Unmanned vehicle heterogeneous actuator adaptive control method and system based on neural network modeling

Abstract

The invention relates to an unmanned vehicle heterogeneous actuator self-adaptive control method and system based on neural network modeling, belongs to the field of unmanned vehicle self-adaptive control, and solves the problems that a traditional control method depends on fixed parameters and manual adjustment and is difficult to adapt to dynamic characteristics and time-varying characteristics of different unmanned vehicle actuators. Comprising the following steps: acquiring and preprocessing actuator offline data of the unmanned vehicle to obtain an actuator offline data set; training based on the offline data set of the actuator to obtain a basic model of the actuator; in the operation process of the unmanned vehicle, when the average prediction error of the control quantity of the actuator in a preset period window exceeds a preset threshold value, performing online incremental training on the actuator basic model by utilizing actuator online data collected in real time to obtain an actuator self-adaptive model; and taking the control quantity output by the adaptive model of the actuator as a feedforward control quantity, carrying out dynamic weighted fusion on the feedforward control quantity and a feedback control quantity to obtain an adaptive control quantity of the actuator, and issuing the adaptive control quantity to the actuator for execution.

Description

Technical Field

[0001] The present invention relates to the technical field of adaptive control of unmanned vehicles, and particularly to an adaptive control method and system for heterogeneous actuators of unmanned vehicles based on neural network modeling. Background Art

[0002] Currently, in a fleet of special unmanned vehicles, components of different individual vehicles are usually produced by different manufacturers during the R & D stage. Generally, there are several types of chassis, including small, medium, and large ones, and each chassis has different versions. Even for the same version, there are several suppliers. This results in significant differences in the hardware architectures, control strategies, and execution characteristics of the drive motors and braking systems of unmanned vehicles. These differences make the vehicle as a whole perform unevenly in terms of dynamic response, control accuracy, and actuator consistency. Existing control methods mainly rely on feedback control with adjustable parameters and require manual adjustment for each vehicle. This method not only increases the cost of algorithm debugging but also is difficult to maintain reliable and stable control performance in complex environments and during long-term operation.

[0003] The changing road conditions further exacerbate the non-linear and time-varying characteristics of the actuators, increasing the uncertainty of control strategies and making the application of traditional control methods in a fleet of unmanned vehicles inefficient. Actuators include drive motors, mechanical braking systems, mechanical steering systems, suspension adjustment systems, and parking systems, etc. The core challenge lies in that actuators from different manufacturers, different batches, and even the same vehicle will have differences in dynamic characteristics due to hardware characteristics, environmental factors (such as temperature, wear), etc. Traditional fixed-parameter control is difficult to adapt to these changes.

[0004] Existing control methods mainly rely on feedback control PID

[0005] (Proportional-Integral-Derivative, proportional-differential-integral controller), which requires manual adjustment or fixed parameters for different vehicles. This not only increases the cost of algorithm debugging but also is difficult to maintain reliable and stable control performance in complex environments and during long-term operation. The non-linear characteristics of the braking system affect control accuracy, which is reflected in that the characteristics of components such as brake boosters and brake pads will drift with the use time, temperature, and load, resulting in fixed control parameters being unable to meet the long-term operation requirements. In addition, the torque-speed relationship of the drive motor is greatly affected by environmental conditions, and existing modeling methods are difficult to accurately capture its dynamic response characteristics, making the control algorithm lack generalization ability among different vehicles. When the performance of the actuator degrades, traditional control strategies need to be re-debugged, reducing the adaptability and cooperative control efficiency of the fleet of unmanned vehicles. Summary of the Invention

[0006] In view of the above analysis, the embodiments of the present invention aim to provide an adaptive control method and system for heterogeneous actuators of an autonomous vehicle based on neural network modeling, so as to solve the technical problems that traditional control methods rely on fixed parameters and manual adjustment, and it is difficult to adapt to the dynamic characteristics and time-varying characteristics of different vehicle actuators of special vehicle autonomous vehicles, resulting in insufficient control accuracy and stability in complex environments and long-term operations.

[0007] The object of the present invention is mainly achieved through the following technical solutions:

[0008] The present invention provides an adaptive control method for heterogeneous actuators of an autonomous vehicle based on neural network modeling, including the following steps:

[0009] Obtain the offline data of the actuator of the autonomous vehicle and perform preprocessing to obtain the actuator offline data set; based on the actuator offline data set, train to obtain the actuator basic model;

[0010] During the operation of the autonomous vehicle, when the average prediction error of the control amount of the actuator within the preset cycle window exceeds the preset threshold, then use the online data of the actuator collected in real time to perform online incremental training on the actuator basic model to obtain the actuator adaptive model;

[0011] Use the control amount output by the actuator adaptive model as the feedforward control amount, and perform dynamic weighted fusion of the feedforward control amount and the feedback control amount to obtain the adaptive control amount of the actuator, and send it to the actuator for execution.

[0012] Further, when the actuator is a braking system, the sample data in the actuator offline data set is the braking torque, and the sample label is the braking percentage; the actuator basic model is constructed based on the RBF neural network model.

[0013] Further, when the actuator is a drive system, the sample data in the actuator offline data set includes the target vehicle speed, motor speed, dq-axis current components, and dq-axis voltage components, and the sample label is the motor drive torque; the actuator basic model is constructed based on the LSTM neural network model.

[0014] Further, the adaptive control amount of the controlled object is obtained by dynamically weighted fusion of the feedback control amount output by the PID controller and the feedforward control amount, as follows:

[0015] u = λ·u ff +(1 - λ)·u fb

[0016] where u is the finally output adaptive control amount; λ is the weight of the feedforward control amount, u ff is the feedforward control amount, 1 - λ is the weight of the feedback control amount, u fbis the feedback control quantity;

[0017] λ is defined as follows:

[0018]

[0019] where e window 、e m 、e h are respectively the average prediction error of the actuator adaptive model within the current cycle window, the maximum average error in the historical cycle data, and the initial prediction threshold;

[0020] When the actuator is a braking system, the input of the PID controller is the braking torque; when the actuator is a driving system, the input of the PID controller is the target vehicle speed.

[0021] Furthermore, if the actuator is a braking system, the feedforward control quantity is as follows:

[0022] u ff =f RBF (T b ;θ b )

[0023] where f RBF (.) is the actuator adaptive RBF neural network model for the predicted feedforward quantity; θ b is the RBF network parameter;

[0024] If the actuator is a driving system, the feedforward control quantity is as follows:

[0025] u ff =f LSTM (v d ,timeseq(ω m ,i d ,i q ,u d ,u q );θ m )

[0026] where f LSTM (.) is the LSTM neural network model, v d is the target vehicle speed, timeseq(.) represents the time series input, ω m is the driving motor speed, i d 、i q are respectively the dq-axis current components, u d 、u q are respectively the dq-axis voltage components, θ m is the LSTM network parameter.

