Data‐driven bounded‐error fault detection

Abstract

In this paper, a new data-driven fault-detection method is proposed. This method is based on a new nonparametric system identification approach, which constitutes the principal contribution to this work. The fault-detection method is a parametric model-free approach that can be applied to nonlinear systems that work at various operating points. Not only can the fault-detection process be applied to the steady state of each operating point, but it can also be applied to the transient state resulting from a change in the operating point. In order to detect faults, the proposed method uses an interval predictor based on bounded-error techniques. The utilization of techniques based on bounded error enables system uncertainties to be included in an explicit way. This in turn leads to the possibility of obtaining interval predictions of the behaviour of the system, which include information on the reliability of the prediction itself. In order to show the effectiveness of the fault-detection method, two examples are presented: in the form of a simulated process (counter-flow shell-and-tube heat-exchanger system) and an example of a real application (two-tanks system). A comparison with two fault-detection methods has also been included.This model can be obtained in several different ways, such as using the physical knowledge of the system or by conventional identification methods. Model-based methods (MBMs) can be grouped into a few approaches: signal-processing methods [14], observer-based methods [15,16], parityrelation methods [17,18], parameter-estimation methods [19,20] and Kalman filter-based methods [21,22].The fault-detection method proposed in this work is parametric model free, so this previous knowledge about the system is not necessary. On the other hand, data-driven fault-detection methods use large volumes of data collected from industrial processes. These methods use an inference system based on statistical techniques [23] or by machine learning [9] applied to the historical data available. The most commonly used statistical techniques are principal component analysis (PCA) [12,24], partial least squares (PLS) [12], Fisher's discriminant analysis [25] and independent component analysis [26]. In general, these methods assume a linear relation in the studied variables of the system. To apply these statistical techniques to nonlinear systems, the corresponding nonlinear versions of the statistical techniques should be used: PCA [27,28] and PLS [5,29]. These nonlinear methods are usually related to a nonlinear optimization problem, and an iterative search algorithm is needed to be applied. These iterative search algorithms can suffer from convergence problems and can be overly sensitive to small data fluctuations [12]. The fault-detection method proposed in this work can be used for nonlinear systems; however, an iterative search algorithm is not needed.The most widely used artificial intelligence techniques include neural networks [30], expert systems [31], support vector machines [32], case-based reasoning...

Kernelized relative entropy for direct fault detection in industrial rotary kilns

“…Data‐based or data‐driven methods, however, do not rely on a mathematical model of the process . However, a large amount of process measurement data is assumed to be a priori available to statistically characterize the healthy mode to the diagnostic system.…”

Robust fault detection and isolation scheme using fuzzy wavelet network with a hybrid design algorithm

“…When the measurement is not consistent with any member of this set, a fault is detected. This type of identi cation method has been applied to the development of several fault detection schemes [25][26][27][28][29].…”