Method for constructing intelligent driver model based on CMAC

Abstract

The present invention discloses a method for constructing an intelligent driver model based on CMAC. The CMAC local neural network and the single-neuron adaptive PID controller are used as the decision-making modules of the driver model. The CMAC neural network has locality, that is, only the activated weights are updated after each learning, and it has good real-time performance. The single-neuron adaptive PID controller is a single-neuron network designed according to the incremental PID that can change the proportional, integral, and differential coefficients in real time, which can improve the stability and anti-interference ability of the controller. The present invention forms a composite controller by combining the CMAC neural network and the adaptive neuron controller. It can not only use the adaptive neuron controller to evaluate the performance of the CMAC controller, but also improve the stability and anti-interference ability of the control system, and has the characteristics of good real-time performance, small output error, and strong robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of automotive automatic control, and particularly relates to a method for constructing an intelligent driver model based on CMAC. Background Technique

[0002] An automotive driver model is a complex control system, which is a mathematical expression of a driver's behavior in operating a vehicle. The driver model includes perception, decision-making, and execution modules. The perception module can obtain road information and vehicle state information. The decision-making module solves the optimal output through a certain mathematical model and according to the information transmitted by the perception module. The execution module simulates the neural and muscle delay behavior of a human driver. Driver models are divided into compensatory tracking driver models, preview driver models, and intelligent driver models. The compensatory tracking driver model decides the output steering angle according to the distance between the lateral position of the vehicle itself and the lane line, without considering the preview link of a real driver; the preview driver model can perceive the distance between the lateral position of the vehicle itself and the lane line at a future moment and make a correction steering angle in advance. Generally, there are single-point preview models, two-point preview models, and multi-point preview models, which can be selected according to different working conditions; the intelligent driver model takes into account multi-layer perception capabilities and can well simulate the thinking mode of the human brain, being more in line with a real driver. Representative ones are neural network driver models and fuzzy control driver models. The neural network driver model can simulate the operation mode of the human driver's brain to the greatest extent, and the fuzzy control driver model can simulate the thinking ability of the human driver. With the continuous development of monitoring technology, communication technology, computer technology, artificial intelligence, and control theory, there are more and more research results on driver models, which have become a hot research issue in the field of autonomous driving.

[0003] The input variables of the driver model in CN 105136469 A are fuzzified, the membership functions of the input and output variables are established using fuzzy values, and a vehicle speed controller based on the improved PSO algorithm and the fuzzy RBF neural network algorithm is built; in CN106202698 B, the relative speed, relative position, time dependence, space dependence, and road surface dependence during vehicle movement are used as uncertainty indicators, and the relevant parameters of the driver model are corrected according to these indicators to characterize the node movement strategy and movement law; the driver model in CN 104260725 B adjusts the performance of the vehicle according to the driving characteristics of the driver and road environment information; in CN 107651010B, an empirical driver operation model, Bang-Bang control, and fuzzy PI control optimized based on reinforcement learning are combined to control the hydraulic servo drive steering system to meet the need for autonomous steering of the driverless vehicle; the intelligent driver model in CN 112874509 B determines the predicted trajectories of the surrounding vehicles of the current vehicle under the influence of each candidate trajectory based on IDM, screens out the candidate trajectories that do not meet the preset constraint conditions, and quantitatively selects the optimal trajectory. Some of the above-established driver models are relatively complex and cannot meet the real-time performance of the control system, and some have poor anti-interference ability and cannot ensure the stability of the system control. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for constructing an intelligent driver model based on CMAC for the defects involved in the background technology.

[0005] The present invention adopts the following technical solutions to solve the above technical problems:

[0006] A method for constructing an intelligent driver model based on CMAC includes the following steps:

[0007] Step 1), a lateral acceleration sensor, a yaw rate sensor, and an inclination sensor are arranged on the vehicle. Among them, the lateral acceleration sensor is used to obtain the lateral speed of the vehicle, the yaw rate sensor is used to obtain the yaw rate of the vehicle, and the inclination sensor is used to obtain the roll angle and roll angle speed of the vehicle.

