A method for torque prediction control of a compression ignition aero-piston engine

Abstract

The application discloses a compression ignition aviation piston engine torque prediction control method, step 1, obtaining and collecting historical operation data; step 2, data min-max normalization processing; step 3, training torque prediction model; step 4, MPC torque prediction control, including step 41, setting reference trajectory; step 42, torque multi-step prediction; step 43, feedback correction; step 44, rolling optimization; step 45, determining future fuel injection amount. The application firstly simulates and controls the engine by using a traditional MAP control mode to obtain engine historical operation data; secondly takes fuel injection amount, intake amount, output torque and its differential as prediction model input to obtain predicted torque; finally, the SQP is used for rolling optimization of performance indexes to obtain an optimal fuel injection amount sequence, and the first component is applied to an aviation piston engine simulation model to control the output torque of the engine.

Description

Technical Field

[0001] This invention belongs to the technical field of torque prediction and control methods for compression-ignition aero-piston engines, specifically relating to a torque prediction and control method for compression-ignition aero-piston engines. Background Technology

[0002] On October 16, 2020, the General Aviation Branch of the China Air Transport Association released the "2019-2020 China General Aviation Development Report." The report indicated that the number of general aviation enterprises, drone enterprises, registered general aviation aircraft, and registered general aviation airports nationwide all saw significant growth, signifying a promising future for my country's general aviation industry. Among various types of aero engines, piston engines dominate the application of general aviation, primarily in fixed-wing propeller aircraft and unmanned helicopters. With the widespread use of aviation kerosene in piston engines, compression-ignition piston engines offer superior advantages compared to spark-ignition piston engines, thus gradually becoming one of the research hotspots for piston aero engines.

[0003] The control technology of aero-engines has continuously evolved from the original hydraulic-mechanical control to electronic control. Thanks to the rapid advancement of computer technology, the control technology of aero-engines has now developed to Full Authority Digital Electronic Control (FADEC). Compared with traditional mechanical-hydraulic controllers and analog electronic controllers, digital electronic controllers are not only simple in structure, but also can apply various complex and advanced control methods to the engine control system.

[0004] Currently, a large amount of research has been conducted on engine torque control both domestically and internationally. Commonly used control methods include lookup table control, PID control, optimal control, and model predictive control.

[0005] The first type of control method: lookup table control. This is a relatively traditional engine control method. First, the engine is calibrated through extensive bench testing to generate a lookup table. Then, the lookup table is imported into the engine's FADEC (Feature Dependent Engine). When the engine is running, sensors detect the current crankshaft speed and transmit the data to the FADEC, while simultaneously calculating the required torque. Finally, based on different crankshaft speeds and required torques, the corresponding fuel injection quantity is obtained by looking up the table, thus achieving the purpose of controlling the engine's output torque.

[0006] Although the above method is relatively simple, it requires a large number of calibration tests. Furthermore, the operating conditions of piston-engine aircraft change frequently, external environmental interference is strong, and there are many uncertainties, which can easily lead to a decrease in control accuracy. Therefore, it is necessary to calibrate the MAP data according to different operating conditions.

[0007] The second type of control method is PID control. Traditional PID control is widely used due to its simple structure, ease of implementation, and lack of the need for a precise system model. First, the PID parameters are tuned to obtain suitable control parameters. Then, these parameters are written into the electronic control unit (ECU). Based on the real-time feedback of the difference between the engine's actual output torque and the target torque, the ECU adjusts the amount of fuel injected to control the engine's output torque.

[0008] Based on the operating characteristics of aero engines, their PID parameters are not constant under different operating conditions. Multiple sets of PID parameters need to be set to adapt to different operating states. Furthermore, PID control is a single-input single-output control strategy, while aero engines have a complex structure, and controlling only a single variable cannot meet the requirements.

[0009] The third type of control method: optimal control. Optimal control belongs to modern control theory. First, the system state equation is established; then, performance indicators and constraints are established according to the control objective. These are usually composite performance indicators, consisting of a terminal part and an integral part. The terminal part represents the final state the system aims to achieve, and the first term of the integral part represents the error between the feedback value and the target torque requirement, while the second term represents a measure of control energy consumption. Finally, based on satisfying the extreme value conditions, the minimum value is solved to calculate the optimal total fuel injection quantity.

