Aircraft engine acceleration control plan nested optimization method and device

Abstract

This invention discloses a nested optimization method for aero-engine acceleration control plans, belonging to the field of aero-engine control technology. The method employs a nested inner and outer loop approach to optimize the geometric mechanisms involved in aero-engine control: in the outer loop, global optimization is performed on the geometric mechanisms involved in open-loop control by constructing Bézier curves; in the inner loop, local rolling optimization is performed on the geometric mechanisms involved in closed-loop control; the optimization of the inner loop is based on the outer loop, and the optimization of the outer loop uses the acceleration response time of the inner loop as the evaluation index. This invention also discloses a nested optimization device for aero-engine acceleration control plans. Compared with existing technologies, the control plan optimization process of this invention is more consistent with the actual control process, fully leveraging the advantages of multivariable control, and the acceleration evaluation process is more direct.

Description

Technical Field

[0001] This invention relates to the field of aero-engine control technology, and in particular to an optimization method for aero-engine acceleration control planning. Background Technology

[0002] One of the most important operating processes of an engine is acceleration and deceleration. To ensure that the engine does not overheat, surge, or shut down during acceleration and deceleration, and can respond quickly, it is crucial to design an optimal acceleration and deceleration control law. In the design of acceleration and deceleration control law, the acceleration and deceleration control plan directly affects the response performance and safety of the aero-engine. Therefore, optimization techniques are widely used to optimize the acceleration and deceleration control plan.

[0003] Commonly used acceleration and deceleration (hereinafter referred to as acceleration) control plan optimization methods include methods based on modern artificial intelligence and methods based on classical nonlinear optimization. For example, in the study of turbofan engines, reference [1] combines genetic algorithm and sequential quadratic programming algorithm to find the optimal solution for the acceleration process in which the engine thrust reaches the maximum in the shortest time. Reference [2] establishes an optimization problem for the entire acceleration process of turbofan engines, and uses sequential quadratic programming method to optimize the Bezier curve of the main fuel flow and the throat area of ​​the tail nozzle, and then constructs an acceleration control plan. References [3][4] extend the optimization of the control plan to the full envelope, and formulate a conversion oil-gas ratio control plan based on isotherms and an N-dot control plan based on contour lines. References [5] and [6] use the ISIGHT platform for multidisciplinary optimization design to carry out the design of the transient control plan for variable cycle engines by maximizing the residual power. In these control plan optimization processes, no distinction is made between the open-loop control geometry and the closed-loop control geometry. The same optimization method is used, or only the geometry of the closed-loop control is optimized, which is not conducive to giving full play to the advantages of multivariable control.

[0004] References:

[0005] [1] Shi Peiyan, Gou Linfeng, Guo Jiangwei, et al. Acceleration optimization control of aero-engine based on GA-SQP [J]. Computer and Modernization, 2014, 0(01): 62-66.

[0006] [2]ZHENG Qiangang,ZHANG Haibo.A global optimization control forturbo-fan engine acceleration schedule design[J].Proceedings of the Institution of Mechanical Engineers,Part G:Journal of Aerospace Engineering,2018,232(2):308-316.

[0007] [3] Liu Zihe, Zheng Qiangang, Liu Minglei, et al. Research on improved method of full envelope acceleration control plan for turbofan engine [J]. Propulsion Technology, 2022, 43(01):346-353.

[0008] [4] Li Yuchen, Li Qiuhong, Zhang Xinsheng, et al. N-dot control method for turbofan engine based on active switching logic [J / OL]. Journal of Beijing University of Aeronautics and Astronautics: 1-14 [2023-06-28]. https: / / doi.org / 10.13700 / j.bh.1001-5965.2022.0022.

[0009] [5] Zhu Baibin. Control Plan and Control Algorithm Design for Three-External-Benefit Variable Cycle Engine [D]. Nanjing: Nanjing University of Aeronautics and Astronautics, 2020.

[0010] [6]JIA L,CHEN Y,CHENG R,et al.Designing method of acceleration and deceleration control schedule for variable cycle engine[J].Chinese Journal ofAeronautics,2021,34(05):27-38. Summary of the Invention

[0011] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a nested optimization method for acceleration control plan of aero-engine. The inner loop optimizes the position of the closed-loop geometry to achieve the best control effect, and the outer loop optimizes the position of the open-loop geometry to improve the acceleration performance of the engine.

