Novelty Detection in System Monitoring and Control with HONU

Introduction

The novelty detection in the control loop is important to detect system perturbations as well as to detect system states for which neural network models and controllers has not been properly trained and thus unexpected behavior of a control loop may occur. Such a detection can serve for an early indication of an action to avoid unstable or unpredicted situation in a controlled system.

In principle, novelty carried through measured samples of data may be evaluated either, via probability based approaches as exampled in Markou and Singh (2003a) or, via learning system based approaches, as in the work Markou and Singh (2003b). The first of these streams, i.e. probabilistic, is represented by the statistical approaches of novelty measures and by probabilistic approaches for evaluation of entropy. The Sample Entropy (SampEn) and the Approximate Entropy (ApEn) are very typical and very relevant examples to be mentioned (Pincus, 1991; Richman & Moorman, 2000). These approaches are closely related to the multi-scale evaluation of fractal measures, where further case studies utilizing SampEn, ApEn, and Multiscale Entropy (MSE) can be found in Costa, Goldberger, and Peng (2002) and Yin and Zhou (2012). Further to this, probabilistic entropy approach to the concept shift (sometimes the concept drift) detection in sensory data is reported in Vorburger and Bernstein (2006). The second of the mentioned streams is represented by the utilization of learning systems, such as neural networks and fuzzy-neural systems, and this is also the main area of focus, for the presented work in this paper. During the last three decades of 20th century, the works that were focused in regards to learning systems are that of(Willsky (1976) and Frank (1990) and for incremental learning approach can be referenced for example also the work (Widmer & Kubat, 1996). Then, a particularly focused approach toward the utilization of learning systems, has been rising with works (Polycarpou & Trunov, 2000; Trunov & Polycarpou, 2000). Where, nonlinear estimators and learning algorithm were utilized for the fault detection via the proposed utilization of a fault function that evaluates behavior of residuals of a learning system. Currently, significant research that shall also be referenced is adaptive concept drift detectors, proposed in Alippi, Boracchi, and Roveri (2013). Some readers might also see some analogies of the proposed approach in this paper to the Adaptive Resonance Theory (Grossberg, 2013). Another approach to novelty detection is based on utilization of adaptive parameters of incrementally learning models (neural networks), i.e. the Adaptation Plot (Ivo Bukovsky & Bila, 2010) that has been recently enhanced with multi-scale approach (I. Bukovsky, Kinsner, & Bila, 2012). A most recent method is the Learning Entropy, i.e., a multiscale approach to evaluation of unusual behavior of adaptive parameters of a learning model is introduced in (I. Bukovsky, Oswald, Cejnek, & Benes, 2014).

From recent studies of Bukovsky et al it appears that HONUs are suitable for fast and instant sample-by-sample detection of novel information that each individual sample of data carries. Further, it will be discussed how these principles can be used to distinguish between external and internal perturbations of identified plant or control loops. The simulation result will demonstrates the potentials of the algorithms for single-input plants as well as for some representative of multiple-input plants and for the improvement of their control.

With focus on Higher Order Neural Units (HONUs), this chapter reviews two recently introduced adaptive novelty detection algorithms based on supervised learning of HONU with extension to adaptive monitoring of existing control loops. Further, the chapter also introduces a novel approach for novelty detection in more complex systems via adaptive monitoring of local models with Self-organizing map (SOM) and HONU.