On-line anomaly detection with advanced independent component analysis of multi-variate residual signals from causal relation networks

Abstract

Anomaly detection in todays industrial environments is an ambitious challenge to detect possible faults/problems which may turn into severe waste during production, defects, or systems components damage, at an early stage. Data-driven anomaly detection in multi-sensor networks rely on models which are extracted from multi-sensor measurements and which characterize the anomaly-free reference situation. Therefore, significant deviations to these models indicate potential anomalies. In this paper, we propose a new approach which is based on causal relation networks (CRNs) that represent the inner causes and effects between sensor channels (or sensor nodes) in form of partial sub-relations, and evaluate its functionality and performance on two distinct production phases within a micro-fluidic chip manufacturing scenario. The partial relations are modeled by non-linear (fuzzy) regression models for characterizing the (local) degree of influences of the single causes on the effects. An advanced analysis of the multi-variate residual signals, obtained from the partial relations in the CRNs, is conducted. It employs independent component analysis (ICA) to characterize hidden structures in the fused residuals through independent components (latent variables) as obtained through the demixing matrix. A significant change in the energy content of latent variables, detected through automated control limits, indicates an anomaly. Suppression of possible noise content in residuals—to decrease the likelihood of false alarms—is achieved by performing the residual analysis solely on the dominant parts of the demixing matrix. Our approach could detect anomalies in the process which caused bad quality chips (with the occurrence of malfunctions) with negligible delay based on the process data recorded by multiple sensors in two production phases: injection molding and bonding, which are independently carried out with completely different process parameter settings and on different machines (hence, can be seen as two distinct use cases). Our approach furthermore i.) produced lower false alarm rates than several related and well-known state-of-the-art methods for (unsupervised) anomaly detection, and ii.) also caused much lower parametrization efforts (in fact, none at all). Both aspects are essential for the useability of an anomaly detection approach.