[0012] The specific technical solution adopted in this invention is as follows:

[0013] A nested optimization method for aero-engine acceleration control plans is proposed, which uses an inner and outer loop nesting approach to perform nested optimization on the geometric mechanisms involved in aero-engine control: in the outer loop, global optimization is performed on the geometric mechanisms involved in open-loop control by constructing Bézier curves; in the inner loop, local rolling optimization is performed on the geometric mechanisms involved in closed-loop control; the optimization of the inner loop is based on the outer loop, and the optimization of the outer loop uses the acceleration response time of the inner loop as the evaluation index.

[0014] Preferably, the global optimization is performed using a swarm intelligence optimization method in the outer loop. The individual dimension is determined based on the number of geometric mechanisms participating in the open-loop control and the control points of the Bézier curve, and the generation range of the initial population is determined based on the adjustment range of the geometric mechanisms.

[0015] Preferably, the control point of the Bézier curve is composed of the rotational speed and the position of the geometric mechanism involved in the open-loop control.

[0016] Preferably, in the inner loop, a prediction model is constructed online based on a neural network state-space model, and the alternating direction multiplier method is used to perform local rolling optimization on the closed-loop control variables; the neural network state-space model includes an aero-engine state-space model and a neural network model, the parameters of the aero-engine state-space model are described by the parameters of the neural network model, and are updated as the parameters of the neural network model are updated online.

[0017] More preferably, the neural network includes a hidden layer, an output layer, and a multiplication layer disposed between the hidden layer and the output layer; the multiplication layer uses the state variables and control variables of the aero-engine state-space model as excitation functions; the hidden layer is divided into n+p groups according to the dimension of the state variables and control variables, where n represents the dimension of the state variables and p represents the dimension of the input variables, and each group contains the same number of j hidden layer nodes. The output of each hidden layer is multiplied by the state variable x and the input variable u in the multiplication layer, respectively. The connection weights between the multiplication layer and the output layer are calculated according to the recursive least squares method.

[0018] Based on the same inventive concept, the following technical solutions can also be obtained:

[0019] A nested optimization device for acceleration control planning of an aero-engine includes nested inner and outer loops for nested optimization of the geometric mechanisms involved in aero-engine control. The outer loop is used for global optimization of the geometric mechanisms involved in open-loop control by constructing Bézier curves. The inner loop is used for local rolling optimization of the geometric mechanisms involved in closed-loop control. The optimization of the inner loop is based on the outer loop, and the optimization of the outer loop uses the acceleration response time of the inner loop as the evaluation index.

[0020] Preferably, the outer loop uses a swarm intelligence optimization method for global optimization, determines the individual dimension based on the number of geometric mechanisms participating in open-loop control and the control points of the Bézier curve, and determines the generation range of the initial population based on the adjustment range of the geometric mechanisms.

[0021] Preferably, the control point of the Bézier curve is composed of the rotational speed and the position of the geometric mechanism involved in the open-loop control.

[0022] Preferably, in the inner loop, a prediction model is constructed online based on a neural network state-space model, and the alternating direction multiplier method is used to perform local rolling optimization on the closed-loop control variables; the neural network state-space model includes an aero-engine state-space model and a neural network model, the parameters of the aero-engine state-space model are described by the parameters of the neural network model, and are updated as the parameters of the neural network model are updated online.

[0023] More preferably, the neural network includes a hidden layer, an output layer, and a multiplication layer disposed between the hidden layer and the output layer; the multiplication layer uses the state variables and control variables of the aero-engine state-space model as excitation functions; the hidden layer is divided into n+p groups according to the dimension of the state variables and control variables, where n represents the dimension of the state variables and p represents the dimension of the input variables, and each group contains the same number of j hidden layer nodes. The output of each hidden layer is multiplied by the state variable x and the input variable u in the multiplication layer, respectively. The connection weights between the multiplication layer and the output layer are calculated according to the recursive least squares method.