[0010] However, the above methods overemphasize optimality and have the following problems: 1. They are difficult to solve for nonlinear cases with complex constraints; 2. They require an accurate system model.

[0011] The fourth type of control method: Model Predictive Control. It mainly consists of four parts: reference trajectory, rolling optimization, predictive model, and feedback correction. First, the prediction time domain is determined. and control time domain Then, based on the actual output torque and the target required torque, the future torque is obtained. The target torque reference trajectory at each time point is used, and the difference between the reference trajectory and the predicted torque after feedback correction is used as the first performance indicator for rolling optimization. The optimal fuel injection quantity sequence is calculated; finally, the first component of the fuel injection quantity sequence is applied to the controlled object. Simultaneously, the calculated optimal fuel injection quantity is applied to the prediction model for multi-step prediction.

[0012] Model predictive control (MPC) is increasingly being applied to the control of automotive and aero-engines. Compared to optimal control, MPC optimizes only within a finite time domain, rather than the entire time domain. Aero-engines are complex nonlinear systems operating in highly variable environments, requiring controller design to account for numerous physical constraints and uncertainties. MPC possesses the ability to explicitly handle system constraints and maintains good control performance even when facing external environmental disturbances and uncertainties.

[0013] Therefore, this invention uses the required torque as the target torque and adopts a model predictive control strategy based on SA-PSO trained RBF neural network to achieve engine torque control, which can achieve good control effect.

[0014] Therefore, this paper proposes a torque prediction control method for compression-ignition aero-piston engines. Summary of the Invention

[0015] To address the aforementioned technical problems, this invention designs a torque prediction control method for a compression-ignition aero-piston engine. This invention takes the compression-ignition aero-piston engine as the research object, with the required torque as the target torque, and employs a model predictive control strategy based on SA-PSO training of an RBF neural network to achieve engine torque control. First, the main framework of the compression-ignition aero-piston engine and other accessories are built using the IFP-Engine and Signal-Control libraries in the AMESim simulation platform; and co-simulation is performed with MATLAB / Simulink, using the traditional MAP control strategy to obtain engine operating data as the training dataset for the neural network prediction model. Second, the RBF neural network torque prediction model is trained using the Particle Swarm Optimization (PSO) method, and simulated annealing (SA) is introduced to avoid training getting trapped in local minima; finally, sequential quadratic programming (SQP) is used as a rolling optimization strategy to construct a complete model predictive control framework.

[0016] To achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution: a torque prediction control method for a compression-ignition aircraft piston engine, comprising the following steps: Step 1: Acquire and collect historical operating data: Build a physical model of a compression-ignition aero-piston engine in LMS Imagine.Lab AMESim simulation software, and build a controller in MATLAB / Simulink. The control method is MAP lookup table control. Connect the engine physical model and the controller through the co-simulation interface, and perform simulation operation of the engine start-up phase and different operating conditions to obtain the engine's historical operating data. The historical operating data mainly includes fuel injection quantity, intake air quantity, and output torque and its derivative.

[0017] Step 2, data min-max normalization processing: map all collected historical running data into the interval (0, 1).

[0018] Step 3, Train the torque prediction model: Establish a multi-input single-output mathematical model as follows:

[0019] The torque prediction model has four inputs and one output. The inputs to the prediction model are the intake volume at the previous time step. Fuel injection quantity at the previous moment The actual crankshaft output torque at the previous moment and the derivative of the actual crankshaft output torque The output is the predicted torque at the current moment. .

[0020] The RBF neural network is configured with 4 input neurons, corresponding to fuel injection quantity, intake air quantity, and output torque and their derivatives; and 1 output neuron, corresponding to predicted torque. Simulated annealing (SA) is introduced when training the RBF neural network in the PSO. Finally, the optimal parameter width of the Gaussian kernel function is obtained through iteration. and center and optimal connection weights .

[0021] Step 4, MPC torque prediction control, includes the following steps: Step 4.1, Setting the Reference Trajectory: The purpose is to ensure that the system output torque reaches the target torque along a smooth reference trajectory, thus smoothing the control process. The specific calculation formula is as follows:

[0022] in, The reference trajectory representing the target torque requirement. This indicates the actual output torque of the engine. Indicates the target torque requirement. Indicates the prediction time domain, Represents the reference trajectory coefficients, and .

