An on-line calibration device and method for control parameters of an engine control system

Abstract

The application discloses an online calibration device and method for control parameters of an engine control system, and relates to the technical field of engine control. The device comprises: an upper computer connected with a controller through a communication module; the controller is connected with a direct current motor driving module and an electronic throttle body; the direct current motor driving module is connected with the electronic throttle body; the upper computer is used for monitoring the state of each input and output pin of the controller, receiving calibration index measurement parameters, calculating calibration sampling point data by using an online calibration method for control parameters of an engine control system, and sending the calibration sampling point data to the controller; the controller is used for sending motor operation parameters to the direct current motor driving module according to the calibration sampling point data by using a PID algorithm, and receiving a throttle opening degree signal returned by the electronic throttle body; and the direct current motor driving module is used for controlling the work of the electronic throttle body according to the motor operation parameters. The application shortens the online calibration period of the control parameters of the engine control system.

Description

Technical Field

[0001] This invention relates to the field of engine control technology, and in particular to an online calibration device and method for control parameters of an engine control system. Background Technology

[0002] Engine control systems exhibit strong nonlinearity, strong coupling, and strong dynamic characteristics, making them difficult to describe with a simple mathematical expression, thus hindering the optimization of control parameters. Engine control parameter calibration primarily requires optimizing the controller parameters of the relevant engine control subsystem based on its control requirements, combined with experimental data and statistical optimization methods. Furthermore, statistical studies of controller applications in industry indicate that approximately 90% of products use PID controllers; therefore, this invention's discussion of engine control parameter calibration mainly focuses on the case where PID controllers are used as subsystem controllers in the engine control system.

[0003] Traditional control parameter calibration methods rely on comprehensive analysis of the interactions between control parameters under different operating conditions and repeated debugging tests under various conditions to determine the optimal control parameter settings that meet specific requirements. However, this method faces significant implementation difficulties. First, the constantly changing test conditions make the calibration process difficult to calibrate accurately, and common test bench systems struggle to simulate real dynamic operating environments, leading to discrepancies between test results and actual application scenarios. Second, frequent dynamic tests not only impact the maintenance and lifespan of the test bench equipment but also significantly reduce the repeatability of the calibration process due to operational complexity. These control parameter tuning methods are based on assumptions about the controlled object model and desired outputs, extracting system characteristics through analysis or graphical methods to select controller settings. Because the control parameter calibration is based on assumptions, further adjustments are almost always required during actual use, and the tuned control parameters perform poorly when system characteristics do not conform to the relevant assumptions.

[0004] Online calibration methods for control parameters can be divided into model-based online calibration methods and model-free online calibration methods. Model-free online calibration methods do not require real-time modeling of the controlled object during the calibration process. These methods generally have characteristics such as long parameter calibration cycles and poor performance when operating conditions and environments change significantly.

[0005] Currently, patent literature on online automatic calibration systems for engine control parameters generally focuses on the structure and function of the automatic calibration system, including key components such as the control unit, acquisition unit, and optimization unit. The main function of the control unit is to output the control parameters of the actuators to adjust the engine's operating state. The acquisition unit is responsible for monitoring and recording the engine's performance response, such as key indicators like torque, fuel consumption rate, and emissions. Furthermore, the optimization unit analyzes and compares the engine's performance under different control parameter configurations to identify and determine the optimal control parameter configuration. The design concept of this automatic calibration system is based on the need to improve engine performance and efficiency. By integrating control, acquisition, and optimization functions, the system can achieve real-time adjustment and optimization of engine parameters, which is particularly important for meeting increasingly stringent fuel efficiency and emission standards. The application of automatic calibration technology not only improves the accuracy and efficiency of the engine adjustment process but also paves the way for further intelligentization and automation of engine management systems. These patents demonstrate the application prospects and potential of automatic calibration technology in modern engine technology, emphasizing the core role of integrated automated systems in improving engine performance. However, the calibration cycle is relatively long and involves many sampling points, making it difficult to meet the speed requirements of online calibration.

[0006] In modern engineering applications, especially in the online calibration of engine control systems, optimization algorithms play a crucial role. The two main types of commonly used optimization algorithms are metaheuristic algorithms and extreme value search algorithms. Each of these algorithms has its own advantages and disadvantages, and is suitable for different application scenarios and requirements.

[0007] Metaheuristic algorithms, such as genetic algorithms, particle swarm optimization (PSO), and simulated annealing, are a class of algorithms that find optimal solutions by mimicking natural processes. Their main advantage lies in their strong global search capability, effectively escaping local optima to find the global optimum. This makes metaheuristic algorithms particularly suitable for solving complex optimization problems with large problem spaces and numerous local minima. However, these algorithms also have some drawbacks, such as high computational cost, potentially slow convergence speed, and complex parameter settings that require extensive experimentation.

[0008] Extreme value search algorithms, including the golden section search and binary search, are methods that locate the optimal solution by gradually narrowing the search range. The advantages of these algorithms are their simplicity of implementation and fast convergence speed, especially when the objective function is unimodal and continuous, enabling them to quickly and effectively find extreme points. However, the main drawback of extreme value search algorithms is their reliance on the unimodality of the problem; they perform poorly on multimodal problems and are prone to getting trapped in local optima.

[0009] In practical applications, the above algorithm has drawbacks when solving non-convex optimization calibration problems of control parameters, such as long solution time, large number of sampling points required, and inability to find the global optimum. Summary of the Invention

[0010] The purpose of this invention is to provide an online calibration device and method for engine control system parameters, which can shorten the online calibration cycle of control parameters.

[0011] To achieve the above objectives, the present invention provides the following solution:

[0012] An online calibration device for control parameters of an engine control system, the online calibration device comprising: an electronic throttle body, a DC motor drive module, a controller, a communication module, and a host computer;

[0013] The host computer is connected to the controller via the communication module; the controller is connected to both the DC motor drive module and the electronic throttle body; the DC motor drive module is connected to the electronic throttle body.