[0024] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0025] (1) The control plan optimization process is more consistent with the actual control process: the closed-loop control variables adopt the closed-loop control optimization method, and the open-loop control variables adopt the open-loop optimization method, so the method of obtaining the control plan is closer to the actual use.

[0026] (2) It can give full play to the advantages of multivariable control: Compared with the traditional method of only performing closed-loop control or finite geometric mechanism optimization, all control variables of the present invention can participate in the optimization, which can give full play to the advantages of multivariable control and improve the acceleration performance of the engine.

[0027] (3) More direct evaluation of acceleration process: Compared with traditional optimization methods based on rotational speed or power extraction, this invention directly evaluates individuals based on the shortest acceleration time, and the evaluation index is consistent with the performance evaluation of acceleration process. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of a variable cycle engine.

[0029] Figure 2This is a structural diagram of an acceleration control plan optimization system with nested inner and outer loops;

[0030] Figure 3 It is a neural network state-space model structure;

[0031] Figure 4 It is the fifth-order A obtained through optimization. 224 Bézier curve;

[0032] Figure 5 The acceleration process control effects with and without external loop particle swarm optimization are shown in the dual bypass mode at an altitude of 0 km and Mach number of 0. (a) is the calculated air-fuel ratio control plan curve; (b) is the low-pressure speed N. L Response curve; (c) shows the fuel flow rate W. fm (d) is the variation curve of the nozzle throat area A8; (e) is the variation curve of the rear duct ejector area A8. 163 The curve shows the variation; (f) represents the area A of the front duct ejector. 224 Variation curve; (g) represents the area A of the high-pressure turbine guide vane. HT Variation curve; (h) represents the area A of the low-pressure turbine guide vane. LT Variation curve; (i) represents the high-pressure rotational speed N. H Response curve; (j) is the thrust F response curve; (k) is the total pressure ratio L between the inner and outer inlets. EPR Response curves; (l) is the fan surge margin SMF variation curve; (m) is the CDFS surge margin SMCD variation curve; (n) is the compressor surge margin SMC variation curve; (o) is the compressor outlet pressure P3 variation curve; (p) is the high-pressure turbine inlet temperature T4 variation curve;

[0033] Figure 6 The results of acceleration process control with and without external loop particle swarm optimization are shown in single bypass mode at an altitude of 0 km and Mach number of 0. (a) is the calculated air-fuel ratio control plan curve; (b) is the low-pressure speed N. L Response curve; (c) shows the fuel flow rate W. fm (d) is the variation curve of the nozzle throat area A8; (e) is the variation curve of the rear duct ejector area A8. 163 The curve shows the variation; (f) represents the area A of the front duct ejector. 224 Variation curve; (g) represents the area A of the high-pressure turbine guide vane. HT Variation curve; (h) represents the area A of the low-pressure turbine guide vane. LT Variation curve; (i) represents the high-pressure rotational speed N. H Response curve; (j) is the thrust F response curve; (k) is the total pressure ratio L between the inner and outer inlets. EPRResponse curves; (l) is the fan surge margin SMF variation curve; (m) is the CDFS surge margin SMCD variation curve; (n) is the compressor surge margin SMC variation curve; (o) is the compressor outlet pressure P3 variation curve; (p) is the high-pressure turbine inlet temperature T4 variation curve. Detailed Implementation

[0034] To address the shortcomings of existing technologies, the present invention proposes to use a nested inner and outer loop approach to optimize the geometric mechanisms involved in aero-engine control: in the outer loop, global optimization is performed on the geometric mechanisms involved in open-loop control by constructing Bézier curves; in the inner loop, local rolling optimization is performed on the geometric mechanisms involved in closed-loop control; the optimization of the inner loop is based on the outer loop, and the optimization of the outer loop uses the acceleration response time of the inner loop as the evaluation index.

[0035] Preferably, the global optimization is performed using a swarm intelligence optimization method in the outer loop. The individual dimension is determined based on the number of geometric mechanisms participating in the open-loop control and the control points of the Bézier curve, and the generation range of the initial population is determined based on the adjustment range of the geometric mechanisms.

[0036] Preferably, the control point of the Bézier curve is composed of the rotational speed and the position of the geometric mechanism involved in the open-loop control.

