Dynamic Anomaly Detection based on Recursive Independent Component Analysis of Multi-Variate Residual Signals

We address the problem of anomaly detection in industrial facilities based on multi-sensor measurement systems. Any anomaly may point to an (early) indication of a fault in the system, which may turn latter into a severe failure and thus should be detected as early as possible. Our approach relies on causal relation networks gained from the measurements of process variables, whose influence weights between causes and effects are modelled by a specific non-linear fuzzy systems architecture comprising generalized rules. Residuals are extracted on-line from a bunch of (high-quality) causal relations and are analyzed over time through an {\em advanced independent component analysis (ICA)}, which employs a specific strategy on dominant parts for automatically suppressing noise content in the data/residuals. A further important novelty aspect is that the mixing and demixing matrices which characterize the independent components are recursively updated in incremental and robust (converging) manner (thus, our approach is termed as {\em RICA}). The control signal extracted from the components is therefore also dynamically updated. The second dynamic aspect concerns the update of the causal relation models themselves through the usage of an evolving fuzzy systems approach. This integrates the possibility of structural changes by decreasing or increasing the models' non-linearity degree on demand and on the fly. RICA is compared with static ICA and with several (15) related state-of-the-art (SoA) methods based on data sets from a chip manufacturing system embedding two production stages. They comprise OK process phases with different machining parameters and NOT-OK (abnormal) phases. Results show improved detection capabilities with RICA while achieving lower false alarms rates than static ICA and than the SoA methods.