[0008] Step 2), the current LTR value of the vehicle is derived according to the motion equations of the vehicle three-degree-of-freedom model, that is, the motion equations in the lateral, yaw, and roll directions, and the steady-state LTR value of the vehicle is solved according to the motion equations of the vehicle three-degree-of-freedom model. The LTR is the ratio of the difference between the left and right vertical forces of the vehicle to the sum of the left and right vertical forces.

[0009] The LTR is used to characterize the roll state of the vehicle.

[0010] The vehicle three-degree-of-freedom model is:

[0011] Lateral motion equation:

[0012] Yaw motion equation:

[0013] Roll motion equation:

[0014]

[0015] In the formula, m is the total mass of the vehicle, m s is the sprung mass, δ is the front wheel steering angle, u is the vehicle speed, g is the acceleration due to gravity, a and b are the distances from the center of mass to the front and rear axles respectively, h is the distance from the roll center to the center of mass, B is the track width, a y is the lateral acceleration of the vehicle, k f 、k r are the front wheel cornering stiffness and the rear wheel cornering stiffness respectively, I x is the moment of inertia of the vehicle about the roll center, I z is the moment of inertia of the vehicle about the Z-axis, is the equivalent roll damping of the suspension, is the equivalent roll stiffness of the suspension, v is the lateral velocity, is the lateral acceleration, r is the yaw rate, is the yaw acceleration, is the roll angle of the suspension, is the roll angular velocity of the suspension, is the roll angular acceleration of the suspension;

[0016] Let be 0, solve for the steady-state LTR value, and use the steady-state LTR value as the target value of system control, which can avoid the risk of rollover and not change the driver's driving intention;

[0017] Step 3), make a decision by combining a CMAC neural network (cerebellar model neural network) with a single neuron adaptive PID controller;

[0018] Feedforward and feedback neural networks are global, that is, every time learning is performed, all weights will be updated, while the CMAC neural network is local, that is, only the activated weights are updated after each learning, with good real-time performance; the single neuron adaptive PID controller is a single neuron network designed based on incremental PID that can change the proportional, integral, and differential coefficients in real time, which can improve the stability and anti-interference ability of the controller;

[0019] Step 3.1), at the initial moment, all weights of the CMAC neural network are 0, the initial output steering angle δ2 of the CMAC neural network is 0, and the initial weights of the single neuron adaptive PID controller are preset in advance;

[0020] In step 3.2), the difference between the current LTR and the steady-state LTR is used as the error input, and the errors at three adjacent moments are converted into three state variables of the incremental PID, namely:

[0021] e(k) = ΔLTR(k)

[0022] x1(k) = e(k)

[0023] x2(k) = e(k) - e(k - 1)

[0024] x3(k) = e(k) - 2e(k - 1) + e(k - 2)

[0025] where ΔLTR(k) is the error at moment k, and x1, x2, and x3 are the three state variables of the incremental PID respectively;

[0026] In step 3.3), solve for the output rotation angle δ1 of the current single-neuron adaptive PID controller:

[0027]

[0028] where K is the proportionality coefficient, is the weighting coefficient corresponding to x i (k);

[0029] In step 3.4), use the supervised Hebb algorithm to update the coefficients of the three state variables of the incremental PID according to the error, the three state variables of the incremental PID, and δ1:

[0030] w i (k + 1) = w i (k) + η i e(k)δ1(k)x i (k)

[0031] where w i (k) is the coefficient corresponding to x i (k), and η i is the learning rate;

[0032] In step 3.5), perform normalization processing on the current LTR and the error e(k), and quantize them into the 0-γ interval:

[0033]

[0034]

[0035] where α1 and α2 are the quantization values of the current LTR and the error e(k) respectively, LTR min 、LTR maxare the minimum and maximum values of the current LTR change range, ΔLTR min , ΔLTR max are the minimum and maximum values of the error e(k) change range respectively, and γ is a preset quantization coefficient;