[0023] Step 4.2, Torque Multi-Step Prediction: To achieve multi-step prediction, multiple neural networks need to be connected in series, using the predicted output value of the previous neural network as the input value of the next neural network; After one prediction, the predicted sequence is obtained. .

[0024] Step 4.3, Feedback and Correction: The measured data of engine output torque at any given time are used to perform prediction compensation on the torque prediction sequence obtained in Step 4.2. The specific calculation formula is as follows:

[0025] in, This indicates that the neural network predicts the output torque. This represents the error correction factor. express The deviation between the actual output torque of the engine and the output torque predicted by the neural network at any given time.

[0026] Step 4.4, Rolling Optimization: Using SQP to solve for a constrained quadratic performance index as the loss function, the optimal injection quantity sequence in the finite time domain is obtained. The loss function is chosen as follows:

[0027] in, Indicates the target torque requirement. This represents the predicted torque after feedback correction. Indicates the amount of change in fuel injection. This represents the torque deviation between the target required torque and the predicted torque. Indicates the first Weighting factor for step-by-step torque prediction error Indicates the first Weighting factor for step-by-step fuel injection variation. Indicates the prediction time domain, This indicates the control time domain.

[0028] During engine operation, constraints such as maximum and minimum fuel injection limits, as well as the range of air-fuel ratio or excess air coefficient, also need to be considered. These system constraints are expressed as follows:

[0029]

[0030] in, and This indicates the minimum and maximum fuel injection quantities at different crankshaft speeds. and This indicates the optimal air-fuel ratio range for the engine.

[0031] Step 4.5, Determine the future fuel injection quantity: Take the first component of the optimal fuel injection quantity sequence obtained in Step 4.4 as... The timing control parameters act on the fuel injection module of an aircraft piston engine; Repeat Steps 4.1 to 4.5 continuously until the process ends.

[0032] Furthermore, in Step 2, the formula for min-max normalization is:

[0033] in, This represents the data after min-max normalization. Represents the original data. and This represents the minimum and maximum values ​​of the sample data.

[0034] Furthermore, in Step 3, the activation function for the hidden layer of the RBF neural network is chosen to be a Gaussian kernel function, as shown in the formula:

[0035] Then network output The formula is:

[0036] in, This represents the width of the Gaussian kernel function. This indicates the center of the Gaussian kernel function. This represents the connection weights between the hidden layer and the output layer.

[0037] The position information of each particle is determined by the width of the Gaussian kernel function. and center and connection weights Composition, when the number of neurons in the hidden layer is At that time, the particle position information is represented as:

[0038] in Represents the width of the Gaussian kernel function , Represents the center of the Gaussian kernel function , Represents the connection weights .

[0039] Furthermore, the simulated annealing mechanism introduced when PSO trains the RBF neural network torque prediction model is manifested in the calculation of the jump probability of each particle in each iteration. When the jump probability is greater than the random jump threshold, the current particle is used to replace the globally optimal particle for particle velocity update.

[0040] The formula for calculating the probability of a sudden jump is:

[0041] in, Indicates the first The probability of a particle jumping out. Indicates the first The fitness value of each particle. This represents the fitness value of the globally optimal particle. This indicates the current annealing temperature.

[0042] The fitness function selects the mean absolute error of the network output:

[0043] in, Indicates the first The average absolute error of each particle Indicates the number of training samples. This represents the actual output of the neural network. This represents the expected output of the neural network.

[0044] The formulas for updating particle velocity and position are:

[0045]

[0046] in, Represents particle velocity. Indicates the particle position. Indicates inertia weight, and Represents the learning factor. and Represents a random number. , Represents the optimal particle for an individual. This represents the globally optimal particle.

