CN110219736B - Aero-engine direct thrust control method based on nonlinear model predictive control - Google Patents

Abstract

The invention discloses a method for controlling the direct thrust of an aircraft engine based on nonlinear model predictive control. The method directly takes the thrust as a control target, and is not a method that the traditional control method adopts an undetectable parameter as the control target. The online sliding window deep neural network is used as a prediction model, the model adopts a deep learning structure, the model precision can be improved, and the sensitivity to the noise of the training data is reduced by selecting the training data online based on a sliding window method. Compared with the current popular control method, the method shortens the acceleration time by 0.425 seconds and improves the response speed by about 1.14 times.

Description

Aero-engine direct thrust control method based on nonlinear model predictive control

Technical Field

The invention relates to an aero-engine control method, in particular to an aero-engine direct thrust control method based on nonlinear model predictive control.

Background

The main function of an aircraft engine is to provide thrust quickly and accurately to an aircraft. Conventional engine control strategies are sensor-based, i.e., indirectly controlling thrust by controlling engine measurable parameters such as rotor speed, Engine Pressure Ratio (EPR), or other measurable parameters. However, the correspondence between engine thrust and measurable parameters may vary due to degradation, manufacturing and manufacturing tolerances, etc. Therefore, if the conventional control concept is adopted, an engine thrust control error inevitably exists. In addition, in order to ensure that the engine can safely and stably operate at the worst operating point, a large safety margin is often maintained in the conventional control strategy, and the performance of the engine at other operating points is greatly limited by the conventional control strategy.

In response to these shortcomings, researchers have invented a new control strategy, Model Based Engine Control (MBEC), to directly control engine performance. The Model Predictive Control (MPC) is an important key technology and research field in model-based engine Control. MPC has better acceleration performance than conventional controllers, and has attracted extensive research interest in the field of aeroengine control. VroemenBG et al applied MPC technology to a laboratory gas turbine. Brunell et al studied the feasibility of state-estimated constrained Nonlinear Model Predictive Control (NMPC) and applied it to a simulation model of a turbojet engine. Decastro proposes an NMPC based on linear variable parameters and is applied to the active tip clearance of a commercial turbofan engine. Richter proposes a multiplexing implementation method, which greatly reduces the calculation amount of the NMPC optimization algorithm. Di Cairano et al developed an MPC strategy to control engine speed to ground slow via electronic throttle and spark timing. The working model prediction model is mainly focused on a linear model, and a good control effect is achieved. However, the aero-engine is a strong nonlinear power device, and the modeling error of the linear model is inevitable. Furthermore, the control targets of conventional NMPCs are measurable parameters. The corresponding relation between the measurable parameters and the thrust or power of the engine changes along with the service time of the engine, so that the thrust control precision is reduced.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provide a method for controlling the direct thrust of an aircraft engine based on nonlinear model predictive control, which can effectively improve the response capability of the engine.

The invention specifically adopts the following technical scheme to solve the technical problems:

a direct thrust control method of an aircraft engine based on nonlinear model predictive control takes thrust as a direct control target and uses a nonlinear model predictive control method for control; specifically, the method is realized by solving the following rolling optimization problem:

wherein r is an engine control command,

for control target prediction, u is the control variable vector, N_f、N_cRespectively the fan speed, the compressor rotor speed, S_mf、S_mcRespectively fan surge margin, compressor surge margin, T₄₁For high pressure turbine inlet temperature, Q and R are positively symmetrical, and N is a positive integer_uAnd N_pRespectively a control time domain and a prediction time domain.

Preferably, a pre-trained online sliding window deep neural network is used as the prediction model of the nonlinear model prediction control method, and the loss function of the online sliding window deep neural network is described as follows:

where x is the input vector, y is the output vector, j is the jth time, L is the length of the rolling interval, f_DNNAnd (5) expressing the mapping of the deep neural network, wherein W is a weight matrix, and b is a bias vector.