[0027] Further, the actuator adaptive model outputs a control quantity and a prediction error;

[0028] When the actuator is a braking system, the output control quantity is the braking percentage, and the output prediction error is the braking prediction error;

[0029] When the actuator is a drive system, the output control quantity is the motor drive torque, and the output prediction error is the drive prediction error.

[0030] Further, at the start of online incremental training, the prediction error of the actuator basic model is used as the initial prediction threshold e h ;

[0031] After each round of online incremental training, calculate the average error within the current cycle window to obtain the maximum average error in the historical cycle window;

[0032] If the average error within the current cycle window is different from the current predicted average error e window , then use the average error within the current cycle window to update the average predicted error e window ;

[0033] If the maximum average error in the historical cycle window is greater than the maximum average error e m , then use the maximum average error in the historical cycle window to update the maximum average error e m .

[0034] Further, perform simulation fitting based on the braking percentage - brake caliper pressure curve of different braking systems to obtain multiple groups of braking torque and braking percentage data pairs, as follows:

[0035]

[0036] Among them, T b is the finally generated braking torque, r is the effective braking radius of the brake, G is the reduction ratio of the transmission system, I is the motor input current, k I is the conversion coefficient between current and torque, k η is the conversion coefficient from braking percentage to current, P b is the braking percentage, A is the area of the hydraulic master cylinder piston, and P is the brake caliper pressure;

[0037] The braking torque and braking percentage data pairs form the offline dataset samples of the braking system.

[0038] Further, configure the vehicle simulation running state parameters in MATLAB. During the simulation of vehicle operation, collect the motor target vehicle speed, motor speed, dq - axis current components, dq - axis voltage components, and motor drive torque;

[0039] The simulation operation state parameters include drive motor parameters, flux linkage parameters, and environmental parameters;

[0040] The motor parameters include the number of pole pairs, stator resistance, stator inductance, peak flux linkage, magnetic saturation curve, and back electromotive force coefficient;

[0041] The flux linkage parameters include DC bus voltage, rated output current, overload capacity, and efficiency curve;

[0042] The environmental parameters include road surface gradient and road surface adhesion coefficient;

[0043] The data pairs of the motor target vehicle speed, motor speed, dq-axis current components, dq-axis voltage components, and motor drive torque form the offline data set samples of the drive system.

[0044] On the other hand, the present invention provides an adaptive control system for heterogeneous actuators of an autonomous vehicle based on neural network modeling, including:

[0045] An offline data acquisition and basic model training module, configured to obtain the offline data of the actuators of the autonomous vehicle and perform preprocessing to obtain an actuator offline data set; based on the actuator offline data set, train to obtain an actuator basic model;

[0046] An online data acquisition and model incremental training module, configured to, during the operation of the autonomous vehicle, when the average prediction error of the control amount of the actuator within a preset period window exceeds a preset threshold, use the online data of the actuator collected in real time to perform online incremental training on the actuator basic model to obtain an actuator adaptive model;

[0047] A feedforward-feedback fusion adaptive control module, configured to use the control amount output by the actuator adaptive model as a feedforward control amount, and perform dynamic weighted fusion on the feedforward control amount and the feedback control amount to obtain an adaptive control amount of the actuator, and send it to the actuator for execution.

[0048] Compared with the prior art, the present invention can at least achieve one of the following beneficial effects:

[0049] 1. Through the online incremental training mechanism of the present invention, the adaptive model can adapt to the dynamic changes and wear of the actuator in real time, ensuring the control accuracy and stability during long-term operation, and the model has strong self-adaptability.

[0050] 2. Based on rich offline data training and online data update, the model of the present invention has good generalization ability for actuators of different manufacturers and different batches, reducing the manual debugging cost.

[0051] 3. The present invention combines a feedforward-feedback cooperative control strategy, uses the predicted control quantity of the neural network model as the feedforward control quantity, and the feedback control quantity of the PID controller for real-time dynamic weighted fusion, significantly improving the control accuracy and reducing errors.

[0052] 4. The present invention ensures the completion of data acquisition, adaptive model prediction, and the issuance and execution of the final adaptive control quantity within a preset control period through an efficient neural network inference and adaptive control algorithm, meeting the real-time control requirements.

[0053] 5. The present invention can maintain stable control performance and enhance the robustness of the system by dynamically adjusting the feedforward and feedback control weights in the face of complex working conditions and changes in actuator performance.

[0054] 6. Based on the predicted control quantity of the existing PID controller, the present invention only increases the gain adjustment of the feedforward and feedback control quantities, with very little modification, reducing the modification cost.

[0055] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the subsequent specification, and some advantages can be made obvious from the specification or understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained from the content specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The drawings are only for the purpose of showing specific embodiments and are not considered to be a limitation of the present invention. Throughout the drawings, the same reference numerals represent the same components.

[0057] Figure 1 It is a flowchart of an adaptive control method for an unmanned vehicle heterogeneous actuator based on neural network modeling in an embodiment of the present invention.

[0058] Figure 2 It is a schematic diagram of the system implementation in an embodiment of the present invention.

[0059] Figure 3 It is a schematic diagram of the system architecture in an embodiment of the present invention.

[0060] Figure 4 It is a schematic diagram of feedforward-feedback adaptive control based on neural network in an embodiment of the present invention.

[0061] Figure 5 It is a schematic diagram of obtaining offline data through adaptive control simulation in an embodiment of the present invention.

[0062] Figure 6 It is a schematic diagram of an adaptive control system for an unmanned vehicle heterogeneous actuator based on neural network modeling in an embodiment of the present invention. Detailed implementation manners

[0063] The following combines the accompanying drawings to specifically describe the preferred embodiments of the present invention. Among them, the accompanying drawings form a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, rather than to limit the scope of the present invention.

[0064] Embodiment 1:

[0065] To solve the above problems, the present invention proposes an adaptive control method based on a neural network. By fitting the characteristics of specific actuator components (drive system and braking system) through a neural network, the present invention provides an adaptive control method that can adapt to the characteristics of different vehicle actuators and dynamically adjust control parameters, enabling the same set of algorithms to be applicable to different vehicles in a special unmanned vehicle fleet. This method can not only reduce the cost of algorithm debugging, but also improve the adaptability and robustness of the control system, ensuring stable control performance during variable road conditions and long-term operation. Solve the problem of inconsistent control caused by differences in actuators from different manufacturers and different batches.