[0037] Preferably, in the inner loop, a prediction model is constructed online based on a neural network state-space model, and the alternating direction multiplier method is used to perform local rolling optimization on the closed-loop control variables; the neural network state-space model includes an aero-engine state-space model and a neural network model, the parameters of the aero-engine state-space model are described by the parameters of the neural network model, and are updated as the parameters of the neural network model are updated online.

[0038] More preferably, the neural network includes a hidden layer, an output layer, and a multiplication layer disposed between the hidden layer and the output layer; the multiplication layer uses the state variables and control variables of the aero-engine state-space model as excitation functions; the hidden layer is divided into n+p groups according to the dimension of the state variables and control variables, where n represents the dimension of the state variables and p represents the dimension of the input variables, and each group contains the same number of j hidden layer nodes. The output of each hidden layer is multiplied by the state variable x and the input variable u in the multiplication layer, respectively. The connection weights between the multiplication layer and the output layer are calculated according to the recursive least squares method.

[0039] The proposed aero-engine acceleration control plan nested optimization device includes nested inner and outer loops for nested optimization of the geometric mechanisms involved in aero-engine control. The outer loop is used for global optimization of the geometric mechanisms involved in open-loop control by constructing Bézier curves. The inner loop is used for local rolling optimization of the geometric mechanisms involved in closed-loop control. The optimization of the inner loop is based on the outer loop, and the optimization of the outer loop uses the acceleration response time of the inner loop as the evaluation index.

[0040] Preferably, the outer loop uses a swarm intelligence optimization method for global optimization, determines the individual dimension based on the number of geometric mechanisms participating in open-loop control and the control points of the Bézier curve, and determines the generation range of the initial population based on the adjustment range of the geometric mechanisms.

[0041] Preferably, the control point of the Bézier curve is composed of the rotational speed and the position of the geometric mechanism involved in the open-loop control.

[0042] Preferably, in the inner loop, a prediction model is constructed online based on a neural network state-space model, and the alternating direction multiplier method is used to perform local rolling optimization on the closed-loop control variables; the neural network state-space model includes an aero-engine state-space model and a neural network model, the parameters of the aero-engine state-space model are described by the parameters of the neural network model, and are updated as the parameters of the neural network model are updated online.

[0043] More preferably, the neural network includes a hidden layer, an output layer, and a multiplication layer disposed between the hidden layer and the output layer; the multiplication layer uses the state variables and control variables of the aero-engine state-space model as excitation functions; the hidden layer is divided into n+p groups according to the dimension of the state variables and control variables, where n represents the dimension of the state variables and p represents the dimension of the input variables, and each group contains the same number of j hidden layer nodes. The output of each hidden layer is multiplied by the state variable x and the input variable u in the multiplication layer, respectively. The connection weights between the multiplication layer and the output layer are calculated according to the recursive least squares method.

[0044] The technical solutions of this invention are applicable to, but are not limited to, turboshaft engines, turboprop engines, turbofan engines, variable cycle engines, and turbine-based ramjet combined engines.

[0045] To facilitate public understanding, the technical solution of the present invention will be described in detail below using a certain type of dual-bypass variable cycle engine as an example, in conjunction with the accompanying drawings:

[0046] The structure of the variable cycle engine is as follows: Figure 1As shown, the main components include: fan, core-driven fan stage (CDFS), mode selection valve (MSV), compressor (COM), front bypass ejector, combustion chamber, high-pressure turbine (HT), low-pressure turbine (LT), rear bypass ejector (RVABI), mixing chamber, afterburner, and exhaust nozzle. Figure A 224 For the area of ​​the adjustable duct ejector, A 163 For the area of ​​the adjustable duct ejector, A HT A LT The area of ​​the high- and low-pressure turbine guide vanes is specified. The engine employs a three-variable closed-loop control method, controlled by the main fuel flow rate W. fm The area of ​​the nozzle throat A8 and the area of ​​RVABI A 163 Closed-loop control of thrust F and low-pressure speed N L The outlet pressure ratio L of the outer bypass duct and the inner bypass duct EPR .

