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

2025-07-22

AI Summary

To address poor generalization in fault detection under dynamic industrial operating conditions—caused by train-test distribution shift—this paper proposes TARD, a novel test-time domain adaptation method operating continuously during inference. TARD introduces a pioneering dual-branch architecture that decouples system parameters from sensor measurements, applying distinct adaptive strategies to each: dynamic batch normalization for the parameter branch, and unsupervised domain alignment coupled with temporal feature disentanglement for the sensor branch. The method enables lightweight, online model updating without requiring labeled target-domain data. Evaluated on two real-world multiphase flow systems, TARD significantly improves early fault detection accuracy and robustness compared to state-of-the-art domain adaptation approaches. It effectively mitigates distribution shifts arising from both data scarcity and evolving operational conditions, demonstrating strong practical applicability in industrial monitoring scenarios.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Systems and Infrastructure for Web, Mobile and WoT: Federated Web and WoT systems, including distributed, federated and edge-based data processing
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Fault detection is essential in complex industrial systems to prevent failures and optimize performance by distinguishing abnormal from normal operating conditions. With the growing availability of condition monitoring data, data-driven approaches have seen increased adoption in detecting system faults. However, these methods typically require large, diverse, and representative training datasets that capture the full range of operating scenarios, an assumption rarely met in practice, particularly in the early stages of deployment. Industrial systems often operate under highly variable and evolving conditions, making it difficult to collect comprehensive training data. This variability results in a distribution shift between training and testing data, as future operating conditions may diverge from previously observed ones. Such domain shifts hinder the generalization of traditional models, limiting their ability to transfer knowledge across time and system instances, ultimately leading to performance degradation in practical deployments. To address these challenges, we propose a novel method for continuous test-time domain adaptation, designed to support robust early-stage fault detection in the presence of domain shifts and limited representativeness of training data. Our proposed framework --Test-time domain Adaptation for Robust fault Detection (TARD) -- explicitly separates input features into system parameters and sensor measurements. It employs a dedicated domain adaptation module to adapt to each input type using different strategies, enabling more targeted and effective adaptation to evolving operating conditions. We validate our approach on two real-world case studies from multi-phase flow facilities, delivering substantial improvements over existing domain adaptation methods in both fault detection accuracy and model robustness under real-world variability.

Problem

• Detect faults under evolving industrial operating conditions
• Address distribution shifts between training and testing data
• Improve fault detection with limited training data representativeness

Innovation

• Continuous test-time domain adaptation method
• Separates input features into parameters and measurements
• Dedicated module for targeted domain adaptation