Dynamic learning-based fault tolerant control for robotic manipulators with actuator faults

Abstract

In this paper, a learning-based active fault-tolerant control (FTC) scheme for robot manipulators with uncertainties and actuator faults is proposed. Unlike traditional FTC methods, with dynamic learning theory, both uncertainties and actuator faults can be accurately identified/learned by radial basis function networks. Based on the learned knowledge, dynamical classifiers and experience-based controllers corresponding to different fault modes are constructed. With the help of dynamical classifiers, fault detection and isolation can be obtained rapidly and accurately, and the correct experience-based controller (instead of the controller reconfigured online) corresponding to the current fault system is selected to compensate for faults, and superior control performance is achieved, even in the presence of faults. The simulation studies demonstrate the feasibility of the proposed FTC method.

Introduction

With the increasing level of industrial automation, robotic systems, as an important component of several complex engineering systems, play an irreplaceable role. In such systems, owing to increasingly strict requirements for operation and productivity, robot manipulators generally approach their design limits. Such conditions may often cause failures in the robot system, such as sensor faults, actuator faults, or unpredictable nonlinear dynamic changes. Therefore, automated health monitoring diagnostics of robot systems and intelligent fault-tolerant control (FTC) of faulty robot systems play a vital role in using robot manipulators without human intervention [1].

Over the past decades, numerous fault diagnosis (FD) algorithms for robot systems have been investigated based on model-based approaches [2], [3] in which an accurate system model is required. However, this requirement is generally difficult to satisfy in most practical applications, particularly in complex nonlinear systems [4]. Subsequently, using the general approximation characteristics of neural networks (NNs), approximation-based learning methodologies have been proposed to learn the anomalies and faults from a monitored robot system [5], [6], and the residual vector is then generated. If a fault is detected, the residual vector is classified using NNs, and then the occurring fault is isolated [7], [8]. In [9], robust FD methods using the learning method for robot manipulators were proposed. By using NNs, not only the occurrence of faults could be detected but also the types of faults could be identified and isolated.

Fault information (FI) obtained from fault detection and isolation (FDI) procedures is useful for FTC, and the influence of a fault is compensated by reconfiguring the controller with the obtained FI. Thus, even if a fault occurs, the faulty system can continue to operate reliably [10]. The basic idea of the FTC scheme based on FDI is that an FD module is employed to detect and isolate the fault (in which a series of online adaptive estimators is included, one is used for fault detection, and the others are used for isolation) based on the FI acquired during the FDI procedure. Then, the FTC component is implemented to reconfigure the controller for a better control performance [11], [12], [13]. Various FTC schemes have been studied for robot systems [9], [14], [15], [16], [17], [18]. For instance, a reconfigurable FTC approach for a robot arm was developed using a sliding-mode observer technique [16]. In [9], [14], based on FI, a nominal computed torque controller (CTC) was reconfigured to control a faulty robot system. However, the use of CTC often cannot guarantee the desired control performance owing to system uncertainties and faults [15]. Some adaptive-based control methods have been proposed to improve the control performance of robot systems during normal (fault-free) operations [19], [20]. In [16], [21], reconfigurable FTC schemes were presented for mechanical systems in which both actuator faults and uncertainty were considered. Huang et al. [22] studied the fault compensation control problem of robot manipulators with actuator failures and disturbances. For autonomous underwater vehicles, Zhu et al. [23] provided a model parameter-free control strategy to address the trajectory-tracking problem under actuator failures. With NNs [24], [25], Shen et al.[26] proposed an NN-based active FTC scheme with nonlinear faults, and an observer-based FTC method was proposed for induction motors with possible actuator faults [27]. In [28], a robust adaptive FTC problem was investigated for robot systems with actuator failures using the sliding mode technique in which NNs were used to compensate for failures and uncertainties. For the aforementioned FD and FTC schemes, NNs are often utilized to estimate the fault model and reconfigure the controller online. However, the learning ability in the adaptive process, that is, the real intelligent capabilities for knowledge acquisition and re-utilization, has not been fully explored. As a result, when performing even the same control task, the estimated parameters have to be re-adapted online, which leads to the consumption of computing resources and limits the implementation in practical applications.

Learning ability, that is, acquiring knowledge from a dynamic environment and using learned knowledge to improve control performance, is the core of an intelligent control system [29], [30]. In recent years, Wang and Hill proposed a deterministic learning (DL) [also known as dynamic learning] method to acquire knowledge from system identification and adaptive neural control (ANC) [31], [32]. Based on DL, the inherent unknown dynamics of the system can be accurately approximated by RBF NNs in the system identification or stable control processes. Using the DL method, numerous interesting results regarding learning control and applications of small FD have been proposed [33], [34], [35], [36], [37], [38]. However, these learning control schemes are mainly considered in a single control environment; that is, it is often assumed that no faults would occur. For robot systems, Chen et al. [39] attempted to investigate FD for robotic manipulators, in which a conventional CTC was used as the executive controller. Although fault isolation is achieved in some sense, owing to the lack of decision-making and control strategies, it is impossible to achieve compensation control for the faulty system. To realize high-performance control in fault environments, it is important to develop intelligent FTC for complex robot systems.

In this paper, we present a novel active FTC scheme for robot manipulators with actuator faults, which will provide a new concept for intelligent FTC of systems in fault environments. The proposed scheme comprises two modules: a rapid FDI module and an FTC module. In the FDI module, first, under the normal controller, the uncertainties and unknown dynamics of each actuator fault (including the fault-free mode) are accurately identified via DL, utilizing the acquired knowledge, and a series of dynamical estimators are constructed to represent different fault modes of the robot manipulator. Then, by comparing the monitored robot system with the constructed estimators, a set of residuals is generated, and rapid FDI can be achieved simultaneously using the minimum residual principle. In the FTC module, an adaptive neural controller is first designed for the fault-free mode and all possible actuator fault modes, and the closed-loop inherent dynamics of the faulty manipulators are accurately identified by the RBF NNs. Using the learned constant NNs, a set of candidate experience-based fault-tolerant controllers (EBFTCs) is constructed, where each controller corresponds to a specific type of actuator fault mode. Then, when a fault reoccurs, based on the rapid FDI results, the correct EBFTC can be directly selected and used to achieve stability and improve control performance.

The main contributions of this study are as follows: 1) under the proposed FTC scheme, both the uncertainties and actuator faults can be accurately modeled/learned by RBF networks with the satisfaction of persistent excitation (PE), accurate classification of different fault modes can be achieved, and a specific knowledge-based controller library is constructed for all possible fault modes; 2) no online adjustment of any estimated parameters is required in the processes of both FDI and FTC, the calculation burden is significantly reduced, and rapid FD can be achieved; and 3) with the help of rapid FDI, the correct experience-based controller corresponding to the current fault can be directly selected, and the knowledge embedded in the controller can rapidly and correctly compensate for the effect of the actuator fault. Different from existing FTC approaches [11], [12], [24], [26], the proposed scheme is based on the acquisition of fault knowledge, rapid and accurate FD, and precise compensation control. Under the action of the experience-based controller, the FI can be effectively retained, avoiding the cover-up of the fault under the action of adaptive control, particularly for small faults.

The remainder of this paper is organized as follows: Section 2 presents the problem formulation and the preliminaries. The rapid FDI strategy is designed in Sections 3 and 4, fault-tolerant controllers are designed, and the implementation of the rapid FTC scheme is addressed. Section 5 presents simulation studies to illustrate the validity of the proposed FTC scheme. Finally, Section 6 concludes the paper.