[0047] To obtain the optimal control plan during acceleration, this embodiment uses the component-level mathematical model (CLM) of the variable cycle engine as a basis, and calculates the area A of the front bypass ejector. 224 High-pressure turbine guide area A HT and low-pressure turbine guide area A LT To optimize the invention, the present invention proposes the following: Figure 2 The accelerated control plan optimization structure is shown. The outer loop employs the Particle Swarm Optimization (PSO) algorithm, a commonly used swarm intelligence optimization method, to optimize A. 224 A HT A LT Optimization is performed, with the CLM as the controlled object in the inner loop. Model predictive control is used with alternating direction multiplier method (ADMM) to optimize the current A. 224 A HT A LT The corresponding W fm A8 and A 163 Input is used to obtain the optimal acceleration performance under the constraints. To ensure a smooth transition in the geometric mechanism, Bézier curves are used to construct A. 224 A HT A LT The acceleration process variation curve is used to optimize the control point of the Bézier curve using particle swarm optimization, and low-pressure speed N is used to achieve the same result. L acceleration time t ac To evaluate an individual's fitness.

[0048] The acceleration control plan in this embodiment adopts an offline optimization method. In order to improve the optimization efficiency, a linear prediction model based on the state space model is established, a quadratic performance index is constructed, and the alternating direction multiplier (ADMM) method is used for optimization.

[0049] This embodiment adopts a neural network-based state-space model (NN-SSM) method, which introduces a multiplication layer on the classic neural network, so that the NN output has the expression form of SSM, and significantly improves the modeling accuracy.

[0050] The state-space model of a variable-cycle engine used for model predictive control is expressed as follows:

[0051]

[0052] Where x = [x1, x2] T =[N L N H ] T Here, u represents the state variables, specifically the rotational speeds at low and high pressure, respectively, where u = [u1, u2, u3]. T =[W fm A8, A 163 ] T Let y = [y1, y2, y3] be the closed-loop control variable. T =[N L F, L EPR ] T As the controlled variable, y c =[y c1 ,y c2 ,y c3 ,y c4 ,y c5 ] T =[SMF,SMCD,SMC,P3,T4] T The constraint vectors represent the fan surge margin, CDFS surge margin, compressor surge margin, compressor outlet pressure, and high-pressure turbine inlet temperature, respectively, where α = N. L These are the scheduling parameters for the state variable model. A, B, C, D, C c D c These are the adaptive model matrices, and the subscript k represents the sampling time.

[0053] In order for the neural network to have an expression for the mathematical model of state variables, this embodiment adopts the following... Figure 3 The network structure is shown. A multiplication layer is added between the hidden layers and the output layer. The multiplication layer uses state variables and control variables as activation functions. The hidden layers are divided into n+p groups according to the dimensions of the state variables and control variables, where n represents the dimension of the state variables and p represents the dimension of the input variables. Each group contains j hidden layer nodes. The output of each group is multiplied by the state variable x and the input variable u in the multiplication layer. The connection weights between the multiplication layer and the output layer are calculated using recursive least squares.

[0054] The activation function of the network multiplication layer is the state variable N at time k. L,k NH,k The input W at time k fm,k A 8,k and A 163,k The m-th output of the neural network can then be represented as:

[0055]

[0056] Among them, h i,k =f(W i α k +b i ) represents the output of the i-th hidden layer node, and the scheduling parameter α = N. L Similarly, the input is a neural network. W is the weight between the input layer and the hidden layer, β is the weight between the multiplication layer and the output layer, b is the bias of the hidden layer, and the subscript i indicates that the variable is related to the i-th hidden layer node or the multiplication layer node.

[0057] Due to N L Since the outputs of the SSM in equation (1) are both outputs and state variables, the corresponding C and D matrix elements can be directly determined. Therefore, the nine outputs of the improved neural network correspond to the SSM.

[0058] At this point, the SSM in equation (1) can be written as:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065] Wherein, the subscript l represents the parameter related to the l-th state variable, and the subscript s represents the parameter related to the s-th constraint variable.

[0066] As shown in equation (3), the SSM parameters can be described by neural network parameters and can be updated as the neural network input and weight parameters are updated. If the neural network adopts an online training method, the obtained SSM can characterize the dynamic characteristics of the engine in real time.