[0066] The method in the present invention aims at the differences in actuators of different manufacturers and different hardware architectures in an unmanned vehicle fleet, and improves the consistency and robustness of vehicle control.

[0067] The actuators in the present invention include the braking system and the drive system of the unmanned vehicle; the main technical principle can be divided into two parts: one is to respectively perform neural network fitting on the drive system and the braking system, that is, adaptive modeling, to fit the input-output characteristics of the drive system and the braking system; the other is to dynamically weight and fuse the fitted control quantity as the feedforward control quantity and the feedback control quantity output by the PID controller to achieve adaptive control to obtain the control quantity of the actuator and send it to the actuator for execution.

[0068] A specific embodiment of the present invention, as Figure 1 shown, discloses an adaptive control method for heterogeneous actuators of an unmanned vehicle based on neural network modeling, including the following steps:

[0069] Step S1, obtain the offline data of the actuator of the unmanned vehicle and perform preprocessing to obtain the actuator offline data set; based on the actuator offline data set, train to obtain the actuator basic model;

[0070] Step S2, during the operation of the unmanned vehicle, when the average prediction error of the control quantity of the actuator within a preset period window exceeds a preset threshold, then use the online data of the actuator collected in real time to perform online incremental training on the actuator basic model to obtain the actuator adaptive model;

[0071] Step S3: Use the control quantity output by the actuator adaptive model as the feedforward control quantity. Dynamically weight and fuse the feedforward control quantity and the feedback control quantity to obtain the adaptive control quantity of the actuator, and send it to the actuator for execution.

[0072] Step S1 includes steps S11 - S13.

[0073] Obtain the offline data of the unmanned vehicle's actuators (drive system and braking system) and perform preprocessing to train the drive system basic model and the braking system basic model.

[0074] Step S11: Obtain the offline data of the braking system and drive system of the unmanned vehicle, and perform preprocessing to obtain the braking system offline data set and the drive system offline data set.

[0075] As Figure 5 shown, establish a vehicle dynamics simulink simulation model to obtain the offline data of the braking system and drive system. Obtain the offline data of the braking system and drive system for training the actuator basic model. Collect offline data: including the braking percentage - brake caliper pressure of the braking system and the torque - speed data pairs of the drive system.

[0076] Exemplarily, build the vehicle dynamics models of the corresponding braking system and drive system based on simulink VDB (Vehicle Dynamics Blockset); the components of the vehicle include tires, braking system, drive system, transmission system, etc.; the parameters of the drive motor are mainly the power output curve, as well as magnetic flux, inductance, equivalent resistance, and number of pole pairs; establish the corresponding braking system and drive system models according to the technical manuals of different manufacturers. Environmental parameters such as road surface gradient and road surface adhesion coefficient are related to the test scenario, but ultimately affect the dynamic effect in the simulink model. The basic controller is the basic controller directly associated with the actuator;

[0077] Write a matlab script to collect the data of simulink, record the drive torque, motor speed, voltage and current status issued by the drive system of the unmanned vehicle, and the target vehicle speed. The offline data is stored in the csv data set, and the python script defines the RBF (Radial basis function network) and LSTM (Long Short - Term Memory) neural network models for the training script of machine learning.

[0078] (1) Simulate to obtain the offline data of the braking system.

[0079] Based on the simulation fitting of the braking percentage - brake caliper pressure curves of different braking systems, multiple sets of braking torque and braking percentage data pairs are obtained as follows:

[0080]

[0081] Among them, T b is the finally generated braking torque, r is the effective braking radius of the brake, G is the reduction ratio of the transmission system, I is the motor input current, k I is the conversion coefficient between current and torque, k η is the conversion coefficient from braking percentage to current, P b is the braking percentage, A is the piston area of the hydraulic master cylinder, P is the brake caliper pressure; P is an intermediate variable;

[0082] The braking torque and braking percentage data pairs form the offline dataset samples of the braking system.

[0083] The braking percentage - brake caliper pressure curves of different braking systems can be obtained from the technical manuals of different manufacturers. The simulation system fits the braking percentage - brake caliper pressure data pairs, and at the same time, the simulation system also brings some interference noise to the simulation data of the braking system.

[0084] Based on the data manuals of different manufacturers of different braking systems, the parameters in the formula can be determined, and a Simulink simulation model can be built. Multiple sets of offline datasets of the braking system are generated through the simulation model.

[0085] The sample data in the offline dataset is the braking torque, and the sample label is the braking percentage.

[0086] Exemplarily,

[0087] Simulation time: 100 seconds;

[0088] Sampling frequency: 100 Hz (100 samples per second)

[0089] Different input conditions: 10 different braking systems

[0090] Then the total sample data volume is estimated as:

[0091] Total number of samples = simulation time × sampling frequency × number of input condition combinations

[0092] Total number of samples = 100 seconds × 100 Hz × 10 = 100000 samples

[0093] In practical applications, the sample data volume can be adjusted according to specific requirements and computing power resources. Usually, more sample data can improve the generalization ability and accuracy of the model, but it will increase the training time and computing cost.

[0094] Fitting the input and output of the braking system using the offline dataset of the braking system.

[0095] (2) Simulating to obtain the offline data of the drive system.

[0096] Obtain the torque - speed curve of the drive motor from the technical manual provided by the manufacturer of the drive motor, establish a drive system simulation model in the simulation system based on this curve, and collect data in time series.

[0097] The torque - speed relationship of the drive motor is affected by vehicle load, road conditions, temperature changes, etc., and it is difficult to accurately describe through a static model. The present invention uses an LSTM network for modeling and utilizes its memory ability for time - series data to capture the dynamic characteristics of the drive system.

[0098] Configure the vehicle simulation running state parameters in MATLAB. During the running process of the simulated vehicle, collect the motor target vehicle speed, motor speed, dq - axis current components, dq - axis voltage components, and motor driving torque;

[0099] The simulation running state parameters include drive motor parameters, flux linkage parameters, and environmental parameters;

[0100] The motor parameters include the number of pole pairs, stator resistance, stator inductance, peak flux linkage, magnetic saturation curve, and back - electromotive force coefficient;

[0101] The flux linkage parameters include DC bus voltage, rated output current, overload capacity, and efficiency curve;

[0102] The environmental parameters include road surface gradient and road surface adhesion coefficient;

[0103] The pairs of data of the motor target vehicle speed, motor speed, dq - axis current components, dq - axis voltage components, and motor driving torque form the sample of the offline dataset of the drive system.

[0104] Obtain the offline dataset of the drive system for training the basic model of the drive system and reducing the cost of real - vehicle testing.

[0105] (3) Pre - process the offline data of the braking system and the drive system to obtain the offline dataset of the braking system and the offline dataset of the drive system.