[0047] The beneficial effects of this invention are: This invention first trains an RBF neural network using a particle swarm optimization method incorporating simulated annealing to construct a torque prediction model for a compression-ignition aero-piston engine. Compared to gradient descent and conventional particle swarm optimization, this method yields a torque prediction model with higher accuracy and better performance. Then, this torque prediction model is used for multi-step prediction. Next, a performance index is constructed using the difference between the target torque demand and the torque after feedback correction, along with the fuel injection increment. Sequential quadratic programming is then used to perform rolling optimization of this performance index within a finite-time domain to solve for the optimal fuel injection sequence, thus achieving torque control of the compression-ignition aero-piston engine. This method leverages the advantages of the SA-PSO method and the RBF neural network to construct an accurate engine torque prediction model; utilizing the advantages of model predictive control, it achieves excellent torque control performance for the compression-ignition aero-piston engine under different operating conditions. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of 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.

[0049] Figure 1 The flowchart of the torque prediction control method for compression ignition aero-piston engines based on SA-PSO optimized RBF neural network of the present invention is shown. Figure 2 The main simulation framework of a compression-ignition aero-piston engine built on the AMESim platform is shown. Figure 3 The flowchart of the training process for the RBF neural network based on SA-PSO is shown. Figure 4 This shows a schematic diagram of multi-step prediction using an RBF neural network; Figure 5 The overall framework for engine model predictive control based on demand torque is shown; Figure 6 A model predictive control framework for engines based on MATLAB / Simulink is shown. Detailed Implementation

[0050] 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.

[0051] Example 1

[0052] See Figures 1 to 6 As shown, a torque prediction control method for a compression-ignition aero-piston engine includes the following steps: Step 1: Acquire and collect historical operating data: Build a physical model of a compression-ignition aero-piston engine in LMS Imagine.Lab AMESim simulation software, and build a controller in MATLAB / Simulink. The control method is MAP lookup table control. Connect the engine physical model and the controller through the co-simulation interface, and perform simulation operation of the engine start-up phase and different operating conditions to obtain the engine's historical operating data. The historical operating data mainly includes fuel injection quantity, intake air quantity, and output torque and its derivative.

[0053] Step 2, data min-max normalization processing: map all collected historical running data into the interval (0, 1).

[0054] Step 3, Train the torque prediction model: Establish a multi-input single-output mathematical model as follows:

[0055] The torque prediction model has four inputs and one output. The inputs to the prediction model are the intake volume at the previous time step. Fuel injection quantity at the previous moment The actual crankshaft output torque at the previous moment and the derivative of the actual crankshaft output torque The output is the predicted torque at the current moment. .

[0056] The RBF neural network is configured with 4 input neurons, corresponding to fuel injection quantity, intake air quantity, and output torque and their derivatives; and 1 output neuron, corresponding to predicted torque. Simulated annealing (SA) is introduced when training the RBF neural network in the PSO. Finally, the optimal parameter width of the Gaussian kernel function is obtained through iteration. and center and optimal connection weights .

[0057] Step 4, MPC torque prediction control, includes the following steps: Step 4.1, Setting the Reference Trajectory: The purpose is to ensure that the system output torque reaches the target torque along a smooth reference trajectory, thus smoothing the control process. The specific calculation formula is as follows:

[0058] in, The reference trajectory representing the target torque requirement. This indicates the actual output torque of the engine. Indicates the target torque requirement. Indicates the prediction time domain, Represents the reference trajectory coefficients, and .

[0059] Step 4.2, Torque Multi-Step Prediction: To achieve multi-step prediction, multiple neural networks need to be connected in series, using the predicted output value of the previous neural network as the input value of the next neural network; After one prediction, the predicted sequence is obtained. .

[0060] Step 4.3, Feedback and Correction: The measured data of engine output torque at any given time are used to perform prediction compensation on the torque prediction sequence obtained in Step 4.2. The specific calculation formula is as follows:

[0061] in, This indicates that the neural network predicts the output torque. This represents the error correction factor. express The deviation between the actual output torque of the engine and the output torque predicted by the neural network at any given time.

[0062] Step 4.4, Rolling Optimization: Using SQP to solve for a constrained quadratic performance index as the loss function, the optimal injection quantity sequence in the finite time domain is obtained. The loss function is chosen as follows:

[0063] in, Indicates the target torque requirement. This represents the predicted torque after feedback correction. Indicates the amount of change in fuel injection. This represents the torque deviation between the target required torque and the predicted torque. Indicates the first Weighting factor for step-by-step torque prediction error Indicates the first Weighting factor for step-by-step fuel injection variation. Indicates the prediction time domain, This indicates the control time domain.