[0067] A prediction model is built using online learning of SSM, with n in the control time domain and n in the prediction time domain. u and n y Online learning of SSM is based on future inputs, i.e., the control sequence U = [uk T ,u k+1 T ,…,u k+nu-1 T ] T The future output sequence Y = [y] of the engine is calculated by SSM. k+1 ,y k+2 ,…,y k+ny ] T and the limit quantity prediction sequence Y C =[y c,k+1 ,y c,k+2 ,y c,k+3 ,…,y c,k+ny ] T .

[0068] At each sampling time k, the current state variable x is known. k The system moments A obtained from online learning of the SSM model k B k C k D k C c,k and D c,k We can obtain the future n y The state variables of the step are shown in equation (4):

[0069]

[0070] Similarly, we can obtain the future n y The output variables and constraint variables of the step are shown in equations (5) and (6):

[0071]

[0072]

[0073] Therefore, in the future n y The output vector and constraint vector of the step can be written in the form of equations (7) and (8):

[0074]

[0075] Y C =P C x k +H C U (8)

[0076] Among them, the P and H matrices of the output quantity and the P matrix of the constraint quantity are calculated. C H C The matrices are as follows:

[0077]

[0078]

[0079] Based on the above online prediction model, the following constrained quadratic programming problem is constructed:

[0080]

[0081] Where r is the reference trajectory, and the input reference trajectory sequence R = [r k+1 r k+2 …r k+ny] T Let E be the difference between the reference trajectory and the predicted output, then E = R – Y. I is an identity matrix of appropriate dimension, where the subscript max represents the maximum value and the subscript min represents the minimum value.

[0082] in,

[0083]

[0084]

[0085] u max u min These are the maximum and minimum limits for the closed-loop control quantity.

[0086]

[0087]

[0088] ΔU is the constraint on the variation of the actuator, Δu max The maximum permissible change of the actuator at each simulation moment satisfies:

[0089]

[0090] To meet the restrictive requirements of the implementing agency, H u The matrix can be represented as equation (17) and satisfies equation (18):

[0091]

[0092] H u U≤ΔU (18)

[0093] Therefore, the final constructed matrix satisfies MU≤L.

[0094] To solve the quadratic programming problem with inequality constraints in equation (11), this invention employs the ADMM algorithm and introduces feedback correction to obtain the current A. 224 A HT ALT The optimal input and output sequence of the engine.

[0095] In the process of optimizing the outer loop geometry, A is constructed based on Bézier curves. 224 A HT and A LT The acceleration process curve is optimized with the goal of minimizing the acceleration time at the controlled low-pressure speed. The particle swarm optimization (PSO) algorithm is used to optimize the control points of the Bézier curve, which characterizes the changes in geometric parameters.

[0096] Bézier curves are mathematical curves used in two-dimensional graphics applications. A curve containing n... p +1 control point P i (i = 0, 1, ..., n) p The equation of the Bézier curve can be expressed as a linear combination of nth-order Bernstein polynomials:

[0097]

[0098] Among them, P i (N i A i The parameters are: N represents the rotational speed on the horizontal axis, A represents the geometric area on the vertical axis, and the subscript i represents the (i+1)th control point. t is the proportional parameter; as t changes from 0 to 1, the point on the Bézier curve moves from the starting point to the ending point.

[0099] The particle swarm optimization algorithm is used to optimize the control points. Each geometric mechanism uses 5 control points, corresponding to the low-pressure speed [76%, 82%, 88%, 94%, 102%]. The three open-loop controlled geometric mechanisms have a total of 15 control points. The position vector X of the i-th particle is then represented as X. i =(X i1 ,X i2 ,…,X i15 The velocity vector V is V i =(V i1 V i2 ,…,V i15 The optimal position searched so far for the i-th particle is p. best,i The optimal position found by the particle swarm search is g. best The particle's velocity and position are updated as follows:

[0100]

[0101] Where g represents the current iteration number, ω is the inertia factor, c1 and c2 are non-negative acceleration factors, and r1 and r2 are random numbers between 0 and 1.

[0102] Acceleration time t under the combined action of inner and outer loops ac To determine the fitness of particles.