[0106] Pre - processing the offline data includes:

[0107] Perform low - pass filtering on the offline datasets of the braking system and the drive system respectively to remove high - frequency noise, and obtain the sample data of the braking system and the drive system after noise removal;

[0108] The sample data of the braking system and the driving system after noise removal are respectively normalized. The normalized sample data is suitable for the training of the neural network model, and the preprocessed offline datasets of the braking system and the driving system are obtained accordingly.

[0109] Step S12: Based on the offline dataset of the braking system, train an RBF neural network model to obtain the basic model of the actuator (braking system).

[0110] The braking system exhibits significant non-linear characteristics, and as the actuator of the braking system wears, the braking performance gradually changes. To improve the adaptability of braking control, the present invention uses an RBF neural network model to model the braking system in order to obtain the basic model of the braking system and fit the non-linear logical relationship between the input and output of the braking system (braking torque - braking percentage). RBF can accurately fit the non-linear curve of the braking system with fewer parameters, and the training and inference calculations are efficient, and it is suitable for online adjustment in subsequent steps.

[0111] The offline dataset of the braking system is used for the training of the basic model of the actuator (braking system).

[0112] When the actuator is the braking system, the sample data in the actuator offline dataset is the braking torque, and the sample label is the braking percentage; the actuator basic model is constructed based on the RBF neural network model.

[0113] When the actuator is the driving system, the sample data of the actuator offline dataset includes the motor target vehicle speed, motor speed, dq-axis current components, and dq-axis voltage components, and the sample label is the motor driving torque; the actuator basic model is constructed based on the LSTM neural network model.

[0114] Exemplarily, the present invention uses the Kernel Ridge Regression (KRR) model and selects the RBF kernel (radial basis function kernel) as the kernel function.

[0115] After being trained with the offline dataset of the braking system and through model parameter tuning, exemplarily, the final parameter settings of the braking system basic model are as follows:

[0116] The regularization parameter alpha = 0.1;

[0117] The width parameter gamma of the RBF kernel = 5;

[0118] The specific form of the kernel function is the Gaussian function.

[0119] The input of the RBF neural network model is the expected braking torque, which is the expected control target; the output is the braking percentage, which is the output control quantity.

[0120] The expression of the RBF neural network model is as follows:

[0121] P b = f RBF (T b ; θ b ) Formula (2)

[0122] Wherein, P b is the braking percentage, which is the output of the RBF neural network model and is used to control the execution force of the braking system. T b is the applied braking torque, which is the input of the RBF neural network model and is the torque that the braking system is expected to generate. θ b are all the parameters of the RBF neural network model, including parameters such as the center vector, width parameter, and output weight.

[0123] During the operation of the driverless vehicle system, the in-vehicle chassis domain control computer records the braking percentage P b issued by the braking system, the required braking torque T b and the brake caliper pressure P c .

[0124] Exemplarily, the in-vehicle chassis domain control computer is a computer with the Linux operating system and is responsible for the motion control at the whole vehicle level of the driverless vehicle. The NVIDIA Orin GPU is used for neural network inference and interacts with the actuator through the CAN (Controller Area Network) bus.

[0125] The brake caliper pressure has a linear relationship with the braking torque and can be converted into the actual braking torque:

[0126]

[0127] Wherein, is the actual braking torque; kappa is a proportionality coefficient determined by the characteristics of the vehicle braking system, and P c is the actual pressure of the brake caliper.

[0128] Regarding kappa: for different types of driverless vehicles, the k value is different because of the different brake calipers and installation conditions, and this parameter can be obtained from the manufacturer's technical manual.

[0129] The physical quantity transfer of the driverless vehicle is: the braking system issues the braking percentage → the brake booster system pressurizes → the brake caliper pressure, and the caliper friction brakes → the braking torque.

[0130] Taking the braking percentage P b as the label and the actual braking torque as the input of the RBF neural network model to construct an offline dataset Store it in the buffer area of the vehicle chassis domain control computer.

[0131] If the average prediction error e within the recent preset period window (e.g., the most recent 5 s) b exceeds the set threshold, trigger online incremental training to optimize the RBF network parameters.

[0132]

[0133] where, e b,i is the prediction error at the i-th time point within the period window, are the actual torque and predicted torque at the i-th time point respectively, and N is the number of data points within one period time window.

[0134]

[0135] where, is the RBF network parameter at the (t + 1)-th iteration, is the RBF network parameter at the t-th iteration, η is the learning rate, which is a hyperparameter and determines the step size of parameter update, represents the gradient of the loss function L with respect to the parameter ; is the braking percentage predicted by the neural network, is the actual braking percentage;

[0136] Exemplarily, for example, when the RBF neural network model fits the input-output characteristics of the braking system, the initial prediction error of the offline dataset is 0.12%.

[0137] For the training of the basic model of the braking system, the loss function L1 is as follows:

[0138]

[0139] where, α1, β1 are the weight hyperparameters of the RBF neural network model, and α1 + β1 = 1, which are used to adjust the contributions of MSE (Mean Squared Error) and MAE (Mean Absolute Error).

[0140] Use the offline dataset of the braking system (divided into a test set and a validation set according to a ratio of 8:2) to train the RBF neural network model. After training multiple epochs; at the end of each epoch, record the MAE and MSE on the training set and the validation set; select the epoch with the best MAE and MSE on the validation set, and retain the structure and parameters of the RBF neural network model of this epoch as the basic model of the braking system.

[0141] Exemplarily, 10 ecophs were trained, and the MAE and MSE for each epoch are shown in Table 1:

[0142] Table 1: MAE and MSE for training each epoch

[0143] Epoch Training MAE Training MSE Validation MAE Validation MSE 1 0.12 0.0144 0.15 0.0225 2 0.10 0.0100 0.13 0.0169 3 0.08 0.0064 0.11 0.0121 4 0.07 0.0049 0.10 0.0100 5 0.06 0.0036 0.09 0.0081 6 0.05 0.0025 0.08 0.0064 7 0.05 0.0025 0.08 0.0064 8 0.06 0.0036 0.09 0.0081 9 0.07 0.0049 0.10 0.0100 10 0.08 0.0064 0.11 0.0121

[0144] In the above example, the MAE and MSE of the validation set in the 6th epoch are both the lowest (MAE = 0.08, MSE = 0.0064). Therefore, the structure and parameters of the model in the 6th epoch are selected as the basic model of the braking system.

[0145] In this step, a basic model of the braking system based on the RBF neural network model is obtained.

[0146] Step S13: Based on the offline dataset of the drive system, train an LSTM neural network model to obtain a basic model of the actuator (drive system).