[0036] Step 3.6), perform address mapping according to the quantized current LTR and error e(k), and activate the weights;

[0037] The address mapping here is equivalent to a table lookup method, that is, each address corresponds to a weight. According to α1 and α2, the corresponding address is found. The two-dimensional input is divided into m layers, and each layer is divided into nb blocks. In this way, after the quantization process of every two input quantities, the unique address of each layer can be determined, and the weight corresponding to the address will be activated. The address is found according to the address formula, and the weight is activated, that is:

[0038] b x = int((α1 + p) / m);

[0039] b y = int((α2 + p) / m);

[0040] s(q) = b x + b y ×nb + (q - 1)×nb 2 + 1.

[0041] In the formula, b x , b y are address coefficients respectively, m is the number of layers, nb is the number of blocks, q = 1 to 4, p = m - q, and s(q) is the weight address;

[0042] Step 3.7), sum the weights as the output rotation angle δ2 of the CMAC neural network, that is:

[0043]

[0044] In the formula, ω j is the activated weight, and a j is the CMAC memory address selection vector;

[0045] Step 3.8), use the sum of the output rotation angle δ1 of the single neuron adaptive PID controller and the output rotation angle δ2 of the CMAC neural network as the front wheel rotation angle δ. According to the difference between δ and δ2, use the gradient descent method to update the weights of the activated address. The weight update formula is:

[0046]

[0047] In the formula, β is the preset learning rate of the CMAC neural network;

[0048] Step 4), considering that the driver cannot execute the operation immediately when the brain issues an instruction, there will be a certain neural delay and muscle delay. Two transfer functions G1(s) and G2(s) are used to simulate the neural and muscle delays of the human driver respectively. The output steering angle δ is used as the input of G1(s), the output of G1(s) is used as the input of G2(s), and the output of G2(s) is used as the target steering angle of the vehicle steering wheel;

[0049] The two transfer functions are respectively:

[0050]

[0051]

[0052] Among them, T d is the delay time of the driver's nervous system, and T n is the delay time of the driver's muscle system.

[0053] Compared with the prior art, the present invention adopts the above technical solutions and has the following technical effects:

[0054] (1) The present invention proposes a method for constructing a CMAC intelligent driver model for vehicle rollover prevention, and designs a thinking operation mode that is closer to the real driver to prevent vehicle rollover;

[0055] (2) The present invention introduces a single neuron adaptive PID controller, which can not only make its control parameters change with the time-varying of the system, but also improve the overlearning phenomenon of the CMAC neural network;

[0056] (3) The present invention constitutes a composite controller by combining the CMAC neural network and the adaptive neuron controller. It can not only use the adaptive neuron controller to evaluate the performance of the CMAC controller, but also improve the stability and anti-interference ability of the control system, and has the characteristics of good real-time performance, small output error and strong robustness. Description of the Drawings

[0057] Figure 1 is the overall control structure diagram of the present invention.

[0058] Figure 2 is the corresponding relationship diagram between the address formula and the weight table of the present invention. Detailed Embodiment

[0059] The technical solutions of the present invention will be further described in detail below with reference to the drawings:

[0060] The present invention can be implemented in many different forms and should not be construed as limited to the embodiments described herein. On the contrary, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. In the drawings, components are enlarged for clarity.

[0061] As Figure 1 shown, the present invention discloses a method for constructing an intelligent driver model based on CMAC, comprising the following steps:

[0062] Step 1), a lateral acceleration sensor, a yaw rate sensor, and an inclination sensor are arranged on the vehicle. Among them, the lateral acceleration sensor is used to obtain the lateral speed of the vehicle, the yaw rate sensor is used to obtain the yaw rate of the vehicle, and the inclination sensor is used to obtain the roll angle and roll angle speed of the vehicle;

[0063] Step 2), the current LTR value of the vehicle is derived according to the motion equations of the vehicle's three-degree-of-freedom model, that is, the motion equations in the lateral, yaw, and roll directions, and the steady-state LTR value of the vehicle is solved according to the motion equations of the vehicle's three-degree-of-freedom model. The LTR is the ratio of the difference between the left and right vertical forces of the vehicle to the sum of the left and right vertical forces;

[0064] The LTR is used to characterize the roll state of the vehicle.

