Deep Root Cause Analysis: Unveiling Anomalies and Enhancing Fault Detection in Industrial Time Series

Hao Huang, Tapan Shah, John Karigiannis, Scott Evans

Abstract: In industrial systems, ensuring operational reliability hinges on effective fault detection and root cause analysis. Traditional data-driven fault detection, employing AI to construct prediction models, often falters in accurately identifying true root causes. This limitation arises when variables with high prediction residuals are merely symptomatic rather than direct causes. We present Deep Root Cause Analysis (DRA), a hierarchical neural network. The first model predicts time series and identifies anomalies with high residuals, while the second regresses these residuals from the first model, generating a saliency map that highlights potential root causes that contribute to the residuals. DRA exhibits promising capabilities in providing detailed insights, overcoming the shortcomings of traditional deviation-based approaches. Experiments conducted on the Tennessee Eastman process dataset showcase DRA’s superiority in detecting various process faults and pinpointing their root causes, thereby enhancing precision in complex industrial settings.