Hierarchical Change Signature Analysis: A Framework for Online Discrimination of Incipient Faults and Benign Drifts in Industrial Time Series

Keywords: fault detection, concept drift, industrial time series, anomaly detection, latent space analysis, unsupervised classification, online learning, change signature, human-in-the-loop, process monitoring

TL;DR: We propose a hierarchical framework that distinguishes incipient faults from benign drifts in industrial time series using multi-scale change signatures and an online normality baseline.

Abstract: Industrial fault detection systems struggle to differentiate between benign operational drifts (e.g., tool wear, recipe changes) and incipient faults, often adapting to faults as new “normal” states and causing catastrophic failures. This work introduces a hierarchical framework that decouples change detection from change characterization. Upon detecting a drift, the system generates a Multi-Scale Change Signature (MSCS) quantifying geometric and statistical transformations in the primary detector's latent space. An unsupervised Drift Characterization Module (DCM), trained on an Online Normality Baseline (ONB), classifies the signature as benign or a potential fault. Benign drifts are ignored, while potential faults are flagged for review; confirmed benign drifts are added to the ONB for future reference. The framework is model-agnostic, computationally efficient, and scalable via a tiered human-in-the-loop system. Experiments on the Tennessee Eastman Process dataset with injected faults and drifts demonstrate the potential to achieve high fault detection rates, reduced false alarms, and efficient adaptation to novel benign changes.

Submission Number: 112