[0014] The host computer is used to monitor the status of each input and output pin of the controller, receive calibration index measurement parameters, and calculate calibration sampling point data based on the calibration index measurement parameters using the online calibration method for the engine control system control parameters, and send the calibration sampling point data to the controller; the calibration index measurement parameters are calculated by the controller based on the throttle opening signal of the electronic throttle body; the online calibration method for the engine control system control parameters is a grid adaptive Bayesian online calibration algorithm;

[0015] The controller is used to send motor operating parameters to the DC motor drive module based on the calibration sampling point data using a PID algorithm, and to receive the throttle opening signal returned by the electronic throttle body; the motor operating parameters include motor speed and motor rotation direction;

[0016] The DC motor drive module is used to control the operation of the electronic throttle body according to the motor operating parameters.

[0017] Optionally, the host computer includes an edge computing device and a controller host computer;

[0018] Both the edge computing device and the host computer of the controller are connected to the controller.

[0019] The edge computing device is used to receive calibration index measurement parameters, and calculate calibration sampling point data based on the calibration index measurement parameters using a grid adaptive Bayesian online calibration algorithm, and send the calibration sampling point data to the controller;

[0020] The host computer of the controller is used to monitor the status of each input and output pin of the controller.

[0021] An online calibration method for control parameters of an engine control system, applied to the aforementioned online calibration device for control parameters of an engine control system, the online calibration method comprising:

[0022] Define the feasible region of the control parameters and the initial sampling point values ​​of the control parameters; the control parameters are the proportional parameters, derivative parameters, and integral parameters of the PID algorithm;

[0023] Based on the mean function, covariance function, and the initial sampling point values ​​of the control parameters, the next sampling point value of the control parameters is determined by applying a prior Gaussian process model and a Bayesian sampling strategy.

[0024] Based on the next sampling point value of the control parameters, a posterior Gaussian process model is applied to determine the value of the objective function; the objective function is the integral of the absolute value of the difference between the throttle opening of the electronic throttle body under the control parameters and the target opening.

[0025] Determine whether the value of the objective function satisfies the preset convergence condition;

[0026] When the value of the objective function does not meet the preset convergence condition, the posterior Gaussian process model is updated by applying the method based on maximizing the logarithmic marginal probability, and the next sampling point value of the control parameter is input into the updated posterior Gaussian process model to obtain the updated value of the objective function, and the process returns to the step "determine whether the value of the objective function meets the preset convergence condition".

[0027] When the value of the objective function meets the preset convergence condition, the value of the next sampling point of the control parameter is used as the online calibration result.

[0028] Optionally, the mean function is μ(x;c)≡c;

[0029] Where x is the calibration parameter, i.e., the three parameters of the PID; c is a constant.

[0030] Optionally, the covariance function is:

[0031]

[0032] Among them, cov f For the covariance of different calibration parameters, x i x j For the i-th and j-th components of the calibration parameters, l1, ..., l D It is the kernel length along each dimension, α is the shape parameter; D represents the data dimension; This represents the variance of the calibration parameters.

[0033] Optionally, the objective function is:

[0034]

[0035] Where μ(x) is the mean function, κ(x, x′) is the covariance kernel function, f(x) represents the objective function value, x and x′ are the data points of two sets of calibration parameters, E represents the mean operation, and R p It represents a p-dimensional real space.

[0036] Optionally, the updated posterior Gaussian process model is:

[0037]

[0038] Where, x * f(x) * ) represent the calibration parameter data value and the corresponding objective function value for the next sampling point, respectively, μ * and κ * These are the updated mean function matrix and covariance kernel function matrix, respectively. X and Y are the calibration parameters of the previous sampling point and the obtained objective function data, respectively. μ(x * ) represents the mean of the calibration parameters for the next sampling point.

[0039] Optionally, the Bayesian sampling strategy is obtained by optimizing the acquisition function using a grid adaptive search algorithm.

[0040] Optionally, the sampling function is a sampling function with a lower confidence bound.

[0041] Optionally, the preset convergence condition is: Any one of them;

[0042] Where, Δ mesh M represents the actual mesh size tolerance. size The grid size tolerance value determines the accuracy of the calibration parameters; y n and y n+1 F represents the objective function for the nth and (n+1)th evaluations. c The minimum improvement in the objective function depends on the accuracy requirements of the calibration index; T no I represents the actual number of iterations in which the objective function value does not change significantly. s F represents the maximum number of iterations at which the objective function value does not change significantly. eval and T represent the number of evaluations of the actual objective function and the number of actual algorithm iterations, respectively, and F max with I max These are the maximum number of evaluations of the objective function and the maximum number of iterations of the algorithm, respectively; C vThe required threshold is to be set.

[0043] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0044] This invention provides an online calibration device and method for control parameters of an engine control system. Specifically, it is an automated online calibration device for control parameters of the electronic throttle subsystem in an engine control system. Based on Bayesian inference and Bayesian decision theory, the sampling process for online calibration of electronic throttle control parameters is guided. Under fixed throttle following conditions, the control parameters are continuously updated using a Bayesian optimization algorithm, and the calibration objective function is solved to achieve automated online calibration of the control parameters. This invention uses a Bayesian optimization algorithm in the online calibration process, employing a grid adaptive algorithm to optimize the sampling function to solve the non-convex optimization problem in the calibration process. A Gaussian process model is used as a surrogate model, which greatly shortens the online calibration cycle of the control parameters and reduces the number of sampling points required for calibration. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of the overall scheme for physical testing and verification provided in Embodiment 1 of the present invention;

[0047] Figure 2 This is a schematic diagram of the grid adaptive optimization algorithm process;

[0048] Figure 3 This is a schematic diagram of the online calibration algorithm simulation verification process;

[0049] Figure 4 A schematic diagram of the Bayesian optimization process;

[0050] Figure 5 This is a schematic diagram of the grid adaptive algorithm process;

[0051] Figure 6 Flowchart of the main control program for the electronic throttle valve test platform;

[0052] Figure 7 This diagram illustrates the process of online calibration of control parameters using the online calibration method provided by this invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] The purpose of this invention is to provide an online calibration device and method for engine control system parameters, aiming to shorten the online calibration cycle of control parameters.

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] Example 1

[0057] like Figure 1 As shown, the online calibration device for the control parameters of the engine control system in this embodiment includes: an electronic throttle body, a DC motor drive module, a controller, a communication module, and a host computer.