[0147] The torque-speed relationship of the drive motor is affected by factors such as load, temperature, and road conditions, and has strong dynamic characteristics. It is difficult to use traditional fixed-parameter modeling. The present invention uses an LSTM neural network to model the drive motor. LSTM is suitable for time series modeling and can capture the time-varying characteristics of the speed-torque relationship and adapt to different road conditions and vehicle load changes.

[0148] The offline dataset of the drive system is divided into a training set and a validation set according to a ratio of 8:2, and the sample data are all sampled during continuous operation.

[0149] Exemplarily, the LSTM model structure in the present invention is LSTM-FNN (Feedforward Neural Network) - output layer.

[0150] Specifically: 3 layers of LSTM, with 128 hidden units in each layer, the input window length is 64, which can capture historical features within 6.4 seconds; FNN has 4 layers; the training optimization algorithm uses the Adam algorithm, the learning rate is 0.001, dropout is 0.1, and L2 regularization is used to avoid overfitting.

[0151] The time series input into the LSTM neural network model of the drive system includes: the target vehicle speed v d ; the historical motor torque T at the previous k moments m ; the motor speed ω m ; the dq-axis current components i d , i q ; the dq-axis voltage components voltage u d , u q .

[0152] The application scenario of the drive system is the speed tracking control of an autonomous vehicle. The goal is to keep the vehicle speed at the target vehicle speed of xx m / s. In terms of the hardware output, it means that the drive system outputs a torque of xx magnitude, and the vehicle maintains the target vehicle speed. The obtained effect is to issue the most reasonable drive torque based on the target vehicle speed to maintain the target vehicle speed.

[0153] Strictly speaking, the drive torque-speed characteristic of an autonomous vehicle is not only related to the drive system. This characteristic is also related to the vehicle load and the road surface conditions (bumpiness, adhesion coefficient, slope). However, this characteristic is most closely related to the drive system. Since the load and road surface conditions obviously change rapidly according to the working conditions, a time series is used to input into the LSTM neural network model to adapt to this dynamic characteristic.

[0154] The data of the drive motor controller is collected in real time through the CAN bus of the domain controller, including historical torque, motor speed, and the motor dq current components i d ,i q 、and the motor dq voltage components u d ,u q .

[0155] Record the input data {(v d , T m , ω m , i d , i q , u d , u q )} and the actual executed real torque and store them in the buffer area of the in-vehicle chassis domain control computer.

[0156] Output the optimal torque command for the drive motor and send it as the control quantity to the drive system.

[0157] The expression of the LSTM neural network model is as follows:

[0158]

[0159] Among them, is the optimal torque command, which is the output of the LSTM neural network model and is used to control the output torque of the drive motor; f LSTM (.) is the LSTM neural network model, which is used to capture the dynamic characteristics of the drive motor; v d is the target vehicle speed, which is one of the inputs of the LSTM neural network model and represents the speed that the autonomous vehicle is expected to reach; timeseq(.) represents the time series input; θ m is the LSTM network parameter, including the weight matrix and the bias vector.

[0160] For v d, many sensors on the driverless vehicle can calculate speed. Exemplarily, it mainly relies on the accelerometer-based vbox to provide real-time speed. Additionally, speed v can also be obtained through Kalman filter data fusion based on wheel speed d .

[0161] The prediction error e of the LSTM neural network model m , is calculated as follows:

[0162]

[0163] where, is the actually executed driving torque, is the driving torque predicted by the LSTM model. The prediction accuracy of the LSTM model is evaluated by calculating the mean absolute error MAE between the predicted value and the true value.

[0164] Exemplarily, the initial prediction threshold for training the basic model of the drive system with the offline dataset is set to 0.3%;

[0165] Optimize the LSTM network parameters:

[0166]

[0167] where: are the updated network weights and the network weights before update respectively; η is the learning rate of the LSTM model; is the torque value predicted by the LSTM.

[0168] For the training of the basic model of the drive system, the loss function L2 is as follows:

[0169]

[0170] where, α2, β2 are the weight hyperparameters of the LSTM neural network model, and α2 + β2 = 1.

[0171] Use the offline dataset of the drive system (divided into a test set and a validation set according to a ratio of 8:2) to train the LSTM neural network model. After training multiple epochs; after each epoch ends, record the MAE and MSE on the training set and the validation set; select the epoch with the best MAE and MSE on the validation set, and retain the structure and parameters of the LSTM neural network model of this epoch as the basic model of the drive system.

[0172] Train an RBF neural network model based on the offline dataset of the braking system to obtain the basic model of the braking system; train an LSTM neural network model based on the offline dataset of the drive system to obtain the basic model of the drive system; obtain a reasonable basic model structure and parameters to ensure that the neural network has good fitting ability and reduce the burden of subsequent online incremental training.

[0173] The function of step S1 is to obtain and preprocess the offline datasets of the braking system and drive system of the unmanned vehicle, and train the actuator basic model, including the basic models of the braking system and drive system of the unmanned vehicle, to provide a basis for subsequent actuator adaptive control.

[0174] Step S2, specifically:

[0175] If the average prediction error e of the trained braking system basic model for predicting the braking percentage b exceeds the initial prediction threshold e h then trigger online incremental training;

[0176] During the operation of the vehicle, online data continuously collects the operating state data of the motor through the CAN bus, including target vehicle speed, torque, rotational speed, current, voltage, etc.

[0177] Utilize the real-time data obtained during the operation of the unmanned vehicle for incremental learning to adapt to the wear and dynamic changes of the drive system, and continuously and real-time optimize the parameters of the RBF neural network model. The real-time data of the braking system is the same as the offline data, which will not be elaborated here.

[0178] Adopt an incremental training strategy, training 10 epochs every 1 second to ensure the efficiency and real-time nature of online updates.

[0179] Exemplarily, perform online incremental training on the real-time data for 10 epochs, optimize the parameters, and obtain a braking system adaptive model and a drive system adaptive model that are better than the basic model.

[0180] The actuator adaptive model outputs a control quantity and a prediction error;

[0181] When the actuator is the braking system, the output control quantity is the braking percentage, and the output prediction error is the braking prediction error;

[0182] When the actuator is the drive system, the output control quantity is the drive torque, and the output prediction error is the drive prediction error.

[0183] The initial prediction threshold of the braking system is set to 0.12%, and the initial prediction threshold of the drive system is set to 0.3%.

[0184] At the beginning of online incremental training, the prediction error of the actuator basic model is used as the initial prediction threshold eh ;

[0185] After each round of online incremental training, calculate the average error within the current cycle window to obtain the maximum average error in the historical cycle window;

[0186] If the average error within the current cycle window is different from the current predicted average error e window , then use the average error within the current cycle window to update the predicted average error e window ;

[0187] If the maximum average error in the historical cycle window is greater than the maximum average error e m , then use the maximum average error in the historical cycle window to update the maximum average error e m ;

[0188] Among them, the prediction error of the actuator adaptive model is three times that of the initial prediction threshold e h triple.

