimated values ​​(x^, y^, ψ^, β^) of the vehicle's position coordinates, azimuth angle, and sideslip angle at each sampling time according to equations (4) to (7).

[0033] The online learning unit 56 uses the observed values ​​(v, δ, ω, x, y, ψ, β) and estimated values ​​(v^, δ^, ω^, x^, y^, ψ^, β^) of the vehicle speed, steering angle, yaw rate, position coordinates, azimuth angle, and sideslip angle at each sampling time from the current time to a predetermined time ago during the control period to learn the neural network 52 online, i.e., in real time while the vehicle is traveling, so as to minimize the evaluation function E expressed by equation (13).

[0034] The model update unit 58 inputs the observed values ​​(v, δ, ω) of the vehicle speed, steering angle, and yaw rate at the current time into the trained neural network 52 to estimate the modeling errors α^ and γ^, and updates the modeling errors α^ and γ^ of the vehicle model 42 to the estimated modeling errors α^ and γ^.

[0035] Next, the procedure for adaptive identification of the vehicle model 42 will be described with reference to a flowchart. Figure 4 is a flowchart explaining adaptive identification of the vehicle model. Adaptive identification of the vehicle model consists of two independent processing loops: a sampling loop and a control loop.

[0036] The sampling loop is executed, for example, with a sampling period of 0.05 seconds (sampling frequency of 200 Hz). At each sampling period, the ECU records the observed values ​​(v, δ, ω, x, y, ψ, β) of the vehicle's speed, steering angle, yaw rate, position coordinates, azimuth angle, and sideslip angle in a memory (not shown) (S11). The estimated value calculation unit 54 estimates the vehicle's position coordinates, azimuth angle, and sideslip angle (x^, y^, ψ^, β^) using the vehicle's steering angle δ^ and speed v^ corrected by the modeling errors α^ and γ^ based on the vehicle model 42, and records the estimated values ​​in a memory (not shown) (S12).

[0037] If the vehicle is not equipped with a sideslip angle sensor, a value estimated from the yaw rate ω may be used as the observed value β of the sideslip angle.

[0038] The control loop is executed, for example, with a control period of 0.005 seconds (control frequency of 2000 Hz). At each control period, the online learning unit 56 reads from a memory (not shown) the observed and estimated values ​​from the current time k at the start of the control period up to a predetermined time W, and uses these observed and estimated values ​​to train the neural network 52 online (S21). For example, if the predetermined time is 5 seconds, the neural network 52 can learn from 1000 pieces of sample data.

[0039] When the neural network 52 has finished learning, the model update unit 58 inputs the observed values ​​(v[k], δ[k], ω[k]) of the vehicle speed, steering angle, and yaw rate at the current time k into the learned neural network 52 to estimate the modeling errors α^[k] and γ^[k] (S22), and updates the modeling errors α^ and γ^ of the vehicle model 42 to the estimated modeling errors α^[k] and γ^[k] (S23).

[0040] The model predictive control device 40 predicts future vehicle behavior using the vehicle model 42, which is updated at the control cycle as described above, and determines the current optimal speed and steering angle based on this prediction using an optimizer 44, and issues operation instructions to the operation system 10.

[0041] Effect According to this embodiment, the neural network 52 can be used to predict and correct in real time the deviation of the vehicle model 42 from a target trajectory at a future time. This enables automatic driving of the vehicle without prior learning. Furthermore, in this embodiment, the only parameter that needs to be set in advance is the wheelbase of the vehicle, and once the wheelbase is set, the modeling error can be corrected while the vehicle is running, even if the vehicle model changes.

[0042] Furthermore, according to this embodiment, if driving data of the driver's usual steering operation and acceleration / deceleration is acquired, when the driver needs driving assistance or an autonomous vehicle, the system can reproduce the driver's own driving habits up to now. This makes it possible for an elderly person whose driving ability has declined to drive a car autonomously with the feeling that they are driving the car themselves.