[0064] During engine operation, constraints such as maximum and minimum fuel injection limits, as well as the range of air-fuel ratio or excess air coefficient, also need to be considered. These system constraints are expressed as follows:

[0065]

[0066] in, and This indicates the minimum and maximum fuel injection quantities at different crankshaft speeds. and This indicates the optimal air-fuel ratio range for the engine.

[0067] Step 4.5, Determine the future fuel injection quantity: Take the first component of the optimal fuel injection quantity sequence obtained in Step 4.4 as... The timing control parameters act on the fuel injection module of an aircraft piston engine; Repeat Steps 4.1 to 4.5 continuously until the process ends.

[0068] In Step 2, the formula for min-max normalization is:

[0069] in, This represents the data after min-max normalization. Represents the original data. and This represents the minimum and maximum values ​​of the sample data.

[0070] In Step 3, the activation function for the hidden layer of the RBF neural network is chosen to be a Gaussian kernel function, as shown in the formula:

[0071] Then network output The formula is:

[0072] in, This represents the width of the Gaussian kernel function. This indicates the center of the Gaussian kernel function. This represents the connection weights between the hidden layer and the output layer.

[0073] The position information of each particle is determined by the width of the Gaussian kernel function. and center and connection weights Composition, when the number of neurons in the hidden layer is At that time, the particle position information is represented as:

[0074] in Represents the width of the Gaussian kernel function , Represents the center of the Gaussian kernel function , Represents the connection weights .

[0075] The simulated annealing mechanism introduced when PSO trains the torque prediction model of the RBF neural network is manifested in that the jump probability of each particle is calculated in each iteration. When the jump probability is greater than the random jump threshold, the current particle is used to replace the global optimal particle for particle velocity update.

[0076] The formula for calculating the probability of a sudden jump is:

[0077] in, Indicates the first The probability of a particle jumping out. Indicates the first The fitness value of each particle. This represents the fitness value of the globally optimal particle. This indicates the current annealing temperature.

[0078] The fitness function selects the mean absolute error of the network output:

[0079] in, Indicates the first The average absolute error of each particle Indicates the number of training samples. This represents the actual output of the neural network. This represents the expected output of the neural network.

[0080] The formulas for updating particle velocity and position are:

[0081]

[0082] in, Represents particle velocity. Indicates the particle position. Indicates inertia weight, and Represents the learning factor. and Represents a random number. , Represents the optimal particle for an individual. This represents the globally optimal particle.

[0083] Example 2

[0084] This invention first trains an RBF neural network using a particle swarm optimization method incorporating simulated annealing to construct a torque prediction model for a compression-ignition aero-piston engine. Compared to gradient descent and conventional particle swarm optimization, this method yields a torque prediction model with higher accuracy and better performance. Then, this torque prediction model is used for multi-step prediction. Next, a performance index is constructed using the difference between the target torque demand and the torque after feedback correction, along with the fuel injection increment. Sequential quadratic programming is then used to perform rolling optimization of this performance index within a finite-time domain to solve for the optimal fuel injection sequence, thus achieving torque control of the compression-ignition aero-piston engine. This method leverages the advantages of the SA-PSO method and the RBF neural network to construct an accurate engine torque prediction model; utilizing the advantages of model predictive control, it achieves excellent torque control performance for the compression-ignition aero-piston engine under different operating conditions.

[0085] Example 3

[0086] like Figure 1As shown, a torque prediction control method for a compression-ignition aero-piston engine includes the following steps: Step 1: Obtain and collect historical running data.

[0087] To train a torque prediction model for an aero-piston engine, historical operating data of the engine is needed as a training sample set. First, the main frame of a compression-ignition aero-piston engine is built in the LMS Imagine.Lab AMESim simulation platform, and simulation parameters are set according to actual engine parameters, such as... Figure 2 As shown; then, a controller was designed in MATLAB / Simulink using a lookup table control method and simulated operation under different working conditions for a total of 220s with a sampling time of 0.01s. The operation data mainly included fuel injection quantity, intake air quantity, output torque and its derivative, etc.

[0088] Step 2: Data min-max normalization processing.