[0065] The vehicle's three-degree-of-freedom model is:

[0066] Lateral motion equation:

[0067] Yaw motion equation:

[0068] Roll motion equation:

[0069]

[0070] In the formula, m is the total mass of the vehicle, m s is the sprung mass, δ is the front wheel steering angle, u is the vehicle speed, g is the acceleration due to gravity, a and b are the distances from the center of mass to the front and rear axles respectively, h is the distance from the roll center to the center of mass, B is the wheelbase, a y is the lateral acceleration of the vehicle, k f 、k r are the front wheel cornering stiffness and the rear wheel cornering stiffness respectively, I x is the moment of inertia of the vehicle about the roll center, I z is the moment of inertia of the vehicle about the Z axis, is the equivalent roll damping of the suspension, is the equivalent roll stiffness of the suspension, v is the lateral speed, is the lateral acceleration, r is the yaw rate, is the yaw acceleration, is the roll angle of the suspension, is the roll angular velocity of the suspension, is the roll angular acceleration of the suspension;

[0071] Let be 0, solve the steady-state LTR value, and use the steady-state LTR value as the target value of system control, which can not only avoid the rollover risk but also not change the driver's driving intention;

[0072] Step 3), make a decision by combining the CMAC neural network (cerebellar model) with the single-neuron adaptive PID controller;

[0073] The feedforward and feedback neural networks are global, that is, every time learning is carried out, all weights will be updated, while the CMAC neural network is local, that is, only the activated weights are updated after each learning, with good real-time performance; the single-neuron adaptive PID controller is a single-neuron network designed according to the incremental PID that can change the proportional, integral, and differential coefficients in real time, which can improve the stability and anti-interference ability of the controller;

[0074] Step 3.1), at the initial moment, all weights of the CMAC neural network are 0, the initial output angle δ2 of the CMAC neural network is 0, and the initial weights of the single-neuron adaptive PID controller are preset;

[0075] Step 3.2), take the difference between the current LTR and the steady-state LTR as the error input, and convert the errors at three adjacent moments into three state variables of the incremental PID, that is:

[0076] e(k) = ΔLTR(k)

[0077] x1(k) = e(k)

[0078] x2(k) = e(k) - e(k - 1)

[0079] x3(k) = e(k) - 2e(k - 1) + e(k - 2)

[0080] In the formula, ΔLTR(k) is the error at the kth moment, and x1, x2, and x3 are the three state variables of the incremental PID respectively;

[0081] Step 3.3), solve the output angle δ1 of the current single-neuron adaptive PID controller:

[0082]

[0083] In the formula, K is the proportionality coefficient, is the weighting coefficient corresponding to x i (k);

[0084] In step 3.4), the coefficients of the three state variables of the incremental PID are updated according to the error, the three state variables of the incremental PID, and δ1 using the supervised Hebb algorithm:

[0085] w i (k + 1) = w i (k) + η i e(k)δ1(k)x i (k)

[0086] where w i (k) is the coefficient corresponding to x i (k), and η i is the learning rate;

[0087] In step 3.5), the current LTR and the error e(k) are normalized and quantized into the 0-γ interval:

[0088]

[0089]

[0090] where α1 and α2 are the quantization values of the current LTR and the error e(k) respectively, LTR min , LTR max are the minimum and maximum values of the change range of the current LTR respectively, ΔLTR min , ΔLTR max are the minimum and maximum values of the change range of the error e(k) respectively, and γ is the preset quantization coefficient;

[0091] In step 3.6), address mapping is performed according to the quantized current LTR and error e(k) to activate the weights;

[0092] Here, the address mapping is equivalent to a table lookup method, that is, each address corresponds to a weight. According to α1 and α2, the corresponding address is found. The two-dimensional input is divided into m layers, and each layer is divided into nb blocks. In this way, after the two input quantities are quantized, the unique address of each layer can be determined, and the weight corresponding to the address will be activated. The address is found according to the address formula to activate the weight, that is:

[0093] b x = int((α1 + p) / m);

[0094] b y = int((α2 + p) / m);

[0095] s(q) = b x + by ×nb+(q - 1)×nb 2 +1.

