Multi-step predictive compensated intelligent control for aero-engine wireless networked system with random scheduling

Abstract

Within the future application of wireless network for the aero-engine control problem, resource constraints (caused by the limitation of hardware), network traffic restriction, interference and long time-delay, must be considered as one of the difficulties to be solved, thus, the network connection and transmission efficiency can be ensured. In order to ensure the performance of aero-engine nonlinear wireless networked control system in consideration with complicated flight conditions and adverse network environments, a SVM (Support Vector Machine) intelligent inverse controller with a reference model was established. To reduce the offline modeling error with time-varying random scheduling strategy and disturbance, a LSTM (Long Short Term Memory) online multi-step error compensation was designed. Training result, on the four optimal conditions chosen in the flight envelop, provided the model and controller establishing is completed. After introduced into the overall system, checking result, showed the errors and the output with/without LSTM compensation on the other four random conditions, provided the control performance under the three kind of adverse wireless network environments. for comparison, other intelligent algorithms were also adopted to predict the multi-step errors in the reference model, then the accuracy and computing time provided the advantage of LSTM algorithm. The method and strategy proposed in this paper ensure the aero-engine safe working at the price of some control performance loss on the adverse wireless network environments.

Introduction

With the advent of 5 G era, wireless communication technology has been further developed. For reducing the weight of aircraft and avoiding the unreliability of electric transmission medium, the application of wireless control in aircraft is an inevitable development trend. Once wireless communication is used, anti-jamming capability and low bandwidth coping capability will become the first problem to be solved as soon as possible. If the network cannot run normally, in order to ensure the control performance, signal scheduling must be active. Therefore, it is of great significance for the embedded design of scheduling and compensation for nonlinear control.

At present, the aero-engine controller is still in the process of linear control and centralized control, which needs to change the parameters according to different flight conditions. Over the past few decades, tremendous advances in aerothermodynamics and materials have improved the performance and efficiency of engines to an unprecedented degree, challenging the ability of control systems on the other hand. Due to the complexity of flight conditions, the uncertainty of complex wireless network constraints, disturbances and control nonlinearity becomes more and more serious.

In the design of aero-engine controller, flight envelope segmentation is a common method to obtain multiple regions under certain performance indexes. Representative nominal working points of each region are selected to establish the state space model of each small deviation and design the corresponding linear controller. However, when the selected working point is far away from the designed working point, the actual working performance of the aero-engine will produce a large deviation. In addition, due to the complex change of aeroengine operating conditions, system uncertainties such as parameter perturbation and external interference inevitably occur, which have strong nonlinear characteristics and have a great impact on the adaptability of the controller to the flight envelope region. In addition, for the networked control system of aero-engine, scheduling strategy and packet loss cannot be ignored, so the controller design of nonlinear uncertain network constraint system has become a research hotspot in this field.

Nonlinear controller parameter is essential condition for nonlinear aero-engine model, and according to the time-delay combination sufficient condition proposed in Ref. [1], output and control time-delays cannot be combined approximately for aero-engine networked control system. However, it is necessary to consider the two time-delays in reference models, when real-time fault-tolerant control strategies are adopted for the nonlinear aero-engine system. In recent years, researchers have a series of studies on the control problems for nonlinear time-delay systems, and predictive control is widely used to predict the future outputs with the current output and control signals, then the accuracy can be ensured by rolling optimizations for control parameters [2], [3], [4].

From references, controllers are always designed for some certain kinds of systems, which may not possess generalization to be adopted in the nonlinear aero-engine networked control systems directly. Ref. [5] predicted the future outputs by a LS-SVM identifier, and designed an adaptive controller based on PID neural network. However, the single-step prediction method cannot be employed in a random bounded delay condition. Ref. [4,6] achieved the output multi-step prediction for time-delay systems based on NARMA (Nonlinear Auto-Regressive and Moving Average) and neural network, but control delay is not in the consideration. Although both the two delays were studied in the multi-step prediction of Ref. [7,8], the single-step predicted results were processed as certain values which uncertainties were ignored, so larger errors may exist.