[0103] mint ac =t2-t1 (21)

[0104] Where t1 is the acceleration start time, and t2 is the time to reach 99% of the target speed.

[0105] By changing the particle's velocity, the particle's position can be altered to obtain different A values. 224 A HT and A LT The change curve, with A 224 For example, the Bezier curve corresponding to the optimal particle is as follows: Figure 4 As shown. By Figure 4 It can be seen that during the process of increasing low-pressure speed, A 224 It exhibits a trend of first decreasing and then increasing, with faster changes at low speeds and slower changes at high speeds. It shows a smooth transition with low-pressure speeds and can be adjusted in an open-loop manner with low-pressure speeds during acceleration.

[0106] Through nested optimization of inner and outer loops, using the Bezier curve constructed by the particle with the shortest acceleration time as the open-loop control scheme, the optimal open-loop and closed-loop input and output sequences of the acceleration process are obtained. Based on the optimal sequence, the fuel and high-pressure rotor speeds N are calculated. H,cor The compressor outlet pressure is used to construct an oil-gas ratio acceleration control plan.

[0107]

[0108] In this context, the subscript cor represents the conversion parameter.

[0109] To verify the effectiveness of the acceleration control plan, simulation verification was conducted in both single-bypass and double-bypass modes of the variable cycle engine. In the double-bypass mode, the acceleration process control plan was optimized at the ground point (H=0km, Ma=0) and the subsonic cruise operating point (H=8km, Ma=0.9). Figure 5 The simulation results for the ground points are given, and A is also shown in the figure. 224 A HT and A LT The simulation results were obtained using a constant (without PSO) model and only the inner-loop predictive control. During the simulation, the fan surge margin was limited to SMFmin = 0.1, the CDFS surge margin was limited to SMCDmin = 0.1, the compressor surge margin was limited to SMCmin = 0.1, the compressor outlet pressure was limited to P3max, and the high-pressure turbine inlet temperature was limited to T4max. Figure 5It can be seen that in the initial stage of acceleration, the engine's output parameters are far from the constraint boundaries. During this stage, the fuel flow rate increases rapidly, and the input is mainly limited by the actuators. As the fuel flow rate continues to increase, the engine's high and low pressure speeds increase, and the operating points of the fan and compressor gradually move towards the surge boundary. In the later stage of acceleration, the engine's high and low pressure speeds continue to increase, and the operating points of the compression components begin to gradually move away from the surge boundary, but the compressor outlet pressure and turbine inlet temperature gradually approach the constraint boundaries. At this point, the low-pressure speed approaches 100%, and the acceleration process ends. Figure 5 (b) It can be seen that after PSO optimization, the low-pressure speed N of the outer loop input trajectory is... L The response time was reduced from 6.7 seconds to 5.55 seconds, achieving 99% accuracy. The results demonstrate that the accelerated control scheme optimized using model predictive control and particle swarm optimization can safely and quickly achieve accelerated process control.

[0110] At the subsonic cruise operating point H=8km and Ma=0.9, the acceleration control plan was also optimized, and the simulation results are consistent with... Figure 5 Similarly, low-pressure speed N L The response time was reduced from 5.575 seconds to 4.949 seconds, achieving 99% response time. This verifies the superiority of the designed acceleration control scheme.

[0111] In single bypass mode, the variable cycle engine underwent acceleration process control plan optimization at both the ground point (H=0km, Ma=0) and the supersonic cruise operating point (H=11km, Ma=1.4). The results at the ground operating point are shown below. Figure 6 . Figure 6 In the single-bypass mode, the acceleration process is similar to that of the double-bypass mode. In the initial stage of acceleration, far from the constraint boundary, the input increases rapidly. In the later stages of acceleration, as the compressor outlet pressure and turbine inlet temperature gradually approach the constraint boundary, the rate of increase in input decreases slightly, and eventually the low-pressure speed stabilizes at 100% of the speed. After optimization, N... L The response time was reduced from 4.15 seconds to 3.9 seconds, achieving 99% response. At the supersonic cruise operating point of H=11km, Ma=1.4, the low-pressure speed response time was reduced from 8.825 seconds to 7.575 seconds, verifying the effectiveness of the inner and outer loop nested acceleration control plan optimization method of the present invention.