[0058] The host computer is connected to the controller via the communication module; the controller is connected to the DC motor drive module and the electronic throttle body respectively; the DC motor drive module is connected to the electronic throttle body.

[0059] The host computer is used to monitor the status of each input and output pin of the controller, receive calibration index measurement parameters, and calculate calibration sampling point data based on the calibration index measurement parameters using the online calibration method for the engine control system control parameters, and send the calibration sampling point data to the controller; the calibration index measurement parameters are calculated by the controller based on the throttle opening signal of the electronic throttle body; the online calibration method for the engine control system control parameters is a grid adaptive Bayesian online calibration algorithm.

[0060] The controller is used to send motor operating parameters to the DC motor drive module based on the calibration sampling point data using a PID algorithm, and to receive the throttle opening signal returned by the electronic throttle body; the motor operating parameters include motor speed and motor rotation direction.

[0061] The DC motor drive module is used to control the operation of the electronic throttle body according to the motor operating parameters.

[0062] In one specific implementation, the host computer includes an edge computing device and a controller host computer.

[0063] Both the edge computing device and the host computer of the controller are connected to the controller.

[0064] The edge computing device is used to receive calibration index measurement parameters, and calculate calibration sampling point data by applying a grid adaptive Bayesian online calibration algorithm based on the calibration index measurement parameters, and send the calibration sampling point data to the controller.

[0065] The host computer of the controller is used to monitor the status of each input and output pin of the controller.

[0066] In practical applications, this invention relates to the field of online calibration technology for engine control system control parameters, and particularly to an online calibration device for control parameters based on grid adaptive Bayesian optimization. This invention utilizes a grid adaptive Bayesian optimization algorithm incorporating extended Kalman filtering as the online calibration algorithm for an online calibration device for engine control system control parameters. The device includes an electronic throttle body, an L298N DC motor drive board, a MicroAutoBox II controller, an engine controller host computer, a VectorVN1630ACAN communication device, and a PC or edge computing device.

[0067] The computer algorithm software platform MATLAB / SIMULINK runs on a PC or edge computing device. It communicates with the MicroAutoBox II controller via CAN communication technology to transmit parameters related to calibration index calculation (message IDs are set to 90 and 91 respectively). At the same time, it sends the latest sampling points through the grid adaptive Bayesian online calibration algorithm to control the electronic throttle system and collect its calibration index for the next cycle. The MicroAutoBox II platform outputs control signals for the H-bridge drive circuit by running discrete PID control and collects signals from the throttle body sensor.

[0068] The electronic throttle body includes a throttle valve plate, a DC drive motor, a return spring, and a Hall effect position sensor. The L298N drive board consists of four MOSFETs, which are divided into two groups to control one direction of the motor. The forward and reverse rotation of the motor is achieved by controlling the conduction of different combinations of MOSFETs. At the same time, the H-bridge PWM drive circuit also allows the motor speed to be controlled by adjusting the duty cycle of the PWM (Pulse Width Modulation) signal. The MicroAutoBox II controller acts as the controller of the electronic throttle, collecting the throttle opening and calculating the control signal, and communicating with the computer algorithm software platform (PC or edge computing device via CAN bus). The engine controller host computer is used to monitor the status of each input and output pin of the MicroAutoBox II controller in real time and observe the CAN communication status and the control effect of the current algorithm iteration. The Vector VN1630A CAN communication device is used to transmit the sampling points of the grid adaptive Bayesian optimization algorithm on the edge computing device or PC to the MicroAutoBox II controller, and send the calibration index calculation related signals from the MicroAutoBox II controller to the PC. Technical solutions such as Figure 1 As shown.

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the following will be combined with... Figure 1 The embodiments of the present invention will be described in further detail below.

[0070] The main process of this example involves setting the initial throttle controller control parameters in MATLAB / SIMULINK software on a PC or edge computing device, while the host computer monitors the control effect. At this point, parameters related to calibration index calculation are transmitted to the PC or edge computing device via VectorVN1630A. The PC or edge computing device runs a grid-adaptive Bayesian optimization algorithm to select the next set of control parameter sampling points, updates the control parameters via CAN communication, and simultaneously collects the parameters required for the next iteration of optimization (message IDs are set to 90 and 91 respectively). After several rounds of optimization, the optimal PID parameters for throttle control are found. These optimal PID parameters are then sent to the electronic throttle controller and stored in the controller for subsequent throttle control.

[0071] Example 2

[0072] The present invention also provides an online calibration method for control parameters of an engine control system, applied to the online calibration device for control parameters of the engine control system described in Embodiment 1, wherein the online calibration method includes:

[0073] Step S1: Set the feasible region of the control parameters and the initial sampling point values ​​of the control parameters; the control parameters are the proportional parameters, derivative parameters and integral parameters of the PID algorithm.

[0074] Step S2: Based on the mean function, covariance function, and the initial sampling point values ​​of the control parameters, apply the prior Gaussian process model and Bayesian sampling strategy to determine the next sampling point value of the control parameters. The Bayesian sampling strategy is obtained by optimizing the acquisition function using a grid adaptive search algorithm.

[0075] Step S3: Based on the next sampling point value of the control parameters, apply the posterior Gaussian process model to determine the value of the objective function; the objective function is the absolute integral of the difference between the throttle opening of the electronic throttle body under the control parameters and the target opening.

[0076] Step S4: Determine whether the value of the objective function satisfies the preset convergence condition.

[0077] Step S5: When the value of the objective function does not meet the preset convergence condition, the posterior Gaussian process model is updated by applying the method based on maximizing the logarithmic marginal probability, and the next sampling point value of the control parameter is input into the updated posterior Gaussian process model to obtain the updated value of the objective function, and then the process returns to step S4.

[0078] Step S6: When the value of the objective function satisfies the preset convergence condition, the value of the next sampling point of the control parameter is taken as the online calibration result.

[0079] In this embodiment, the method first collects the relevant parameters required for the calculation of the control index, solves the calibration index, i.e., the size of the objective function, and then designs a sampling function based on the Bayesian sampling strategy. The optimal value of the sampling function is searched using a grid adaptive search algorithm to determine the next set of optimal control parameter sampling points. The process is iterated repeatedly until the improvement of the calibration index is no longer significantly reflected in the actual control effect or the convergence conditions of other algorithms are met.

