Small oscillation fault detection for a class of nonlinear systems with output measurements using deterministic learning☆
Introduction
Developments of fault detection (FD) approaches for nonlinear systems have received more and more attention in the past decades, principally due to an increasing complexity of modern engineering systems and a higher reliability requirement. One mainstream for fault detection of nonlinear systems is model-based analytical redundancy approach (see, e.g. [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16] and the references therein). There are various approaches for FD using analytical redundancy, such as observer based [4], [5], [6], parameter estimation based [7], [8] and parity space based [9], [10].
In the literature of fault diagnosis, early detection of small faults is an important issue in avoiding catastrophic consequences [3], [12], and is helpful in minimization of maintenance activities and costs [17], [18]. However, small faults are difficult to be detected since they may be hidden by modeling uncertainties. To solve this problem, one promising method is to use neural networks to learn the unknown system dynamics and the fault functions. In [12], [13], [14], adaptive approximation based fault diagnosis approaches have been developed, in which a fault can be detected and learned by using the measured variables. In [15], these techniques are extended to the case that the nonlinearity is modeled as a nonlinear function of the system input and state variables and satisfies a Lipschitz condition.
One key problem of learning the unknown system dynamics and fault functions is that convergence of NN weights to their optimal values requires the satisfaction of the persistent excitation (PE) condition [14], [19], [20]. However, the PE condition is generally very restrictive to be satisfied [14]. Recently, a deterministic learning (DL) theory was proposed for NN approximation of nonlinear dynamical systems with periodic or recurrent trajectories [21], [22], [23]. It is shown that by using localized radial basis function (RBF) neural networks, almost any periodic or recurrent trajectory can lead to the satisfaction of a partial PE condition. This partial PE condition leads to exponential stability of a class of linear time-varying adaptive systems, and accurate NN approximation of the system dynamics is achieved in a local region along the periodic or recurrent trajectory. Further, a DL based observer technique is presented for a class of nonlinear systems undergoing periodic or recurrent motions with only output measurements in [24]. It is showed that learning of the unknown system dynamics is achieved by combining a high gain observer and the DL techniques. The knowledge obtained through DL can be reused in state observation process to achieve a nonhigh-gain design.
Based on the DL theory, in previous works the authors develop a rapid detection scheme for small oscillation faults in the case that the measurements of all system states are available [25]. In practice, measurements of all system states may not be available. In this paper, we extend the previous results by considering a class of nonlinear systems with only output measurements. First, accurate approximation of unknown system dynamics and fault functions are achieved in a local region along the estimated state trajectory by using the DL based observer technique developed in [24]. The obtained knowledge of system dynamics for both normal and fault modes is stored in constant RBF networks.
In the diagnosis phase, a bank of dynamical estimators are constructed for all the normal mode and oscillation faults. The knowledge obtained through DL are reused with a nonhigh-gain design. The detection decision is made based on average norms of the residuals and a smallest residual principle. The occurrence of a fault can be detected if a residual average norm of the residual of a fault estimator becomes smaller than that of the normal estimator in a finite time. A rigorous analysis of the detectability properties of the proposed fault detection scheme is also given, which includes the fault detectability condition and the fault detection time. The attractions of the paper lie in that with output measurements, the knowledge of modeling uncertainty and nonlinear faults is obtained and then utilized to enhance the sensitivity to small faults.
The rest of the paper is organized as follows: Section 2 presents problem formulation and preliminary results. In Section 3, a DL-based scheme for training and rapid detection of oscillation fault modes with output measurements is presented. A rigorous analysis of the performance of the proposed detection scheme is also provided. Section 4 presents the simulation results, and Section 5 concludes the paper.
Section snippets
Problem formulation
Consider a class of oscillation faults generated from the following class of nonlinear systems where is the state vector of the system, is the control input vector, , , are continuous nonlinear functions, with representing the dynamics of the nominal model, the unknown system dynamics, and the deviation in system dynamics due to fault ; represents the disturbances. It is
Training and rapid detection of oscillation faults
In this section, a DL-based scheme for training and rapid detection of oscillation fault modes with output measurement is presented. A rigorous analysis of the performance of the proposed detection scheme is also provided.