[0089] Because the operating data has different units and ranges, the accuracy of neural network training may decrease and convergence may be difficult. Before importing historical operating data into the neural network training program, the fuel injection quantity, intake air quantity, output torque and their derivatives need to be normalized to the range of (0, 1) using min-max normalization, which can improve the training efficiency and accuracy of the RBF neural network.

[0090] The formula for min-max normalization is:

[0091] in, This represents the data after min-max normalization. Represents the original data. and This represents the minimum and maximum values ​​of the sample data.

[0092] Step 3, train the torque prediction model. Establish a multi-input single-output mathematical model as follows:

[0093] The input to the prediction model is the intake volume at the previous time step. Fuel injection quantity at the previous moment The actual crankshaft output torque at the previous moment and the derivative of the actual crankshaft output torque The output is the predicted torque at the current moment. .

[0094] The activation function for the hidden layers of an RBF neural network is a Gaussian kernel function, and the formula is as follows:

[0095] Then network output The formula is:

[0096] in, This represents the width of the Gaussian kernel function. This indicates the center of the Gaussian kernel function. This represents the connection weights between the hidden layer and the output layer.

[0097] The training flowchart for the torque prediction model based on the RBF neural network is as follows: Figure 3 As shown.

[0098] First, initialize parameters such as particle position and velocity. The position information of each particle is determined by the width of the Gaussian kernel function. and center and connection weights Composition, when the number of neurons in the hidden layer is At that time, the particle position information is represented as:

[0099] in Represents the width of the Gaussian kernel function , Represents the center of the Gaussian kernel function , Represents the connection weights .

[0100] Next, the fitness function is set and calculated to determine the initial optimal individual particle and the global optimal particle. The mean absolute error of the network output is chosen as the fitness function.

[0101] in, Indicates the first The average absolute error of each particle Indicates the number of training samples. This represents the actual output of the neural network. This represents the expected output of the neural network.

[0102] Then, a simulated annealing (SA) mechanism is introduced to update particle position and velocity information. During each iteration, the jump probability of each particle is calculated. When the jump probability is greater than the random jump threshold, the current particle is used to replace the globally optimal particle for particle velocity update.

[0103] The formula for calculating the probability of a sudden jump is:

[0104] in, Indicates the first The probability of a particle jumping out. Indicates the first The fitness value of each particle. This represents the fitness value of the globally optimal particle. This indicates the current annealing temperature.

[0105] The formulas for updating particle velocity and position are:

[0106]

[0107] in, Represents particle velocity. Indicates the particle position. Indicates inertia weight, and Represents the learning factor. and Represents a random number. , Represents the optimal particle for an individual. This represents the globally optimal particle.

[0108] Finally, determine if the termination condition is met. Training ends when the method reaches the maximum number of iterations; otherwise, repeat the above training process until training ends.

[0109] Step 4, MPC torque predictive control. For example... Figure 5 As shown, the overall control framework is roughly divided into three parts. The first part is the demand torque calculation part, which consists of modules such as throttle pushrod torque, propeller demand torque, starting demand torque, idle speed demand torque, accessory and friction demand torque, and cruise control demand torque correction. Finally, the demand torque coordinator calculates the target demand torque under different operating conditions. The second part is the torque prediction control part, which mainly consists of reference trajectory, rolling optimization, prediction model, and feedback correction. This part completes the conversion of the target demand torque into the total fuel injection quantity. The third part is the fuel injection module, which mainly consists of fuel injection pressure MAP, injection advance angle MAP, and injection pulse width MAP. This part converts the calculated total fuel injection quantity into the corresponding three injection parameters: fuel injection pressure, injection advance angle, and injection pulse width, and controls the engine torque output. Step 4 includes the following steps.

[0110] Step 41, Set the reference trajectory: The purpose is to ensure that the system output torque reaches the target torque along a smooth reference trajectory, thus making the control process smooth. The specific calculation formula is as follows:

[0111] in, The reference trajectory representing the target torque requirement. This indicates the actual output torque of the engine. Indicates the target torque requirement. Indicates the prediction time domain, Represents the reference trajectory coefficients, and .

[0112] Step 4.2, Multi-step Torque Prediction: (e.g.) Figure 4 As shown, to achieve multi-step prediction, multiple neural networks need to be connected in series, using the prediction output of the previous neural network as the input value of the next neural network; After one prediction, the predicted sequence is obtained. .