[0080] The basic computational process of the grid adaptive Bayesian optimization algorithm is as follows: Figure 2 As shown, for the online calibration problem of electronic throttle control parameters, after designing the discrete PID controller, the optimal sampling points are designed for the controller's control parameters to be optimized (proportional gain, integral gain, and derivative gain), and the process is iterated until the optimal parameter combination is found. The specific implementation steps are as follows:

[0081] Step 1: Determine the feasible region [LB, UB] of the parameters to be calibrated, give the initial calibration point x0, and initialize the algorithm.

[0082] Step 2: Train the prior Gaussian process model based on the mean function μ(x) and covariance function κ(x, x′), combined with the initial point and the information of random initial sampling points generated by random sequence.

[0083] Step 3: Based on the probability information of the Gaussian process model, optimize the sampling strategy to obtain the next suitable sampling point x. n+1 .

[0084] Step 4: For the new sampling point x n+1 Sampling is performed to obtain the objective function value y n+1 .

[0085] Step 5: Determine if the algorithm convergence condition is met. If it is, output the parameter values ​​and objective function values ​​corresponding to the current point as the online calibration results; if not, update the posterior Gaussian process model using the updated dataset D based on the method of maximizing the logarithmic marginal probability and execute Step 3.

[0086] In practical applications, taking the PID parameters of throttle control as an example, such as... Figure 3 and Figure 4 As shown, the Bayesian optimization process is as follows:

[0087] Step 1: Using the three initial PID parameters Kp, Ki, and Kd as control parameters for throttle control, a 4-second sinusoidal signal is used for calibration testing. The sinusoidal signal is in the form of 1.474sin(2πt) + 0.096. The measured parameter is the throttle position. The objective function value is then calculated, set as the integral of absolute value error (IAE). The controller tracks the sinusoidal signal to control the throttle position. After the operation, a Gaussian process model is trained using the three PID parameters as input and the IAE as output via the maximum likelihood method. This involves finding the parameter θ (the mean and covariance functions are parameterized by θ) that maximizes the probability of the actual function value. In practice, maximizing the log-likelihood function is commonly used, and maximum likelihood estimation can be achieved through gradient ascent. The trained Gaussian process model is then used for subsequent IAE prediction.

[0088] Step 2: Using a Bayesian sampling strategy, select the next PID parameter set and use it for the experiment.

[0089] Step 3: Evaluate the integral of actual absolute value error (IAE) of the throttle position tracking effect of this set of parameters, and add this set of experimental data to the training dataset of the Gaussian process to update the Gaussian process model.

[0090] The above three steps are repeated until the convergence condition is met, at which point the optimization is considered complete, and the optimal parameter set obtained by optimization is the final calibration result.

[0091] The implementation process of this invention applies Bayesian optimization surrogate model selection, Gaussian process parameter update, Gaussian process model prediction principle, acquisition function design and optimization, algorithm convergence condition design, controller and control strategy, and algorithm hyperparameter design for the electronic throttle control parameter calibration problem.

[0092] 1. Proxy model selection for Bayesian optimization.

[0093] The core of Bayesian optimization algorithms lies in using prior knowledge and Bayesian probability and decision theory to determine the next sampling action. Therefore, the surrogate model in Bayesian optimization algorithms needs to reflect certain probabilistic characteristics. The role of the surrogate model is to establish a preliminary model of the objective function from known prior knowledge, and to continuously refine it through each iterative sampling to obtain a posterior model that better reflects the characteristics of the objective function.

[0094] Surrogate models used in Bayesian optimization are typically divided into parametric surrogate models and non-parametric surrogate models. Parametric surrogate models require explicit parameters directly related to the objective function; this type of surrogate model is called a parametric surrogate model. Updating the posterior of the surrogate model is actually equivalent to updating the posterior probabilities of its explicit parameters. This type of surrogate model is usually used for problems with small search spaces and relatively simple objective functions. For the engine control system control parameter calibration problem studied in this application, using a non-parametric model as the surrogate model is more suitable. This type of model has no explicit parameters or its parameters are not directly related to the objective function. Among non-parametric surrogate models, random forest models and Gaussian process models are popular. However, random forest models are generally only applicable to discretized parameter spaces, require a large amount of data for each iteration, and cannot optimize their parameters based on gradients. Therefore, the predictive surrogate model used in this application adopts a Gaussian process (GP) model. GP has a unique and important characteristic: it does not produce point estimation predictions like other machine learning models, but instead uses a probability distribution for prediction. Prediction in the form of a probability distribution or probability prediction is key to achieving Bayesian optimization, as it can quantify the uncertainty in prediction, thereby improving the risk-reward trade-off in decision-making.

[0095] A Gaussian process model is a stochastic process model, defined as follows: a Gaussian process model is ... n Its joint probability density function is p(f(x1), ..., f(x)). n When all three variables conform to a joint Gaussian distribution, this random process is called a Gaussian process, as shown in the equation. Its characteristics are entirely determined by its mean and covariance function.

[0096]

[0097] Where M is a multivariate Gaussian distribution, θ is the hyperparameter of the GP model structure, Θ represents the range of parameter variation, p(f|θ) is the predicted probability distribution, p(y|x, θ) is the prior distribution, and μ x Cov is the mean function. f Let be the covariance function.

[0098]

[0099] The modeling process using the Gaussian process model involves three key points: using the mean function to control the expected behavior of the GP, using the covariance function to control the smoothness of the GP, and using gradient descent to learn the optimal hyperparameters of the GP. These will be explained in detail below.

[0100] 1) Design of mean function and covariance function.

[0101] The mean function and covariance function are two core components of a Gaussian process (GP). By making specific choices of the mean or covariance function, prior knowledge can be effectively assigned to the GP; therefore, this application requires the selection and design of the mean and covariance functions.

[0102] 2) Design of the mean function.