[0043] Simulation Next, the effects of this embodiment will be explained through simulations. Figure 5 shows a course map used in the simulation and graphs of the speed and steering angle input to the vehicle model. The simulation was performed using a simulation device connected to a computer equipped with vehicle motion numerical analysis software (CarSim) and a game handle device (manufactured by Logitech). The control frequency of CarSim was 2000 Hz (0.005 s), and the sampling frequency was 200 Hz (0.05 s). The vehicle used was a so-called compact car (B-class hatchback) with front-wheel drive, measuring 1.535 m in height, 1.695 m in width, 3.3 m in length, a 2.6 m wheelbase, and a vehicle mass of 1110 kg.

[0044] A slalom course like the one shown in Figure 5 was created in CarSim, and the trajectory the driver took when driving on this course was defined as the correct trajectory. In this experiment, to replicate the driving behavior of an experienced driver, we had a driver who regularly drives a car, has racing experience, and is familiar with CarSim drive. This simulation simulates Japanese roads, so left-hand traffic is assumed. The simulation course is set to a 3.5-meter-wide road, similar to the width of a typical national highway. Because this course requires large steering angle changes, azimuth angle errors are likely to occur. Furthermore, because this course simulates a wet, slippery road surface (surface friction coefficient μ = 0.5), nonlinear vehicle characteristics are likely to occur, potentially resulting in large modeling errors.

[0045] The graph of speed and steering angle shown in Figure 5 shows the speed and steering angle obtained by the driver's driving. As can be seen from the steering angle graph, the times indicated by the arrows in the figure, 3s, 7s, 9s, 12s, 15s, and 17s, correspond to curves, and extreme values ​​appear in the steering angle as the steering direction changes. The speed v and steering angle δ, as well as the vehicle position coordinates (x, y), yaw rate ω, and sideslip angle β obtained by the driver's driving, are input to a KBM that does not include modeling errors, a vehicle model of a prior art example, and the vehicle model 42 according to this embodiment, and verification is performed to determine whether the estimated values ​​(x^, y^, ψ^) by each model match the actual driving trajectory (x, y, ψ). In the vehicle model 42 according to this embodiment, the azimuth angle is estimated using equation (6').

[0046] FIG. 6 is a graph comparing the estimation errors of position, azimuth angle, and sideslip angle between the KBM, the prior art, and the present embodiment. The KBM exhibits large errors at curve points for all of the position error, azimuth angle error, and sideslip angle error. While the prior art exhibits smaller errors at curve points than the KBM, some errors still occur at curve points. In contrast, the present embodiment exhibits such minimal errors at curve points that it is difficult to tell which curve points correspond to them just by looking at the graphs of position error, azimuth angle error, and sideslip angle error.

[0047] As described above, according to this embodiment, the position of the center of gravity and the azimuth angle of the vehicle can be predicted with significantly higher accuracy than in the prior art when the vehicle is traveling in a state where the sideslip angle is large due to a disturbance.

[0048] <<Variations>> The control period may be longer than the sampling period. Furthermore, the control period is not fixed, and the online learning unit 56 may change the control period depending on the vehicle conditions, such as the vehicle type and driving conditions. For example, the control period may be shortened when driving at high speeds or on rough roads, and may be lengthened when driving at low speeds or on paved roads.

[0049] The number of nodes in the intermediate layer of the neural network 52 and the intermediate layer itself may be increased. However, in order to complete learning within a short control period, it is desirable to have as simple a configuration as possible, as shown in Figure 3.

[0050] The estimation of the vehicle position coordinates, azimuth angle, and sideslip angle by the estimated value calculation unit 54 may be performed collectively by batch processing at the start of the control period, rather than at the sampling period. However, if the control period is very short, there is a risk that the calculations will not be able to keep up, so it is preferable to estimate and record the values ​​at each sampling period.

