Continuous Test-time Domain Adaptation for Efficient Fault Detection under Evolving Operating Conditions

TL;DR

This paper tackles fault detection under evolving industrial operating conditions by proposing TAAD, a continuous test-time domain adaptation framework that combines a reconstruction-based anomaly detector with two domain-adaptive modules dedicated to control variables and sensor measurements. By freezing the reconstruction model and only adapting the control-variable module via test-time learning (with AdaBN for stability), TAAD achieves robust early fault detection while mitigating false alarms under continuous domain shifts. The approach is validated on a real-world industrial pump dataset, demonstrating superior detection timeliness and lower false-positive rates compared with AdaBN, MMD, and non-adaptive baselines, including across station transfers. Practically, TAAD enables fleet-level knowledge transfer with minimal labeled fault data, offering a scalable solution for proactive maintenance in dynamic industrial environments.

Abstract

Fault detection is crucial in industrial systems to prevent failures and optimize performance by distinguishing abnormal from normal operating conditions. Data-driven methods have been gaining popularity for fault detection tasks as the amount of condition monitoring data from complex industrial systems increases. Despite these advances, early fault detection remains a challenge under real-world scenarios. The high variability of operating conditions and environments makes it difficult to collect comprehensive training datasets that can represent all possible operating conditions, especially in the early stages of system operation. Furthermore, these variations often evolve over time, potentially leading to entirely new data distributions in the future that were previously unseen. These challenges prevent direct knowledge transfer across different units and over time, leading to the distribution gap between training and testing data and inducing performance degradation of those methods in real-world scenarios. To overcome this, our work introduces a novel approach for continuous test-time domain adaptation. This enables early-stage robust anomaly detection by addressing domain shifts and limited data representativeness issues. We propose a Test-time domain Adaptation Anomaly Detection (TAAD) framework that separates input variables into system parameters and measurements, employing two domain adaptation modules to independently adapt to each input category. This method allows for effective adaptation to evolving operating conditions and is particularly beneficial in systems with scarce data. Our approach, tested on a real-world pump monitoring dataset, shows significant improvements over existing domain adaptation methods in fault detection, demonstrating enhanced accuracy and reliability.

Figures (3)

• Figure 1: General pipeline of reconstruction-based unsupervised anomaly detection
• Figure 2: Test-time domain Adaptation Anomaly Detection (TAAD). a) We first pretrain the reconstruction-based anomaly detection model on the source dataset; b) For domain adaptative prediction, we add domain adaptive module and train it on the target training data.
• Figure 3: Comparison of performance of (a) baseline (b) AdaBN (c) MMD (d) TAAD for fault detection on Pump B-D. The number of predicted faults per day is plotted. The yellow lines mark the starting date of the reported faults. Red regions mark the 14-day time window before the occurrence of reported faults, and the red dots within this region mark the first detected abnormal day within this period.