[0103] When the characteristics of the objective function to be optimized are relatively clear and its variation with the variable to be optimized is relatively simple, the form of the mean function can be directly specified. The main influence of the prior mean on prediction and the optimization strategy made based on these predictions lies in the Bayesian inference principle. However, making inferences without sufficiently clear prior knowledge can be dangerous. Therefore, in Bayesian optimization, the prior mean function is usually regarded as a constant function, as shown in equation (3) below:

[0104] μ(x;c)≡c (3)

[0105] The value of the constant c is usually chosen to be 0 when the distribution of the objective function is unknown. Then, the value is updated by maximizing the logarithmic marginal likelihood using the initial point data and subsequent sampled data.

[0106] 3) Design of covariance function.

[0107] While the mean function defines the expected behavior of the overall objective function, the covariance function or kernel function of the generalized system (GP) plays a more complex role. It expresses the relationship between data points within the domain and controls the structure and smoothness of the GP. When its length scale is set relatively large, the model's predictions become smoother and less uncertain. If it is set relatively small, the GP model's predicted values ​​will change faster and the uncertainty will increase. The size of the shape parameter also affects the uncertainty of the predicted values.

[0108] The selection of its parameters is usually quite complex, so gradient descent is required to configure and update the parameters by maximizing the log marginal likelihood. In addition, different types of covariance functions also play a crucial role in the flexibility and practicality of GP modeling. Among them, the RBF kernel is mainly used to model infinitely differentiable functions. However, since the objective function of the calibration problem usually cannot meet such strict requirements, the Matérn kernel is more suitable. The Matérn 5 / 2 kernel is usually the first choice kernel function for Bayesian optimization because it can ensure that GP has a certain smoothness, thus making the Bayesian optimization algorithm have sufficient stability, and also make GP have sufficient variability, making the modeling of real-world optimization calibration problems more realistic. Its specific form is shown in Equation (4).

[0109] For any set [x1, ..., xn] points of any n, ... n The covariance matrix Σ is defined as follows:

[0110]

[0111] In the above formula, cov f For the covariance of different calibration parameters, x i x j For the i-th and j-th components of the calibration parameters, l1, ..., l D It is the kernel length along each dimension, α is the shape parameter; D represents the data dimension; This represents the variance of the calibration parameters.

[0112] 2. Update Gaussian process parameters.

[0113] After determining the two main components of the Gaussian process, the mean function and the covariance function, the prior model of the Gaussian process needs to be continuously updated after collecting the sampling points determined by the Bayesian decision. The hyperparameters of the mean function and the hyperparameters of the covariance function together constitute the hyperparameters θ of the Gaussian process model. In the process of updating the model, the gradient descent method is mainly used to maximize the log marginal likelihood, thereby ensuring the fit between GP and the sampled data pair, as shown in Equation (5).

[0114]

[0115] Where μ and Σ are the mean vector and covariance matrix, respectively, N is the covariance matrix of the observation noise, n is the amount of data, and x and y are the input and output vectors of the data, respectively.

[0116] 3. Prediction principle of Gaussian process model.

[0117] In the Bayesian optimization process, after the Gaussian process surrogate model is updated, model prediction is required to evaluate each candidate sampling point according to the acquisition function. When performing model prediction for the Gaussian process, since each set of calibration parameters corresponds to the normal distribution of its objective function, when considering the entire function space, the objective function random variable {f(x)} conforms to the following formula.

[0118]

[0119] Where μ(x) is the mean function of the objective function, κ(x, x′) is the covariance kernel function of the objective function, f(x) represents the objective function value, x and x′ are the data points of two sets of calibration parameters, E represents the mean operation, and Rp represents the p-dimensional real space.

[0120] Using the parameter set X = [x1, ..., x2] from the training data N ] T The set of measurement evaluation values ​​corresponding to X is Y = [y1, ..., y2]. N ] T Perform x * Gaussian process prediction of parameter set is obtained by solving the conditional probability through the joint high-dimensional Gaussian distribution of training data and prediction points, as shown in Equation (9).

[0121]

[0122] Where, x * f(x) * ) represent the calibration parameter data value and the corresponding objective function value for the next sampling point, respectively, μ * and κ * These are the updated mean function matrix and covariance kernel function matrix, respectively. X and Y are the calibration parameters of the previous sampling point and the obtained objective function data, respectively. μ(x * ) represents the mean of the calibration parameters for the next sampling point; μ(X) represents the mean of the calibration parameters for the previous sampling point; κ(X,X) represents the variance of the historical sampling points; κ(x) represents the variance of the historical sampling points. * κ(X,x) represents the covariance between the next sampling point and the historical sampling points; * ) represents the covariance between historical sampling points and the next sampling point; κ(x) *,x * ) represents the variance of the next sampling point.

[0123] 4. Design and optimization of data acquisition functions.

[0124] The surrogate model and the sampling function are two main components of the Bayesian optimization process. This application designs a corresponding sampling function and selects an optimization algorithm for the calibration problem of an engine control system. Since the results of observations cannot be known before sampling, the sampling optimization strategy must consider this uncertainty. Furthermore, because the results of each observation not only have a direct impact on the present but also form the basis for all future decisions, the sequential nature of the optimization further exacerbates the importance of this uncertainty. Therefore, developing an effective sampling optimization strategy requires addressing this uncertainty in some way.

[0125] The sampling function quantifies the value of each sampling point given a predicted probability distribution. The location of the next sampling point is determined by finding the point that maximizes the value of the sampling. According to the Bayesian optimization principle, the selection of sampling points is an iterative process. In each iteration of the loop, the general algorithm (GP) is trained based on the data observed from the target. The GP is evaluated according to the sampling strategy to obtain sampling suggestions. Observations are made at the recommended sampling points, and new points are added to the training data. The whole process is repeated until the algorithm's cutoff condition is met.

[0126] 1) Design of the data acquisition function.

[0127] In Bayesian optimization, commonly used acquisition functions include probability-based acquisition functions (PI), expectation-based acquisition functions (EI), upper confidence bound acquisition functions (UCB), and lower confidence bound acquisition functions (LCB).