[0189] When performing online training, the learning rate of the drive system model is relatively large, on the order of 10 -4 , and the learning rate of the braking system model is relatively small, on the order of 10 -5 magnitude.

[0190] The prediction errors of the braking system and drive system adaptive models are three times the initial prediction threshold, and the prediction error of the drive system adaptive model is 0.36%; the prediction error of the braking system adaptive model is 0.9%.

[0191] Perform online incremental training on the braking system basic model and drive system basic model to obtain a braking system adaptive model and a drive system adaptive model that are superior to the basic models.

[0192] The function of step S2 is to use the data collected in real time during the operation of the unmanned vehicle to perform online incremental training on the basic model, optimize the model parameters, so as to adapt to the dynamic changes and wear of the actuator, and improve the prediction accuracy and control performance of the model.

[0193] Step S3, specifically:

[0194] The present invention adopts a feedforward-feedback cooperative control strategy, combines the output of the neural network model with the control method of PID control to implement an adaptive control algorithm.

[0195] This step is an adaptive control algorithm in which the actuator adaptive model is used as the feedforward and the PID controller is used as the feedback, and the two are combined. The feedback control uses PID control to compensate for the feedforward error. The proportional gain controls the error response speed, the integral gain eliminates the steady-state error, and the derivative gain improves the dynamic response.

[0196] The adaptive control quantity of the controlled object is obtained by dynamically weighted fusion of the feedback control quantity output by the PID controller and the feedforward control quantity, as follows:

[0197] u = λ·u ff +(1 - λ)·u fb Formula (13)

[0198] Wherein, u is the finally output adaptive control quantity; λ is the weight of the feedforward control quantity, u ff is the feedforward control quantity, 1 - λ is the weight of the feedback control quantity, u fb is the feedback control quantity;

[0199] λ is defined as follows:

[0200]

[0201] Wherein, e window 、e m 、e h are respectively the average prediction error of the actuator adaptive model within the current cycle window, the maximum average error in the historical cycle data, and the initial prediction threshold;

[0202] When the actuator is a braking system, the input of the PID controller is the braking torque; when the actuator is a driving system, the input of the PID controller is the target vehicle speed.

[0203] Exemplarily, error data within a running window is statistically analyzed, which is 5s for the driving system and 600s for the braking system; this is because the torque - speed characteristic changes rapidly, while the characteristic of the braking system changes slowly.

[0204] If the average prediction error e window in the window is greater than the initial prediction threshold e h , online incremental training is triggered. After each online incremental training, the initial prediction threshold e h is adjusted, the average prediction error in this window is used as the new initial prediction threshold e h , and the maximum average error e m is recorded.

[0205] There are two starting points: one is to avoid consuming resources due to frequent incremental training; the other is to transition more weights to the feedback control when the fitting model error is large. When e m is large, this will result in the same e window , and the feedforward - feedback control weight λ will be smaller, that is, the proportion of the feedforward is smaller. If e window is large, the feedforward weight is reduced to make the feedback control dominant to enhance control stability.

[0206] When adjusting the weight of the feedforward control amount, by calculating the prediction error, if the error is large, the weight of the feedforward control amount is reduced to make the feedback control dominant, and finally the optimization of the control amount is achieved.

[0207] There is already a PID control loop control method on existing unmanned vehicles. The starting point of the present invention is to add the controller adaptive model constructed based on the neural network model as a feedforward term without much modification. Almost no modification to the original PID control method is required. The controller adaptive model of the controlled objects (brake system and drive system) in the present invention is used as the feedforward term, and the feedforward-feedback gain is adjusted with the feedback term of the PID controller to obtain the control amounts (brake percentage and optimal drive torque) of the controlled objects (i.e., actuators).

[0208] If the actuator is a brake system, the feedforward control amount is as follows:

[0209] u ff =f RBF (T b ;θ b ) Formula (15)

[0210] Wherein, f NN (.) is the actuator adaptive model of the predicted feedforward amount; f RBF (.) is the RBF neural network model, and is the RBF network parameter;

[0211] If the actuator is a drive system, the feedforward control amount is as follows:

[0212] u ff =f LSTM (v d ,timeseq(ω m ,i d ,i q ,u d ,u q );θ m ) Formula (16)

[0213] Wherein, f LSTM (.) is the LSTM neural network model, v d is the target vehicle speed, timeseq(.) represents the time series input, ω m is the drive motor speed, i d 、i q are the dq-axis current components respectively, u d 、u q are the dq-axis voltage components respectively, θ m is the LSTM network parameter.

[0214] The PID controller outputs the feedback control amount as follows:

[0215]

[0216] Among them, u fb is the feedback control quantity, and e is the deviation between the input and the actual output of the PID controller; K p , K i , K d are respectively the proportional gain, integral gain, and derivative gain parameters of the PID controller; the input of the PID controller is the same as the input of the actuator adaptive model.

[0217] The feedback control quantity u fb is used to adjust the control instruction to reduce the error;

[0218] e is the error signal (the deviation between the target value and the actual execution value);

[0219] K p : Proportional gain, controlling the direct response of the error;

[0220] K i : Integral gain, eliminating the steady-state error;

[0221] K d : Derivative gain, improving the dynamic response and reducing overshoot.

[0222] As Figure 4 shown, the inputs of the PID controller and the actuator adaptive model are the same; the actuator adaptive model outputs the feedforward term - control quantity (including the braking percentage and driving torque) and the prediction error; the output of the PID controller is used as the feedback term; the feedforward term and the feedback term are weighted and fused to perform feedforward-feedback gain adjustment, and the control quantity u of the final controlled object is output and sent to the controlled object for execution.

[0223] Through the coordinated action of the inner loop and the outer loop, fast response and long-term optimization can be achieved, ensuring stable control performance in complex working conditions and long-term operation.

[0224] The inner loop is the control loop of the existing method, and a feedforward-feedback gain adjustment module is added on the basis of the existing method; the control period of the inner loop is shorter than that of the outer loop. Exemplarily, the control period of the inner loop is 10 ms, and the control period of the outer loop is 100 ms.

[0225] The control quantity predicted by the actuator adaptive model (for example, issuing the optimal torque command and braking percentage of the drive system) compensates for the dynamic characteristics of the actuator in advance; the feedback control uses the PID controller to calculate the error (the deviation between the target value and the actual execution value) and adjusts the control quantity to optimize the control accuracy.