[0096] In the formula, b x and b y are address coefficients respectively, m is the number of layers, nb is the number of blocks, q = 1 - 4, p = m - q, and s(q) is the weight address;

[0097] As Figure 2 shown, the results after quantization of LTR and the error LTR are 3.5 and 3 respectively. There are 4 addresses aA, aB, bA, bB in the first layer, and these 4 addresses correspond to Figure 1 ω1, ω2, ω3, ω4 in the weight table respectively. The corresponding relationships of other layers are similar to those of the first layer. According to the address formula, we can calculate the result 4 of the first layer, that is, the address is bB, and extract the weight ω4. By analogy, the activation weights of other layers can be calculated as ω8, ω 12 , ω 13 , that is, the output corner δ2 of the CMAC neural network = ω4 + ω8 + ω 12 + ω 13 ;

[0098] Step 3.7), sum the weights as the output corner δ2 of the CMAC neural network, that is:

[0099]

[0100] In the formula, ω j is the activated weight, and a j is the CMAC memory address selection vector;

[0101] Step 3.8), take the sum of the output corner δ1 of the single neuron adaptive PID controller and the output corner δ2 of the CMAC neural network as the front wheel corner δ. According to the difference between δ and δ2, use the gradient descent method to update the weights of the activated addresses. The weight update formula is:

[0102]

[0103] In the formula, β is the preset learning rate of the CMAC neural network;

[0104] Step 4), considering that the driver cannot execute the operation immediately when sending an instruction from the brain, there will be a certain neural delay and muscle delay. Use two transfer functions G1(s) and G2(s) to simulate the neural and muscle delays of the human driver respectively. Take the output corner δ as the input of G1(s), take the output of G1(s) as the input of G2(s), and take the output of G2(s) as the target corner of the car steering wheel;

[0105] The two transfer functions are respectively:

[0106]

[0107]

[0108] Among them, T d is the delay time of the driver's nervous system, and T n is the delay time of the driver's muscle system.

[0109] The present invention constitutes a composite controller by combining a CMAC neural network and an adaptive neuron controller. It can not only use the adaptive neuron controller to evaluate the performance of the CMAC controller, but also improve the stability and anti-interference ability of the control system, and has the characteristics of good real-time performance, small output error, and strong robustness.

[0110] Those skilled in the art of this technology can understand that unless otherwise defined, all terms (including technical terms and scientific terms) used here have the same meaning as the general understanding of those of ordinary skill in the field to which the present invention belongs. It should also be understood that terms defined in a general dictionary, such as those, should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as here.