[0128] The probability-based improved acquisition function (PI) uses the probability that an observed point might find a better objective function value than the current point as the evaluation criterion, and can be expressed as:

[0129]

[0130] Among them, f * Let μ(x) be the current optimal value corresponding to the current point, μ(x) be the predicted mean corresponding to the current posterior Gaussian process, σ(x) be the predicted variance corresponding to the current posterior Gaussian process, Φ be the cumulative probability density of the standard normal distribution, D be the currently collected dataset, x be the data point, and α be the current value. PI (x) is the probability-based improved acquisition function.

[0131] Using the PI acquisition function only evaluates the probability of improvement of the objective function without caring about the amount of improvement. This may lead to an overly conservative search strategy, ignoring regions where the improvement may be significant but the probability of improvement is low. Therefore, using this acquisition function for Bayesian optimization often causes the optimization to get stuck in local optima.

[0132] The expected improvement acquisition function (EI) considers not only the probability of improvement at the observation point but also the magnitude of the improvement. Its expression is as follows:

[0133]

[0134] Among them, a EI (x; D) represents the acquisition function based on the expectation improvement, μ(x) is the predicted mean corresponding to the current posterior Gaussian process, σ(x) is the predicted variance corresponding to the current posterior Gaussian process, Φ is the cumulative probability density of the standard normal distribution, D is the currently acquired dataset, and x is the data point.

[0135] This sampling strategy balances exploration and exploitation, but it is not suitable for situations with noise.

[0136] There is no fundamental difference between the upper confidence bound-based acquisition function (UCB) and the lower confidence bound-based acquisition function (LCB). Both use the upper and lower confidence bounds of the posterior Gaussian process as evaluation criteria to determine the location of the next sampling point. This strategy can still make relatively good sampling decisions even under high noise conditions. For the control system calibration problem studied in this application, the objective function value is usually minimized. Therefore, choosing LCB is more reasonable. Its expression is as follows:

[0137]

[0138] Where β l =2ln(D) t 2π 2 / (6δ)), D is the input data space, t is time, v and δ are empirical parameters. According to the research of Srinivas et al., the recommended values ​​for v and δ are 0.2 and 0.1, respectively, to better balance exploration and utilization. If two data points have the same predicted mean but different predicted standard deviations, the data point with higher uncertainty will receive a higher evaluation value; if two data points have the same predicted standard deviation but different predicted means, the data point with the higher mean will receive a higher evaluation value. Where vβ t The smaller the value of vβ, the more the LCB strategy tends to exploit it; conversely, vβ... t The larger the value, the more the LCB strategy leans towards exploration.

[0139] 2) Optimize the acquisition function.

[0140] In Bayesian optimization, the acquisition function guides the selection of the next sampling point, while the optimization method of the acquisition function is the direct step in determining the sampling point. The acquisition function found through the optimization method determines the next sampling point based on the maximum evaluation value calculated by the current posterior Gaussian process. By continuously updating the posterior Gaussian process and performing iterative optimization, the sampling point corresponding to the point where the evaluation value cannot be sufficiently improved is the optimal solution of Bayesian optimization, and its value is the optimal value of Bayesian optimization. Therefore, the optimization method of the acquisition function is very important in Bayesian optimization, as it directly determines whether a global optimum can be found.

[0141] However, acquisition functions are often non-convex and multimodal, sometimes even more difficult to optimize than the objective function itself. Fortunately, since acquisition functions are calculated based on probabilistic information predicted by Gaussian processes, their evaluation cost is much lower than that of the objective function. Commonly used acquisition function optimization algorithms include gradient-based quasi-Newton methods (L-BFGS-B), pattern search algorithms, and metaheuristic algorithms. For gradient-based optimization algorithms, it is often impossible to find the global optimum when dealing with acquisition function optimization problems with non-convex and multimodal characteristics. Metaheuristic algorithms have slow convergence speeds and can only design convergence conditions based on the number of iterations and the objective function value. This approach is very inconvenient for online calibration problems of engine control systems, because online calibration of engine control systems not only requires high calibration cycles, but also needs to consider calibration accuracy, parameter accuracy, and the number of samplings in the convergence condition design. In addition, in order to avoid a series of safety issues caused by the calibration parameters exceeding the feasible region during the actual calibration process, it is necessary to constrain the parameter range while optimizing. Therefore, this application selected the Grid Adaptive Search (MADS) algorithm in the pattern search algorithm. Compared with the direct search algorithm, it can perform dense search in multiple directions in the search space through the vector set, instead of only searching along the grid direction. This characteristic determines its fast optimization speed and strong ability to find the global optimum.

[0142] The MADS algorithm first initializes the initial point and its initial grid parameters. Then, it iteratively searches for the next sampling point that meets the improvement criteria. The iterative search consists of two parts: a global search in the Search phase, and a local refinement search in the Poll phase after the Search phase fails to find an optimal candidate point (i.e., a sampling point). During the search, if the next optimal candidate point is found, the sampling point is updated and the grid parameters are increased to accelerate convergence; otherwise, the grid parameters are decreased for a refinement search. The algorithm terminates when the grid parameters are smaller than the set grid size tolerance. The pseudocode for its optimization process is as follows: Figure 5 As shown.

[0143] 5. Algorithm convergence condition design.

[0144] For the control parameters and control MAP calibration problem of the engine control system studied in this application, considering the accuracy requirements of calibration parameters, the accuracy requirements of calibration indexes, and the evaluation cost of experiments during the calibration process, the following convergence criterion is designed based on the principle of Bayesian optimization algorithm:

[0145]

[0146] Where, Δ mesh M represents the actual mesh size tolerance. size The grid size tolerance value determines the accuracy of the calibration parameters; y n and y n+1 F represents the objective function for the nth and (n+1)th evaluations. c The minimum improvement in the objective function depends on the accuracy requirements of the calibration index; T no I represents the actual number of iterations in which the objective function value does not change significantly. s The maximum number of iterations at which the objective function value does not change significantly allows for fast convergence of the algorithm; F eval and T represent the number of evaluations of the actual objective function and the number of actual algorithm iterations, respectively, and F max with I max These are the maximum number of evaluations of the objective function and the maximum number of iterations of the algorithm, respectively, which depend on the experimental cost that can be tolerated for the calibrated problem; C v To determine the required threshold for calibration, the algorithm needs to be designed based on the specific calibration problem. The algorithm terminates when one of the above convergence conditions is met. The design of the convergence parameters needs to be based on experiments examining the influence of the algorithm's hyperparameters on the specific calibration problem.