[0226] Combining feedforward and feedback control forms an adaptive closed-loop control system that can not only quickly respond to input changes but also eliminate long-term steady-state errors.

[0227] Exemplarily, taking the target vehicle speed tracking as an example:

[0228] The target vehicle speed is, for example, 20 m / s; through the control algorithm of the actuator adaptive control model, a control quantity, that is, the driving torque, is sent to the drive system; at this time, the sensor will collect the actual vehicle speed, which is the feedback on the black line in the figure. With the feedback item, it can be known how much the current actual vehicle speed differs from the target vehicle speed. Then the PID controller can adjust the control quantity through the deviation, with the goal of making the deviation zero.

[0229] However, there will be a lag in feedback. Because the deviation is calculated based on the operating data collected at the previous time node, there is a lag of one sampling period from the current actual value. On the other hand, the dynamic response time of the system is also uncertain. For example, for a driving motor torque of 100 Nm, it also takes some time for the vehicle speed to increase from 0 to 20 m / s. Based on the deviation, the PID controller calculates the control output, which easily leads to overshoot or instability. Therefore, in the existing methods, it is often necessary to perform PID parameter tuning on the actuators of each type of autonomous vehicle in the autonomous vehicle fleet.

[0230] The feedforward is the control quantity output by the actuator adaptive model. For example, when the target vehicle speed is expected to be 20 m / s, under the current load and road conditions, it is calculated through the actuator adaptive model that it is appropriate to maintain the torque at 100 Nm. At this time, even without the feedback control loop, theoretically, the autonomous vehicle can be maintained at 20 m / s after a certain period of time only relying on the actuator adaptive model; however, the real-time in-loop feedback control has stronger robustness, so feedforward + feedback is a better choice. Based on the supplement of feedforward, when designing the PID controller, it can be more conservative, focusing on stability and reducing overshoot, and the output of feedforward will supplement the stability and reliability of the actuator.

[0231] Figure 2 Shows the overall architecture of the autonomous vehicle adaptive control system of the present invention. It includes an autonomous vehicle running in real time, a sensor and data acquisition device, an in-vehicle chassis domain control computer, and a computer actuator adaptive model and an actuator basic model. The sensor collects the real-time operating data of the autonomous vehicle and transmits it to the in-vehicle chassis domain control computer, calculates the control instruction through the adaptive control algorithm, and feeds it back to the autonomous vehicle to optimize the control performance.

[0232] As Figure 3 Shown in the architecture diagram. When the system runs, it collects online information in real time and updates the neural network model and adjusts the controller adaptive algorithm. This figure shows the adaptive control process of vehicle driving and braking, including the whole vehicle motion control, braking system, and driving motor control links.

[0233] The red line in the figure is the output of the neural network actuator adaptive model, and the yellow line is the specific online collected data information.

[0234] There are three lines in the drive system part:

[0235] (1) The torque command sent to the drive motor actuator;

[0236] (2) The rotational speed, voltage and current information of the drive motor is reported by the drive motor actuator to the vehicle chassis domain control computer through the CAN bus;

[0237] (3) Vehicle speed, this information is obtained through data fusion of the accelerometer and the rotational speeds of multiple wheels. The accelerometer is in an independent guidance control computer device. The vehicle chassis domain control computer also carries a gyroscope, Beidou, and Qianxun positioning devices, and establishes a TCP connection with the vehicle chassis domain control computer through the network port. The transmitted content includes vehicle position, speed, acceleration, attitude, angular acceleration, etc. In addition, wheel speed sensors (encoders) will be integrated inside the wheel hubs, and the wheel speed sensors will report the wheel speed information to the devices on the CAN bus through the CAN bus. In the vehicle chassis domain control computer where our algorithm is deployed, the current vehicle body speed will be obtained based on the data of the guidance control computer and the wheel speed information.

[0238] There are two lines in the braking system part:

[0239] (1) The braking percentage signal (PWM) sent to the brake booster;

[0240] (2) The pressure sensor signal at the brake caliper is linearly converted into the brake caliper pressure. A brake pressure sensor will be installed on the wheel brake caliper and will also be connected to the CAN bus to regularly report the brake caliper pressure data.

[0241] The black line is the relevant signal flow or physical quantity flow.

[0242] The function of step S3 is to dynamically weight and fuse the feedforward control of the actuator adaptive model and the feedback control of the PID controller, realize the adaptive control of the braking system and the drive system on different unmanned vehicles, and optimize the control accuracy and system stability of the unmanned vehicle.

[0243] Embodiment 2:

[0244] Another embodiment of the present invention discloses a heterogeneous actuator adaptive control system for an unmanned vehicle based on neural network modeling, so as to realize a heterogeneous actuator adaptive control method for an unmanned vehicle based on neural network modeling in Embodiment 1. The specific implementation manners of each module refer to the corresponding descriptions in Embodiment 1.

[0245] Such as Figure 6As shown in the figure, the system includes an offline data acquisition and basic model training module M1, an online data acquisition and model incremental training module M2, and a feedforward-feedback fusion adaptive control module M3.

[0246] The offline data acquisition and basic model training module M1 is used to obtain the offline data of the actuators of the unmanned vehicle and perform preprocessing to obtain an offline actuator dataset; based on the offline actuator dataset, a basic actuator model is trained.

[0247] The online data acquisition and model incremental training module M2 is used to, during the operation of the unmanned vehicle, when the average prediction error of the control amount of the actuator within a preset period window exceeds a preset threshold, perform online incremental training on the basic actuator model using the real-time acquired online actuator data to obtain an adaptive actuator model.

[0248] The feedforward-feedback fusion adaptive control module M3 is used to use the control amount output by the adaptive actuator model as the feedforward control amount, and perform dynamic weighted fusion of the feedforward control amount and the feedback control amount to obtain the adaptive control amount of the actuator, and send it to the actuator for execution.

[0249] In summary, an adaptive control method and system for heterogeneous actuators of an unmanned vehicle based on neural network modeling according to an embodiment of the present invention has the following beneficial effects:

[0250] 1. Through the online incremental training mechanism of the present invention, the adaptive model can adapt to the dynamic changes and wear of the actuator in real time, ensuring the control accuracy and stability during long-term operation, and the model has strong self-adaptability.

[0251] 2. Based on rich offline data training and online data update of the present invention, the model has good generalization ability for actuators of different manufacturers and different batches, reducing the manual debugging cost.

[0252] 3. By combining the feedforward-feedback cooperative control strategy of the present invention, using the predicted control amount of the neural network model as the feedforward control amount and the feedback control amount of the PID controller for real-time dynamic weighted fusion, the control accuracy is significantly improved and the error is reduced.