[0051] As described above, the embodiments have been described as examples of the technology of the present invention. For this purpose, the accompanying drawings and detailed description have been provided. Therefore, the components described in the accompanying drawings and detailed description may include not only components essential for solving the problem, but also components that are not essential for solving the problem in order to exemplify the above technology. Therefore, the fact that these non-essential components are described in the accompanying drawings or detailed description should not be interpreted as immediately indicating that these non-essential components are essential. Furthermore, because the above-described embodiments are intended to exemplify the technology of the present invention, various modifications, substitutions, additions, omissions, etc. may be made within the scope of the claims or their equivalents. [Explanation of symbols]

[0052] 100 Autonomous Driving System 44 vehicle models 50 Vehicle model adaptive identification device 52 Neural Networks 54 Estimated value calculation section 56 Online Learning Department 58 Model Update Department

Claims

1. An apparatus for adaptively identifying a vehicle model used in a model-based controlled autonomous driving system, comprising: the vehicle model is a model in which modeling errors in the left-right direction and the front-rear direction of the vehicle are added to a kinematic bicycle model, a neural network that receives the vehicle speed, steering angle, and yaw rate as inputs and outputs the modeling error; an estimate calculation unit that calculates estimates of a vehicle's position coordinates, azimuth angle, and sideslip angle at each sampling time using the vehicle's steering angle and speed corrected by the modeling error based on the vehicle model; an online learning unit that performs online learning on the neural network using the observed values ​​and the estimated values ​​of the vehicle speed, steering angle, yaw rate, position coordinate, azimuth angle, and sideslip angle at each sampling time from the current time to a predetermined time before during a control period, so as to minimize an evaluation function including errors in the vehicle position coordinate, azimuth angle, and sideslip angle; a model update unit that inputs observed values ​​of the vehicle speed, steering angle, and yaw rate at the current time into the trained neural network to estimate the modeling error, and updates the modeling error of the vehicle model to the estimated modeling error. An adaptive identification device for a vehicle model, comprising:

2. 2. The adaptive identification device for a vehicle model according to claim 1, wherein the estimated value calculation unit estimates the azimuth angle of the vehicle at each sampling time from an observed value of the azimuth angle of the vehicle one time before and a difference between the observed value of the azimuth angle of the vehicle one time before and the position coordinate of the vehicle from one time before.

3. 3. The adaptive identification device for a vehicle model according to claim 1, wherein the online learning unit changes the control period in accordance with the vehicle situation.

4. 3. The adaptive identification device for a vehicle model according to claim 1, wherein the neural network is a three-input, two-output neural network consisting of three layers: an input layer, an intermediate layer, and an output layer.

5. 4. The adaptive identification device for a vehicle model according to claim 3, wherein the neural network is a three-input, two-output neural network consisting of three layers: an input layer, an intermediate layer, and an output layer.

6. 1. A method for adaptively identifying a vehicle model for use in a model-based controlled automated driving system, comprising: the vehicle model is a model in which modeling errors in the left-right direction and the front-rear direction of the vehicle are added to a kinematic bicycle model, determining estimated values ​​of a vehicle position coordinate, an azimuth angle, and a sideslip angle at each sampling time using the steering angle and the speed of the vehicle corrected by the modeling error based on the vehicle model; In a control cycle, using the observed values ​​and the estimated values ​​of the vehicle speed, steering angle, yaw rate, position coordinate, azimuth angle, and sideslip angle at each of the sampling times from the current time to a predetermined time ago, a neural network that receives the vehicle speed, steering angle, and yaw rate as inputs and outputs the modeling error is trained online so as to minimize an evaluation function that includes errors in the vehicle position coordinate, azimuth angle, and sideslip angle; the trained neural network is input with the observed values ​​of the vehicle speed, steering angle, and yaw rate at the current time to estimate the modeling error; updating the modeling error of the vehicle model to the modeling error estimated by the trained neural network; 1. A method for adaptively identifying a vehicle model, comprising:

7. 7. The method for adaptively identifying a vehicle model according to claim 6, wherein the azimuth angle of the vehicle at each sampling time is estimated from the difference between the observed value of the azimuth angle of the vehicle one time before and the position coordinate of the vehicle from one time before.

8. 8. The method for adaptively identifying a vehicle model according to claim 6, wherein the control period is changed in accordance with a running state of the vehicle.

Images