[0147] 6. Controller and control strategy.

[0148] 1) Design of discrete PID controller for electronic throttle valve.

[0149] Since the actual controller's control process is sampling control, the deviation at the sampling time needs to be calculated before the control quantity can be calculated. This section designs a discrete PID controller. However, because the positional PID control algorithm requires the accumulation of error terms for each control signal output calculation, the calculated control signal may change significantly when there are problems with the measurement signal acquisition, which could damage the DC motor driving the electronic throttle system. Therefore, the following incremental PID controller is designed:

[0150]

[0151] Where, k p k iand k d The three control parameters for a PID controller are k. i =k p / T I k d =k p T D T is the sampling period, k is the kth sampling point, e(k), e(k-1), e(k-2) are the error terms of the throttle position at the kth, k-1th, and k-2th times, respectively, Δu is the control voltage increment, and u is the control voltage value.

[0152] Meanwhile, an integrator clamping method is used to prevent integral saturation. Integral separation is performed when the control signal reaches the actuator saturation voltage limit and the error is in the same direction as the control signal. This algorithm can prevent the control quantity from staying in the saturation region for a long time, so as to ensure the accuracy and speed of electronic throttle valve opening control.

[0153] 2) Control strategy.

[0154] Based on the selection of the drive circuit and the frequency design of the PWM speed control signal, and according to the relevant requirements of the electronic throttle control system, the following is performed: Figure 6 The main control program for the electronic throttle valve test platform is shown.

[0155] In the main control program of the electronic throttle test platform, the sensor voltage signal is measured by the ADC module and the actual angle value is calculated by the sensor characteristic curve. After the duty cycle of the output PWM signal of the discrete PID control is determined, it first passes through the duty cycle threshold judgment part before finally determining the duty cycle of the control signal to be output. This is because the sensor and ADC module will generate measurement errors during the measurement process and process errors will also be generated during actual operation. These errors will cause the DC drive motor to frequently switch between forward and reverse rotation if the threshold judgment module is not set, resulting in rapid vibration of the throttle body.

[0156] 7. Design hyperparameters for the algorithm to address the calibration problem of electronic throttle control parameters.

[0157] For the convergence criteria designed in this application, most of the parameters need to be determined by specific analysis of the calibration problem to be applied. This application designs the convergence parameters as shown in Tables 1, 2 and 3 for calibration accuracy, parameter accuracy and evaluation cost of the control parameters of the electronic throttle control system.

[0158] Table 1. General Hyperparameter Statistics of the Algorithm

[0159]

[0160]

[0161] Table 2. Statistics of Hyperparameters for Algorithm Convergence

[0162]

[0163] Table 3. Convergence parameter statistics for the control parameter calibration problem of electronic throttle control system.

[0164] Convergence hyperparameters symbol Parameter value Mesh size tolerance <![CDATA[M size ]]> 0.1 Minimum improvement of objective function <![CDATA[F c ]]> 0.00698 Maximum number of evaluations of the objective function <![CDATA[F max ]]> 200 Maximum number of iterations <![CDATA[I max ]]> 100 Calibration requirements <![CDATA[C v ]]> 0.0698

[0165] Since the influence of the calibrated control parameters on the control effect is negligible when the variation range is less than 0.1, the grid size tolerance M is not considered. size For example, 0.1 is sufficient; while the calibration index requires C. v Minimum improvement of the objective function F c The design can be carried out by offline calibration of convergence parameters based on the control parameters in Chapter 2. The IAE calculated when the tracking error is less than 1 degree for each sampling under the tracking condition within 4 seconds is selected as C. v The value of F, and F c Then, the IAE (IAE = ∫0τ|e(t)|) is calculated when the tracking error is 0.1 degrees for each sampling, and the sampling time T is taken as the IAE. s =0.001s; The maximum number of evaluations and the maximum number of iterations are calculated based on the specific calibration cost, taking F as the value. max =200, I max =100. Finally, based on as... Figure 7 The online calibration process shown performs online calibration of the control parameters.

[0166] In practical applications, online calibration of control parameters is performed continuously during system operation. It utilizes real-time data to calibrate control parameters, continuously monitoring and adapting to changes in the environment and system to optimize the calibration results, thus adapting to variations in environmental and system characteristics. While non-model-based online calibration methods address the issue of excessive reliance on model accuracy, they still require extensive experimental sampling, making them less suitable for situations with high sampling costs. Therefore, to reduce the number of sampling attempts, this invention, based on Bayesian strategies and Gaussian process probabilistic modeling principles, guides the sampling process during calibration, ensuring the calibration requirements of control parameters are met while minimizing sampling. A grid-adaptive Bayesian optimization-based online calibration algorithm is proposed, and a corresponding online control parameter calibration device is developed. Under fixed throttle following conditions, the control parameters are continuously updated using the Bayesian optimization algorithm, and the calibration objective function is solved, achieving automated online calibration of the control parameters. This invention employs a Bayesian optimization algorithm in the online calibration of control parameters, utilizing a grid adaptive algorithm to optimize the sampling function to address the non-convex optimization problem during calibration. A Gaussian process model is used as a surrogate model, significantly shortening the online calibration cycle and reducing the number of sampling points required for calibration. Furthermore, this method introduces extended Kalman filtering technology to further improve the accuracy of the surrogate model and the algorithm's anti-interference capability. Therefore, this invention not only improves the automation level of online control parameter calibration but also, by integrating Bayesian optimization and extended Kalman filtering algorithms, significantly enhances the anti-interference capability of online calibration, shortens the calibration cycle, and reduces unnecessary sampling points while meeting calibration requirements, providing a more efficient and reliable technical solution for the calibration of engine control system control parameters.

[0167] This invention has the following advantages:

[0168] 1. Compared with traditional calibration methods, it can be achieved without human intervention, reducing the uncertainty caused by the operation of calibration personnel, improving the reliability of calibration results, and enhancing the automation of the calibration process.