[0253] 4. Through the efficient neural network inference and adaptive control algorithm of the present invention, it is ensured that data acquisition, adaptive model prediction, and final adaptive control amount sending and execution are completed within a preset control period, meeting the real-time control requirements.

[0254] 5. By dynamically adjusting the feedforward and feedback control weights, the present invention can maintain stable control performance and enhance the robustness of the system when facing complex working conditions and actuator performance changes.

[0255] 6. Based on the predicted control quantity of the existing PID controller, the present invention only adds the gain adjustment of the feedforward and feedback control quantities, with very small modifications, thereby reducing the modification cost.

[0256] Those skilled in the art can understand that all or part of the processes for implementing the methods of the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a magnetic disk, an optical disk, a read-only memory, or a random access memory, etc.

[0257] As described above, the above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. An adaptive control method for heterogeneous actuators of an autonomous vehicle based on neural network modeling, characterized in that, Including: Obtain the actuator offline data of the driverless vehicle and perform preprocessing to obtain the actuator offline dataset; Based on the actuator offline dataset, train to obtain the actuator basic model; During the operation of the driverless vehicle, when the average prediction error of the control amount of the actuator within the preset cycle window exceeds the preset threshold, use the actuator online data collected in real time to perform online incremental training on the actuator basic model to obtain the actuator adaptive model; Take the control amount output by the actuator adaptive model as the feedforward control amount, and perform dynamic weighted fusion of the feedforward control amount and the feedback control amount to obtain the adaptive control amount of the actuator, and send it to the actuator for execution.

2. The method according to claim 1, wherein When the actuator is a braking system, the sample data in the actuator offline dataset is the braking torque, and the sample label is the braking percentage; the actuator basic model is constructed based on the RBF neural network model.

3. The method according to claim 1, wherein When the actuator is a drive system, the sample data in the actuator offline dataset includes the target vehicle speed, motor speed, dq-axis current components, and dq-axis voltage components, and the sample label is the motor drive torque; the actuator basic model is constructed based on the LSTM neural network model.

4. The method according to any one of claims 1 to 3, characterized in that, Based on the feedback control amount output by the PID controller and the feedforward control amount, perform dynamic weighted fusion to obtain the adaptive control amount of the controlled object, as follows: u = λ·u ff +(1 - λ)·u fb Among them, u is the final output of the adaptive control quantity; λ is the weight of the feedforward control quantity, u ff is the feedforward control quantity, 1 - λ is the weight of the feedback control quantity, u fb is the feedback control quantity; λ is defined as follows: Among them, e window , e m , e h are respectively the average prediction error of the actuator adaptive model within the current cycle window, the maximum average error in the historical cycle data, and the initial prediction threshold; When the actuator is a braking system, the input of the PID controller is the braking torque; when the actuator is a drive system, the input of the PID controller is the target vehicle speed.

5. The method according to claim 4, wherein If the actuator is a braking system, the feedforward control amount is as follows: u ff = f RBF (T b ; θ b ) Among them, f RBF (.) is the RBF neural network model used to predict the feedforward quantity for brake actuator control, T b is the target braking torque, and θ b are the RBF network parameters; If the actuator is a drive system, the feedforward control amount is as follows: u ff = f LSTM (v d , timeseq(ω m , i d , i q , u d , u q ); θ m ) Among them, f LSTM (.) is an LSTM neural network model used to predict the feedforward quantity for driving actuator control, v d is the target vehicle speed, timeseq(.) represents the time series input, ω m is the driving motor speed, i d , i q are the dq-axis current components respectively, u d , u q are the dq-axis voltage components respectively, θ m are the LSTM network parameters.

6. The method according to claim 4, wherein The actuator adaptive model outputs the control amount and the prediction error; If the actuator is a braking system, the output control amount is the braking percentage, and the output prediction error is the braking prediction error; If the actuator is a drive system, the output control amount is the motor drive torque, and the output prediction error is the drive prediction error.

7. The method according to claim 5, wherein At the start of online incremental training, the prediction error of the actuator base model is used as the initial prediction threshold e h ; After each round of online incremental training, calculate the average error within the current cycle window to obtain the maximum average error in the historical cycle window; If the average error within the current cycle window is different from the current predicted average error e window then update the average predicted error e with the average error within the current cycle window window ; If the maximum average error in the historical period window is greater than the maximum average error e m , then update the maximum average error e with the maximum average error in the historical period window m .

8. The method according to claim 2, characterized in that, Based on the braking percentage - brake caliper pressure curve of different braking systems, perform simulation fitting to obtain multiple groups of braking torque and braking percentage data pairs, as follows: Among them, T b is the finally generated braking torque, r is the effective braking radius of the brake, G is the reduction ratio of the transmission system, I is the input current of the motor, k I is the conversion coefficient between current and torque, k η is the conversion coefficient from the braking percentage to current, P b is the braking percentage, A is the piston area of the hydraulic master cylinder, and P is the brake caliper pressure; The braking torque and braking percentage data pairs form the sample of the braking system offline dataset.

9. The method according to claim 3, wherein Configure the vehicle simulation operation state parameters in MATLAB, and during the simulation vehicle operation, collect and obtain the motor target vehicle speed, motor speed, dq-axis current components, dq-axis voltage components, and motor drive torque; The simulation operation state parameters include drive motor parameters, flux linkage parameters, and environmental parameters; The motor parameters include the number of pole pairs, stator resistance, stator inductance, peak flux linkage, magnetic saturation curve, and back electromotive force coefficient; The flux linkage parameters include the DC bus voltage, rated output current, overload capacity, and efficiency curve; The environmental parameters include the road surface gradient and road surface adhesion coefficient; The motor target vehicle speed, motor speed, dq-axis current components, dq-axis voltage components, and motor drive torque data pairs form the sample of the drive system offline dataset.

10. An adaptive control system for heterogeneous actuators of an autonomous vehicle based on neural network modeling, characterized in that, Including: An offline data collection and basic model training module, which is used to obtain the actuator offline data of the driverless vehicle and perform preprocessing to obtain an actuator offline data set; Based on the actuator offline data set, an actuator basic model is trained; An online data collection and model incremental training module, which is used to, during the operation of the driverless vehicle, when the average prediction error of the control amount of the actuator within a preset cycle window exceeds a preset threshold, perform online incremental training on the actuator basic model using the actuator online data collected in real time to obtain an actuator adaptive model; A feedforward-feedback fusion adaptive control module, which is used to use the control amount output by the actuator adaptive model as a feedforward control amount, dynamically weight and fuse the feedforward control amount and the feedback control amount to obtain an adaptive control amount of the actuator, and send it to the actuator for execution.

Images