[0113] Step 4.3, Feedback and Correction: The measured data of engine output torque at any given time are used to perform prediction compensation on the torque prediction sequence obtained in Step 4.2. The specific calculation formula is as follows:

[0114] in, This indicates that the neural network predicts the output torque. This represents the error correction factor. express The deviation between the actual output torque of the engine and the output torque predicted by the neural network at any given time.

[0115] Step 4.4, Rolling Optimization: Using SQP to solve for a constrained quadratic performance index as the loss function, the optimal injection quantity sequence in the finite time domain is obtained. The loss function is chosen as follows:

[0116] in, Indicates the target torque requirement. This represents the predicted torque after feedback correction. Indicates the amount of change in fuel injection. This represents the torque deviation between the target required torque and the predicted torque. Indicates the first Weighting factor for step-by-step torque prediction error Indicates the first Weighting factor for step-by-step fuel injection variation. Indicates the prediction time domain, This indicates the control time domain.

[0117] During engine operation, constraints such as maximum and minimum fuel injection limits, as well as the range of air-fuel ratio or excess air coefficient, also need to be considered. These system constraints are expressed as follows:

[0118]

[0119] in, and This indicates the minimum and maximum fuel injection quantities at different crankshaft speeds. and This indicates the optimal air-fuel ratio range for the engine.

[0120] Step 4.5, Determine the future fuel injection quantity: Take the first component of the optimal fuel injection quantity sequence obtained in Step 4.4 as... The timing control parameters act on the fuel injection module of an aircraft piston engine; Repeat Steps 4.1 to 4.5 continuously until the process ends.

[0121] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0122] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A torque prediction control method for a compression-ignition aero-piston engine, characterized in that: Includes the following steps: Step 1: Obtain and collect historical operation data; Step 2, data min-maX normalization processing; Step 3, train the torque prediction model: Step 4, MPC torque prediction control; In Step 1, a physical model of a compression-ignition aero-piston engine is established in the LMS Imagine.Lab AMESim simulation software, and a controller is established in MATLAB / Simulink. The control method is MAP lookup table control. The engine physical model and the controller are connected through a co-simulation interface, and simulations of the engine start-up phase and different operating conditions are performed to obtain historical engine operating data. The historical operating data mainly includes fuel injection quantity, intake air quantity, and output torque and its derivative. In Step 2, all the collected historical operation data are mapped to the interval (0, 1); In Step 3, a multi-input single-output mathematical model is established as follows: Tp(k)=f[ma(k-1),mf(k-1),Tac(k-1),Ta (k-1)] The torque prediction model has four inputs and one output. The inputs to the prediction model are the intake air volume ma(k-1) and the fuel injection volume mf(k-1) from the previous time step, respectively. The actual crankshaft output torque Tac(k-1) and the differential Ta of the actual crankshaft output torque. (k-1), output It is the predicted torque Tp(k) at the current moment; The RBF neural network is set to have 4 input neurons, corresponding to fuel injection quantity, intake air quantity, and output torque and their derivatives; the output neurons are set to 1, corresponding to predicted torque; simulated annealing (SA) is introduced when training the RBF neural network in PSO; finally, the optimal parameter width b and center c of the Gaussian kernel function and the optimal connection weight w are obtained through iteration. Step 4 includes the following steps: setting a reference trajectory, multi-step torque prediction, feedback correction, rolling optimization, and determining the future fuel injection quantity; In Step 4, a reference trajectory is set so that the system output torque follows a smooth reference trajectory to achieve the target torque, thus ensuring a smooth control process. The specific calculation formula is as follows: Tref(k+i)=αiTac(k)+(1-αi)Tta,i=1,2, ,p Where Tref represents the reference trajectory of the target torque demand, Tac represents the actual output torque of the engine, Tta represents the target torque demand, p represents the prediction time domain, α represents the reference trajectory coefficient, and 0 < α < 1. The torque multi-step prediction requires multiple neural networks connected in series, with the prediction output of the previous neural network serving as the input for the next. After p predictions, a prediction sequence is obtained. p= Tp(k+1), Tp(k+2), ,Tp(k+p) ; The feedback correction involves using the measured engine output torque data at time k to compensate for the torque prediction sequence obtained from the multi-step torque prediction. The specific calculation formula is as follows: Tfc(k+i)=Tp(k+i)+μ.e(k),i=1,2, ,p Where Tp represents the output torque predicted by the neural network, μ represents the error correction factor, and e(k) represents the deviation between the actual output torque of the engine and the output torque predicted by the neural network at time k; The rolling optimization uses SQP to solve for a constrained quadratic performance index as the loss function to obtain the optimal injection quantity sequence in the finite time domain. The loss function is selected as follows: Where Tref represents the target torque demand, Tfc represents the predicted torque after feedback correction, Δmf represents the fuel injection variation, and eTor represents the torque deviation between the target torque demand and the predicted torque, and eTor = Tref - Tfc. This represents the weighting factor for the predicted torque error in the i-th step. The weighting factor represents the change in fuel injection at step j, p represents the prediction time domain, and u represents the control time domain; During engine operation, constraints such as maximum and minimum fuel injection limits, as well as the range of air-fuel ratio or excess air coefficient, also need to be considered. These system constraints are expressed as follows: mf,min <mf(k)<mf,max λmin<λ(k)<λmax Where mf,min and mf,max represent the minimum and maximum fuel injection quantities at different crankshaft speeds, and λmin and λmax represent the optimal air-fuel ratio range of the engine. The determination of the future fuel injection quantity uses the first component of the obtained optimal fuel injection quantity sequence as the control quantity at the k-th moment, which acts on the fuel injection module of the aero-piston engine; repeat Step4 at the (k + 1)-th moment until the operation ends.