Further preferably, the input x of the predictive model_DNNAnd output y_DNNThe method comprises the following specific steps:

wherein, W_fb(k)，N_f(k),N_c(k),S_mf(k),S_mc(k),T₄₁(k) F (k) engine fuel input, fan speed, compressor rotor speed, fan surge margin, compressor surge margin, high pressure turbine inlet temperature, engine thrust, m, at time k, respectively₁,m₂,…,m₇Is a preset positive integer.

Compared with the prior art, the technical scheme of the invention has the following beneficial effects:

the invention can obviously improve the control precision and response speed of the control system of the aero-engine.

Drawings

FIG. 1 is a schematic diagram of a nonlinear model predictive control system according to the present invention;

FIG. 2 is a schematic diagram of the rolling optimization principle;

FIG. 3 is a diagram of a deep neural network;

FIG. 4 is a schematic view of a sliding window principle;

FIG. 5 is a schematic diagram of a back propagation algorithm;

FIGS. 6(a) to 6(g) are the results of simulation of the method of the present invention, wherein F, T₄₁、N_f、N_c、S_mf、S_mcNormalization processing is performed.

Detailed Description

Aiming at the defects in the prior art, the method takes the thrust as a direct control target and uses a nonlinear model predictive control method for control, so that the control precision and the response speed of the control system of the aero-engine are improved.

The technical scheme of the invention is explained in detail in the following with the accompanying drawings:

the structure of the nonlinear model predictive control system of the invention is shown in fig. 1. The engine nonlinear model estimates the unmeasured parameters, the NMP module comprises an optimization submodule and a prediction model submodule, and the online sliding window deep neural network (OL-SW-DNN) is used as a prediction model, so that the method has stronger fitting capability compared with other shallow network structures such as the traditional neural network and support vector regression, and the prediction precision is improved. In the optimization submodule, the optimization objective of the NMPC is the thrust command. It can be seen that the optimization submodule and the prediction model submodule of NMPC are key technologies, as follows.

The nonlinear model predictive control is a closed-loop, limited online optimization control. As shown in fig. 2, at time k, the control targets are:

wherein r is a control command,for control target prediction, u is the control variable vector, and Q and R are positively symmetric. Positive integer N_uAnd N_pRespectively a control interval and a prediction interval.

The control objectives of conventional aircraft engines are always to select measurable parameters such as engine boost ratio or rotor speed. The relationship between thrust and measurable parameters may change over time. Therefore, the control object of the present invention is to directly control the thrust force F.

The operation of the engine is limited by mechanical and component temperature limitations, such as fan and compressor rotor speed limitations, fan and compressor surge margin limitations, high pressure turbine inlet temperature limitations, and the like. In addition, the physical constraints of the actuator should also be considered. Therefore, in order to ensure the safe and stable operation of the engine, the engine operation should satisfy the following constraint conditions:

wherein N is_f、N_cRespectively the rotational speed of the fan and the compressor rotor, S_mf、S_mcSurge margins, T, of fans and compressors, respectively₄₁Is the high pressure turbine inlet temperature.

The control targets and the constraint limits in equations (1) and (2) are calculated by a predictive model. The accuracy of the prediction model will directly affect the control effect. The predictive model of conventional NMPC is a linear model. However, the transient process of an aircraft engine is a strongly nonlinear process. Therefore, the NMPC based on the on-line Sliding Window Deep Neural Network (OL-SW-DNN) as a prediction model is invented.

DNN has a strong fitting ability to nonlinear objects and can be described as:

y＝f_DNN(x) (3)

where x is the input vector and y is the output vector. To maintain engine dynamics, the inputs to the predictive model include current and historical engine fuel inputs W_fbAnd S of historical time_mf,S_mc,N_f,T₄₁And F. The inputs and outputs of the prediction model are:

wherein m is₁,m₂,…,m₇Is related to the non-linearity of the engine. As shown in fig. 3 and 4, the DNN has a multi-layer network structure, and the more hidden layers of the DNN, the stronger the fitting ability of the DNN. Each layer of DNN may be defined as:

a^l+1＝W^lh^l+b^l(5)

h^l+1＝σ(a^l+1) (6)

wherein, W^lIs a weight matrix, b^lIs a bias vector, σ is an activation function, h^l(l>0) Is the output of the l layers, l 1,2, …, n_l,n_lIs the number of implicit nodes per layer.