Recently in nonlinear NCS (Networked Control System) research, Ref. [9] studied the problem of reliable filter problem for a category of sensor networks in the framework of interval type-2 fuzzy model. Ref. [10] proposed a distributed filtering scheme for the fault detection problem of nonlinear T-S (Takagi-Sugeno) stochastic systems with wireless sensor networks. Ref. [11] analyzed the problem of adaptive fuzzy control for a category of SISO (single-input single-output) nonlinear NCSs with network-induced delay and data loss based on adaptive backstepping control approach. Ref. [12] designed the static output feedback controller for a class of T-S switched fuzzy systems with actuator faults. Thus, based on the traditional state space equation, similar to this kind of fuzzy switching method to conduct nonlinear research, our previous research [13] focused on the co-design of dynamic scheduling and robust H-infinity control method analysis by employing a class of aero-engine T-S fuzzy model based on flight envelop division. However, the aero-engine flight envelope has a wide range and the system parameters at each point in the flight envelope change dramatically, it is difficult to describe with some selected system models for switching or weighted integration. Meanwhile, it will inevitably increase the complexity and the probability of failure if the number of system reference models is increased. Thus, it is necessary to study the global intelligent nonlinear control using artificial intelligence and deep learning algorithms. Ref. [14] considered the nonlinear part of the system as external interference, and then a sliding mode tracking controller is designed based on the assumption that the interference satisfies the sliding mode matching conditions. Ref. [15] provided a linearization method to transform the nonlinear function on the convex hull into a linear function, by which the system with nonlinear terms can be transformed into a class of linear systems with time-varying parameters. Based on adaptive technology, Ref. [16] studied the problem of event-triggered communication, scheduling and fault-tolerant of nonlinear NCSs under constraints. However, this kind of the traditional sate space equation with the nonlinear term, which still belongs to the category of linearization of nonlinear problems after satisfying certain assumptions and matching conditions, could not deal with more complex flight conditions, communication conditions and failure occurrence as actively as artificial intelligence. In addition, the traditional nonlinear control theory is mostly based on the explicit model of the controlled object, which often leads to the complexity of control. The setting of control parameters is mainly to ensure the stability of the system, and no efficient performance oriented nonlinear control theory has been proposed. Thus, Ref. [17] proposed a novel model-free dual neural network to address the learning and control of manipulators simultaneously in a unified framework. Ref. [18] provided performance-oriented non-linear control by mastering non-linear discrete optimal control law in an implicit data-learning manner. Based on a RHONN (recurrent high order neural network), Ref. [19] designed the neural controller which is trained with an extended Kalman filter. However, these completely nonlinear control methods do not include co-design of network communication scheduling strategy. In terms of control system switching, Ref. [20,21] designed the quantized feedback controller and the observer-based controller for the nonlinear Markov jump systems, and it is also necessary to study the switching of scheduling strategies in control systems. In summary, it is necessary to study the intelligent nonlinear control methods considering network smoothness and interference for the future aero-engine ZigBee wireless NCS application. Although this paper does not involve the creation of a new artificial intelligence method, its main contribution and novelty focused on the combination and application of deep learning theory in network control systems with scheduling. The novelty of this paper includes: 1) SVM algorithm is used to deal with implicit nonlinear problems of NCSs; 2) more complex conditions with scheduling strategy, time-delay and disturbing contained; 3) the deep learning LSTM algorithm is employed for on-line compensation to the complex conditions, so the networked control are more active and intelligent; 4) simulation technique of the multiple pulses trigger for the combined scheduling of input and output signals; and 5) aero-engine control application within the full flight envelop data.

Therefore, for aero-engine distributed control system with network constraints, interference and long time-delays, a SVM (Support Vector Machine) intelligent inverse controller with a SVM-NARMA reference model was established, and random scheduling strategy with LSTM (Long Short Term Memory) online multi-step error compensation was designed. Training result, on the four optimal (T-S fuzzy rules) conditions chosen in the flight envelop, provided the model and controller establishing is completed. After introducing into the overall system, checking result, showed the errors and the output with/without LSTM compensation on the other four random conditions, provided the control performance under the three kind of adverse wireless network environments. for comparison, SVM [22], BPANN (Back Propagation Artificial Neural Network) [23], GP (Genetic Programming) [22] were also adopted to predict the multi-step errors in the reference model, then the RMSE (Root Mean Square Error) and computing time provided the advantage of LSTM algorithm. The method and strategy proposed in this paper ensure the aero-engine safe working at the price of some control performance loss on the adverse wireless network environments.