Robust Data-Driven Fault Detection in Dynamic Process Environments Using Discrete Event Systems

This chapter is dedicated to the improvement of robustness and performance of continuous data-driven fault detection (FD) systems with the usage of discrete event systems. With data-driven FD it is meant that causal relations and variable dependencies in the system are explored as nominal reference models for detecting atypical occurrences. The models are automatically extracted from data, typically collected within multi-sensor networks, which can be of large-scale nature leading to very high-dimensional learning settings. We will first demonstrate several principal concepts and algorithms for establishing such models in batch, off-line environments, together with advanced anomaly and fault detection strategies based on these models. Then, we will explain concepts how to address system dynamics properly and with sufficient accuracy by updating these models on demand, on the fly and fully autonomously during on-line processing mode. The problem therein is that upcoming failures also trigger dynamic changes in the process, which have to be distinguished from intended (non-failure) changes, as these should not be respected in model adaptation, obviously (as they would induce a deteriorating performance). We thus will demonstrate how signals from discrete event systems can be hybridized with multi-sensor measurement systems (continuously recorded channel signals) in order to properly realize such a distinction. This hybridization can be seen as a new form of hybrid dynamic systems which we see as necessary for preventing the time-intensive a priori collection of typical fault patterns or fault signatures (which are mostly application dependent) and to increase the level of automatization. Therein, a specific automated (data-driven) fault isolation technique acting on the residual signals and on the statistics extracted from causal relation networks serves as methodological back-bone.