2. The torque prediction control method for a compression-ignition aero-piston engine according to claim 1, characterized in that: In the said Step2, the formula for min-max normalization is: where, xnom represents the data after min-max normalization, x represents the original data, and xmin and xmax represent the minimum and maximum values of the sample data.

3. The torque prediction control method for a compression-ignition aero-piston engine according to claim 1, characterized in that: In the said Step3, the activation function of the RBF neural network hidden layer selects the Gaussian kernel function, and the formula is: Then the formula for the network output yp is: where, b represents the width of the Gaussian kernel function, c represents the center of the Gaussian kernel function, and w represents the connection weight value between the hidden layer and the output layer; The position information of each particle consists of the width b and the center c of the Gaussian kernel function and the connection weight value w. When the number of hidden layer neurons is m, the particle position information is expressed as: pos=[pos1, ,posm,posm+1, ,pos5m,pos5m+1, ,pos6m] where pos1~posm represent the widths b1~bm of the Gaussian kernel function, posm+1~pos5m represent the centers c1,1~c4,m of the Gaussian kernel function, and pos5m+1~pos6m represent the connection weights W1~Wm.

4. The torque prediction control method for a compression-ignition aero-piston engine according to claim 1, characterized in that: In the said Step3, when the PSO trains the RBF neural network torque prediction model, a simulated annealing mechanism is introduced, which is manifested as calculating the jump probability of each particle in each iteration process. When the jump probability is greater than the random jump threshold, the current particle is used to replace the global optimal particle for particle velocity update; The formula for the jump probability is: =exp where, Psc(i) represents the jump probability of the i-th particle, Error(i) represents the fitness value of the i-th particle, Errorleast represents the fitness value of the global optimal particle, and Tem represents the current annealing temperature; The fitness function selects the mean absolute error of the network output: where, Error(i) represents the mean absolute error of the i-th particle, NS represents the number of training samples, ym represents the actual output of the neural network, and ys represents the expected output of the neural network; The formulas for particle velocity and position update are: Vi(t + 1)=0.Vi(t)+c1.r1.(pBi(t)−posi(t))+c2.r2.(gB(t)−posi(t)) posi(t + 1)=posi(t)+Vi(t + 1) where, Vi represents the particle velocity, pos represents the particle position, 0 represents the inertia weight, c1 and c2 represent the learning factors, r1 and r2 represent random numbers, r1,r2∈[0,1], pBi represents the individual optimal particle, and gB represents the global optimal particle.

5. A torque prediction control method for a compression ignition aero-piston engine according to any one of claims 1-4 discloses an application of the torque prediction control method for a compression ignition aero-piston engine in the technical field of torque prediction control methods for compression ignition aero-piston engines.

Images