Design example
In order to demonstrate the use of the proposed schemes, the single-link robot system is considered [30]: where is the acceleration due to gravity, is the inertia, is the angle position, is the angle velocity, is the angle acceleration, is the length of the link, is the mass of the link, and is the control force.
By introducing the variable changes , and , the above equation can be rewritten in the following form, including the
Conclusions
In this paper, a fault diagnosis scheme for a class of uncertain nonlinear system with output measurement is presented. Firstly, accurate approximation of unknown system dynamics and fault functions are achieved in a local region by combining a high gain observer and a recently developed DL theory. Secondly, the learned knowledge of system dynamics has been reused in a RBFN-based nonlinear observer to achieve rapid detection of the faults. The sensitivity to small faults is enhanced through
References (30)
Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy—a survey and some new results
Automatica
(1990)- et al.
Parametric design of robust fault isolation observers for linear non-square systems
Systems Control Lett.
(2013) - et al.
Nonlinear system fault diagnosis based on adaptive estimation
Automatica
(2004) Fault detection and isolation using parity relations
Control Eng. Pract.
(1997)- et al.
Fault detection filter design for discrete-time nonlinear systems—a mixed optimization
Systems Control Lett.
(2014) - et al.
Fault diagnosis of a class of nonlinear uncertain systems with Lipschitz nonlinearities using adaptive estimation
Automatica
(2010) - et al.
Persistency of excitation and performance of deterministic learning
Systems Control Lett.
(2011) - et al.
A state observer-based adaptive fuzzy output-feedback control for a class of uncertain nonlinear systems
Inform. Sci.
(2010) Model-Based Fault Diagnosis Techniques
(2008)Fault Diagnosis Applications: Model Based Condition Monitoring, Actuators, Drives, Machinery, Plants, Sensors, and Faulttolerant Systems
(2011)
Observer-based strategies for actuator fault detection, isolation and estimation for certain class of uncertain nonlinear systems
IET Control Theory Appl.
A lifting based approach to observer based fault detection of linear periodic systems
IEEE Trans. Automat. Control
Fast fault estimation and accommodation for dynamical systems
IET Control Theory Appl.
Parity space-based fault detection for Markovian jump systems
Int. J. Syst. Sci.
Incipient fault diagnosis of dynamical systems using online approximators
IEEE Trans. Automat. Control
Cited by (29)
Fault diagnosis for a class of nonlinear uncertain systems using deterministic learning approach
2023, Journal of the Franklin InstituteDynamical pattern recognition for sampling sequences based on deterministic learning and structural stability
2021, NeurocomputingCitation Excerpt :Specifically, this method adopts the following ingredients to enable rapid recognition of dynamical patterns: 1) dynamical patterns are represented in a time-invariant and spatially distributed manner through deterministic learning; 2) similarities of dynamical patterns are characterized by the system dynamics inherently within dynamical patterns; 3) rapid recognition is achieved by the state synchronization mechanism according to a kind of internal and dynamical matching on system dynamics. In recent years, this method has been employed to solve many recognition tasks in dynamic environments, such as small fault detection [22–27], gait recognition [28–32], etc. More recently, [33] extended the rapid dynamical pattern recognition method to sampling sequences and showed the good interpretability of dynamical pattern recognition methods under the deterministic learning framework.
Small fault detection from discrete-time closed-loop control using fault dynamics residuals
2019, NeurocomputingCitation Excerpt :Consequently, locally accurate NN approximation of the system dynamics can be obtained, represented, and stored as a constant NN model [29]. Based on the DL theory, novel small fault detection schemes were developed for nonlinear uncertain systems in [20,30,31], which however are focused on continuous-time systems. Since most continuous dynamical systems are nowadays controlled using digital devices [32], discrete-time systems are more veritable to describe a practical problem in modern engineering systems compared with the continuous-time systems.
- ☆
This work was supported by the National Science Fund for Distinguished Young Scholars (Grant No. 61225014), and by the National Natural Science Foundation of China (Grant Nos. 61403087, 60934001).
- 1
D.J. Hill is also with School of Electrical and Information Engineering, the University of Sydney, Sydney, Australia.