[0111] The specific embodiments described above have further elaborated on the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for constructing an intelligent driver model based on CMAC, characterized in that, Specifically, it includes the following steps: Step 1), a lateral acceleration sensor, a yaw rate sensor, and an inclination sensor are set on the vehicle. Among them, the lateral acceleration sensor is used to obtain the lateral speed of the vehicle, the yaw rate sensor is used to obtain the yaw rate of the vehicle, and the inclination sensor is used to obtain the roll angle and roll rate of the vehicle; Step 2), derive the LTR value of the current vehicle according to the motion equations of the vehicle's three-degree-of-freedom model, that is, the motion equations in the lateral, yaw, and roll directions, and solve the steady-state LTR value of the vehicle according to the vehicle's three-degree-of-freedom model. The LTR is the ratio of the difference between the left and right vertical forces of the vehicle to the sum of the left and right vertical forces; Step 3), make a decision by combining a CMAC neural network and a single-neuron adaptive PID controller; Step 3.1), at the initial moment, all the weights of the CMAC neural network are 0, the initial output angle δ2 of the CMAC neural network is 0, and the initial weights of the single-neuron adaptive PID controller are preset in advance; Step 3.2), take the difference between the current LTR and the steady-state LTR as the error input, and convert the errors at three adjacent moments into three state variables of the incremental PID, that is: e(k) = ΔLTR(k) x1(k) = e(k) x2(k) = e(k) - e(k - 1) x3(k) = e(k) - 2e(k - 1) + e(k - 2) In the formula, ΔLTR(k) is the error at the kth moment, and x1, x2, and x3 are the three state variables of the incremental PID respectively; Step 3.3), solve the output angle δ1 of the current single-neuron adaptive PID controller: where K is the proportionality coefficient, corresponding to x i (k) is the weighting coefficient; Step 3.4), use the supervised Hebb algorithm to update the coefficients of the three state variables of the incremental PID according to the error, the three state variables of the incremental PID, and δ1: w i (k + 1)= w i (k)+ η i e(k)δ1(k)x i (k) where, w i (k) is the coefficient corresponding to x i (k), and η i is the learning rate; Step 3.5), perform normalization processing on the current LTR and the error e(k), and quantize them into the 0-γ interval: where α1 and α2 are the quantization values of the current LTR and the error e(k), respectively, and LTR min and LTR max are the minimum and maximum values of the current LTR change range, respectively, ΔLTR min and ΔLTR max are the minimum and maximum values of the error e(k) change range, respectively, and γ is a preset quantization coefficient; Step 3.6), perform address mapping according to the quantized current LTR and the error e(k), and activate the weights; Step 3.7), sum the weights as the output angle δ2 of the CMAC neural network, that is: where ω j is the activated weight, and a j is the CMAC memory address selection vector; Step 3.8), take the sum of the output angle δ1 of the single-neuron adaptive PID controller and the output angle δ2 of the CMAC neural network as the front wheel angle δ. According to the difference between δ and δ2, use the gradient descent method to update the weights of the activated address. The weight update formula is: In the formula, β is the preset learning rate of the CMAC neural network; Step 4), use the transfer functions G1(s) and G2(s) to simulate the neural delay and muscle delay of a human driver respectively. Take the output angle δ as the input of G1(s), take the output of G1(s) as the input of G2(s), and take the output of G2(s) as the target angle of the vehicle steering wheel; Among them, T d is the delay time of the driver's nervous system, and T n is the delay time of the driver's muscle system.

2. The method for constructing an intelligent driver model based on CMAC according to claim 1, characterized in that, The three-degree-of-freedom model of the vehicle in step 2) is: Lateral motion equation: Yaw motion equation: Rolling motion equation: where m is the total mass of the vehicle, m s is the sprung mass, δ is the front wheel steering angle, u is the vehicle speed, g is the acceleration due to gravity, a and b are the distances from the center of mass to the front and rear axles respectively, h is the distance from the roll center to the center of mass, B is the track width, a y is the lateral acceleration of the vehicle, k f 、k r are the front wheel cornering stiffness and the rear wheel cornering stiffness respectively, I x is the moment of inertia of the vehicle about the roll center, I z is the moment of inertia of the vehicle about the Z-axis, is the equivalent roll damping of the suspension, is the equivalent roll stiffness of the suspension, v is the lateral velocity, is the lateral acceleration, r is the yaw rate, is the yaw acceleration, is the roll angle of the suspension, is the roll angular velocity of the suspension, is the roll angular acceleration of the suspension; Let be 0 and solve for the steady-state LTR value.

3. The method for constructing an intelligent driver model based on CMAC according to claim 1, wherein, The formula for address mapping according to the quantized current LTR and the error e(k) in step 3.6) is as follows: b x = int((α1 + p) / m); b y = int((α2 + p) / m); s(q) = b x + b y × nb+(q - 1)× nb 2 + 1. where b x and b y are address coefficients respectively, m is the number of layers, nb is the number of blocks, q = 1 to 4, p = m - q, and s(q) is the weight address.

Images