Claims

1. A nested optimization method for acceleration control plans of an aero-engine, characterized in that, Nested optimization of the geometric mechanisms involved in aero-engine control is performed using a nested inner and outer loop approach: In the outer loop, global optimization of the geometric mechanisms involved in open-loop control is achieved by constructing Bézier curves; in the inner loop, local rolling optimization is performed on the geometric mechanisms involved in closed-loop control; the optimization of the inner loop is based on the outer loop, and the optimization of the outer loop uses the acceleration response time of the inner loop as the evaluation index; in the inner loop, a prediction model is constructed online based on a neural network state-space model, and the alternating direction multiplier method is used to perform local rolling optimization of the closed-loop control variables; the neural network state-space model includes an aero-engine state-space model and a neural network model, the parameters of the aero-engine state-space model are described by the parameters of the neural network model, and are updated online with the online update of the neural network model parameters.

2. The nested optimization method for aero-engine acceleration control plans as described in claim 1, characterized in that, The global optimization is performed using a swarm intelligence optimization method in the outer loop. The individual dimension is determined based on the number of geometric mechanisms involved in the open-loop control and the control points of the Bézier curve. The generation range of the initial population is determined based on the adjustment range of the geometric mechanisms.

3. The nested optimization method for aero-engine acceleration control plans as described in claim 1, characterized in that, The control point of the Bezier curve is determined by the rotational speed and the position of the geometric mechanism involved in the open-loop control.

4. The nested optimization method for aero-engine acceleration control plans as described in claim 1, characterized in that, The neural network includes a hidden layer, an output layer, and a multiplication layer between the hidden layer and the output layer; the multiplication layer uses the state variables and control variables of the aero-engine state-space model as activation functions; the hidden layer is divided into layers according to the dimension of the state variables and control variables. n + p Group, n The dimension representing the state variable. p The dimension representing the input variables, with the same number of elements in each group. j There are several hidden layer nodes, and the output of each hidden layer is multiplied by the state variables in the multiplication layer. x and input variables u The connection weights between the multiplication layer and the output layer are calculated using the recursive least squares method.

5. A nested optimization device for acceleration control plans of an aero-engine, characterized in that, The system includes nested inner and outer loops for nested optimization of the geometric mechanisms involved in aero-engine control. The outer loop performs global optimization of the geometric mechanisms involved in open-loop control by constructing Bézier curves. The inner loop performs local rolling optimization of the geometric mechanisms involved in closed-loop control. The optimization of the inner loop is based on the outer loop, and the optimization of the outer loop uses the acceleration response time of the inner loop as the evaluation index. In the inner loop, a prediction model is constructed online based on a neural network state-space model, and the alternating direction multiplier method is used to perform local rolling optimization of the closed-loop control variables. The neural network state-space model includes an aero-engine state-space model and a neural network model. The parameters of the aero-engine state-space model are described by the parameters of the neural network model and are updated online with the online updates of the neural network model parameters.

6. The nested optimization device for aero-engine acceleration control plan as described in claim 5, characterized in that, The outer loop uses a swarm intelligence optimization method for global optimization. The individual dimension is determined based on the number of geometric mechanisms participating in the open-loop control and the control points of the Bézier curve. The generation range of the initial population is determined based on the adjustment range of the geometric mechanisms.

7. The nested optimization device for aero-engine acceleration control plan as described in claim 5, characterized in that, The control point of the Bezier curve is determined by the rotational speed and the position of the geometric mechanism involved in the open-loop control.

8. The nested optimization device for aero-engine acceleration control plan as described in claim 5, characterized in that, The neural network includes a hidden layer, an output layer, and a multiplication layer between the hidden layer and the output layer; the multiplication layer uses the state variables and control variables of the aero-engine state-space model as activation functions; the hidden layer is divided into layers according to the dimension of the state variables and control variables. n + p Group, n The dimension representing the state variable. p The dimension representing the input variables, with the same number of elements in each group. j There are several hidden layer nodes, and the output of each hidden layer is multiplied by the state variables in the multiplication layer. x and input variables u The connection weights between the multiplication layer and the output layer are calculated using the recursive least squares method.

Images