The loss function of OL-SW-DNN is described as:

w and b are updated as follows:

wherein the content of the first and second substances,are respectivelyη is the learning rate.

And

it can be calculated by the back propagation algorithm shown in fig. 5:

wherein, delta^lThe method comprises the following steps:

wherein, l ═ n_net,n_net-1,L,2，

Is a Hadamard product

Order to

Comprises the following steps:

wherein n is_netIs the number of network layers.

In order to verify the effectiveness of the method, the NMPC based on the method of the invention and the popular NMPC based on Extended Kalman Filter (EKF) are respectively simulated. The engine acceleration process is selected by the simulation process of both methods. The acceleration process starts at a steady-state operating point with the throttle lever angle PLA of 26 ° and ends at a steady-state operating point with the throttle lever angle PLA of 70 °. Both of these simulated operating conditions were standard atmospheric conditions at a height H of 0 km and a mach number Ma of 0. The figure shows EKF-based NMPC and proposed accelerated process simulation of NMPC, where 'NMPC-EKF' denotes EFK-based NMPC and 'NMPC-DNN' denotes OL-SW-DNN-based NMPC. For convenience, the two methods are described below as conventional NMPC and proposed NMPC, respectively. The engine parameters in the map are normalized.

As shown in fig. 6(a), in the conventional NMPC and the proposed NMPC, the time for the engine thrust to increase to 100% thrust is 3.025 seconds and 2.6 seconds, respectively. Compared with the conventional NMPC, the proposed NMPC reduces the acceleration time by 0.425 seconds, and the acceleration speed is increased almost by 1.14 times. The main reason is that conventional NMPC uses a linear model as a prediction model. However, the transient process of the engine is a strongly nonlinear process. With a linear model as the prediction model, prediction errors are inevitable. And the invented NMPC adopts OL-SW-DNN as a prediction model. The OL-SW-DNN has stronger fitting capability to the nonlinear object, and the prediction precision of the NMPC is improved.

As shown in fig. 6(g), the operating point of the engine during acceleration moves along the surge boundary, which is considered to be the path of fastest engine acceleration response. For other limitations, the engine did not reach over-temperature, overspeed, or surge, as shown in fig. 6(c) -6 (f). Therefore, the control method has higher control precision and response speed.

Claims (3)

1. A direct thrust control method of an aircraft engine based on nonlinear model predictive control is characterized in that thrust is taken as a direct control target, and the control is carried out by using a nonlinear model predictive control method; specifically, the method is realized by solving the following rolling optimization problem:

wherein r is an engine control command,for control target prediction, u is the control variable vector, N_f、N_cRespectively the fan speed, the compressor rotor speed, S_mf、S_mcRespectively fan surge margin, compressor surge margin, T₄₁For high pressure turbine inlet temperature, Q and R are positively symmetrical, and N is a positive integer_uAnd N_pRespectively a control time domain and a prediction time domain.

2. The nonlinear model predictive control-based aircraft engine direct thrust control method according to claim 1, characterized in that a pre-trained online sliding window deep neural network is used as a predictive model of the nonlinear model predictive control method, and a loss function of the online sliding window deep neural network is described as:

where x is the input vector, y is the output vector, j is the jth time, L is the length of the rolling interval, f_DNNAnd (5) expressing the mapping of the deep neural network, wherein W is a weight matrix, and b is a bias vector.

3. The nonlinear model predictive control-based aircraft of claim 2Method for direct thrust control of an empty engine, characterized in that the input x of the prediction model is_DNNAnd output y_DNNThe method comprises the following specific steps:

wherein, W_fb(k)，N_f(k),N_c(k),S_mf(k),S_mc(k),T₄₁(k) F (k) engine fuel input, fan speed, compressor rotor speed, fan surge margin, compressor surge margin, high pressure turbine inlet temperature, engine thrust, m, at time k, respectively₁,m₂,…,m₇Is a preset positive integer.

Images