[0169] 2. The designed grid-adaptive Bayesian optimization algorithm can solve complex non-convex optimization problems quickly and guarantee global optimum. It performs well in terms of convergence speed and accuracy of search results, which greatly improves the real-time performance of online calibration.

[0170] 3. Compared with non-model-based online calibration algorithms, it greatly reduces unnecessary sampling points and shortens the online calibration cycle.

[0171] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0172] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for online calibration of control parameters of an engine control system, characterized in that, An online calibration device for control parameters applied to an engine control system, the online calibration device comprising: an electronic throttle body, a DC motor drive module, a controller, a communication module, and a host computer; The host computer is connected to the controller via the communication module; the controller is connected to both the DC motor drive module and the electronic throttle body; the DC motor drive module is connected to the electronic throttle body. The host computer is used to monitor the status of each input and output pin of the controller, receive calibration index measurement parameters, and calculate calibration sampling point data based on the calibration index measurement parameters using the online calibration method for the engine control system control parameters, and send the calibration sampling point data to the controller; the calibration index measurement parameters are calculated by the controller based on the throttle opening signal of the electronic throttle body; the online calibration method for the engine control system control parameters is a grid adaptive Bayesian online calibration algorithm; The controller is used to send motor operating parameters to the DC motor drive module based on the calibration sampling point data using a PID algorithm, and to receive the throttle opening signal returned by the electronic throttle body; the motor operating parameters include motor speed and motor rotation direction; The DC motor drive module is used to control the operation of the electronic throttle body according to the motor operating parameters; The online calibration method includes: Define the feasible region of the control parameters and the initial sampling point values ​​of the control parameters; the control parameters are the proportional parameters, derivative parameters, and integral parameters of the PID algorithm; Based on the mean function, covariance function, and the initial sampling point values ​​of the control parameters, the next sampling point value of the control parameters is determined by applying a prior Gaussian process model and a Bayesian sampling strategy. Based on the next sampling point value of the control parameters, a posterior Gaussian process model is applied to determine the value of the objective function; the objective function is the integral of the absolute value of the difference between the throttle opening of the electronic throttle body under the control parameters and the target opening. Determine whether the value of the objective function satisfies the preset convergence condition; When the value of the objective function does not meet the preset convergence condition, the posterior Gaussian process model is updated by applying the method based on maximizing the logarithmic marginal probability, and the next sampling point value of the control parameter is input into the updated posterior Gaussian process model to obtain the updated value of the objective function, and the process returns to the step "determine whether the value of the objective function meets the preset convergence condition". When the value of the objective function satisfies the preset convergence condition, the value of the next sampling point of the control parameter is taken as the online calibration result; The covariance function is: Among them, cov f For the covariance of different calibration parameters, x i x j For the i-th and j-th components of the calibration parameters, l1,...,l D It is the kernel length along each dimension, α is the shape parameter; D represents the data dimension; The variance of the calibration parameter, r 2 (x i ,x j ) represents x i and x j The sum of squared distances; x id For the data of the d-th dimension of the i-th component of the calibration parameter, x jd The data for the j-th component and d-th dimension of the calibration parameters.

2. The online calibration method for control parameters of an engine control system according to claim 1, characterized in that, The host computer includes an edge computing device and a controller host computer; Both the edge computing device and the host computer of the controller are connected to the controller; The edge computing device is used to receive calibration index measurement parameters, and calculate calibration sampling point data by applying a grid adaptive Bayesian online calibration algorithm based on the calibration index measurement parameters, and send the calibration sampling point data to the controller; The host computer of the controller is used to monitor the status of each input and output pin of the controller.

3. The online calibration method for control parameters of an engine control system according to claim 1, characterized in that, The mean function is μ(x; c); Where x is the calibration parameter, i.e., the three parameters of the PID; c is a constant.

4. The online calibration method for control parameters of an engine control system according to claim 1, characterized in that, The objective function is: Where μ(x) is the mean function of the objective function, κ(x, x′) is the covariance kernel function of the objective function, f(x) represents the objective function value, x and x′ are the data points of two sets of calibration parameters, E represents the mean operation, and R p It represents a p-dimensional real space.

5. The online calibration method for control parameters of an engine control system according to claim 1, characterized in that, The updated posterior Gaussian process model is as follows: m * =κ(x * ,X)(κ(X,X)) -1 (Y-μ(X))+μ(x * ); k * =κ(x * ,x * )-k(x * ,X)(κ(X,X)) -1 k(X,x * ); Where, x * f(x) * ) represent the calibration parameter data value and the corresponding objective function value for the next sampling point, respectively, μ * and κ * These are the updated mean function matrix and covariance kernel function matrix, respectively. X and Y are the historical sampling point calibration parameters and the obtained objective function data, respectively. μ(x * ) represents the mean of the calibration parameters for the next sampling point; μ(X) represents the mean of the calibration parameters for the historical sampling points; κ(X,X) represents the variance of the historical sampling points; κ(x) represents the variance of the historical sampling points. * κ(X,x) represents the covariance between the next sampling point and the historical sampling points; * ) represents the covariance between historical sampling points and the next sampling point; κ(x) * ,x * ) represents the variance of the next sampling point.

6. The online calibration method for control parameters of an engine control system according to claim 1, characterized in that, The Bayesian sampling strategy is obtained by optimizing the sampling function using a grid adaptive search algorithm.

7. The online calibration method for control parameters of an engine control system according to claim 6, characterized in that, The acquisition function is the acquisition function with the lower confidence bound.

8. The online calibration method for control parameters of an engine control system according to claim 1, characterized in that, The preset convergence condition is: Any one of them; Where, Δ mesh M represents the actual mesh size tolerance. size The grid size tolerance value determines the accuracy of the calibration parameters; y n and y n+1 F represents the objective function for the nth and (n+1)th evaluations. c The minimum improvement in the objective function depends on the accuracy requirements of the calibration index; T no I represents the actual number of iterations in which the objective function value does not change significantly. s F represents the maximum number of iterations at which the objective function value does not change significantly. eval and T represent the number of evaluations of the actual objective function and the number of actual algorithm iterations, respectively, and F max with I max These are the maximum number of evaluations of the objective function and the maximum number of iterations of the algorithm, respectively; C v The required threshold is to be set.

Images