Robust Recursive Eigendecomposition and Subspace-Based Algorithms With Application to Fault Detection in Wireless Sensor Networks

Abstract

Abstract-The principal component analysis (PCA) is a valuable tool in multivariate statistics, and it is an effective method for fault detection in wireless sensor networks (WSNs) and other related applications. However, its online implementation requires the computation of eigendecomposition (ED) or singular value decomposition. To reduce the arithmetic complexity, we propose an efficient fault detection approach using the subspace tracking concept. In particular, two new robust subspace tracking algorithms are developed, namely, the robust orthonormal projection approximation subspace tracking (OPAST) with rank-1 modification and the robust OPAST with deflation. Both methods rely on robust M-estimate-based recursive covariance estimate to improve the robustness against the effect of faulty samples, and they offer different tradeoff between fault detection accuracy and arithmetic complexity. Since only the ED in the major subspace is computed, their arithmetic complexities are much lower than those of other conventional PCA-based algorithms. Furthermore, we propose new robust T 2 score and SPE detection criteria with recursive update formulas to improve the robustness over their conventional counterparts and to facilitate online implementation for the proposed robust subspace ED and tracking algorithms. Computer simulation and experimental results on WSN data show that the proposed fault detection approach, which combines the aforementioned robust subspace tracking algorithms with the robust detection criteria, is able to achieve better performance than other conventional approaches. Hence, it serves as an attractive alternative to other conventional approaches to fault detection in WSNs and other related applications because of its low complexity, efficient recursive implementation, and good performance.Index Terms-Fault detection, orthonormal projection approximation subspace tracking (PAST) (OPAST), outlier detection, PAST algorithm with deflation (PASTd), recursive principal component analysis (R-PCA), robust statistics, subspace eigendecomposition (ED), wireless sensor networks (WSNs).

Distributed Real-Time Anomaly Detection in Networked Industrial Sensing Systems

“…However, they either depend on static routing trees, or require accurately assigned thresholds to ensure their detection accuracy. Other approaches, such as [21]- [23], although they provide efficient anomaly-detection solutions, they do not focus on identifying anomalies with respect to each sensor observation, but the anomaly condition of samples (which is a sets of observations) and sensor devices.…”

A Two-Stage Fault Detection and Isolation Platform for Industrial Systems Using Residual Evaluation

“…These measurements are processed under different assumptions about the process, which produces them. From this perspective, FDI algorithms can be categorized into two groups: 1) model-based FDI where the mathematical model of the process is assumed to be known and used [1]- [3] and 2) model-free FDI, which considers FDI as a pattern recognition problem [4]- [6].…”

“…As long as u is determined, (5) can be used to obtain α and b, thus the model (6). In addition, a shrinking grid search algorithm is introduced in [17] for tunning γ and σ , which is used in this paper.…”

“…Then, the final classifier has the form of z( y) = sign( N i=1 α i z i K ( y, y i ) + b). Similar to regression in (6), again RBF kernel is chosen. The kernel and hyperparameters, σ and γ , are assigned by a shrinking grid search [17].…”

Adaptive Online Learning With Regularized Kernel for One-Class Classification

“…KPCA based online one-class classifier (OKPCA) has been developed by Chatzigiannakis and Papavassiliou [23], where the model is constructed in a batch mode for all new samples. Later, an incremental way of model construction for KPCA has been developed for new samples [20], [21].…”