Semi-supervised learning for industrial fault detection and diagnosis: A systemic review

Area of Science:

• Industrial Automation
• Machine Learning
• Data Science

Background:

• Fault Detection and Diagnosis (FDD) is critical for industries, with Machine Learning (ML) methods being leading contenders.
• Traditional ML approaches include supervised and unsupervised learning, but often struggle with scarce labeled data in industrial settings.

Purpose of the Study:

• To systematically review and organize the existing literature on Semi-Supervised Learning (SSL) for FDD applications.
• To identify the most and least utilized SSL algorithms within the FDD domain.
• To provide practical recommendations for implementing SSL in industrial FDD.

Main Methods:

• Literature review and systematic organization of SSL methods for FDD.
• Classification of SSL algorithms based on the taxonomy of van Engelen & Hoos.
• Analysis of algorithm usage in relation to fault detection tasks and dataset structures.

Main Results:

• SSL emerges as a promising solution for FDD, particularly where labeled data is scarce.
• Identified the prevalence of different SSL algorithms and their suitability for various FDD tasks.
• Highlighted common dataset structures encountered in FDD research.

Conclusions:

• SSL offers significant potential to improve the accuracy of ML-based FDD systems in industrial environments.
• Proposed best practices for the practical implementation of SSL for FDD to mitigate common industrial faults.
• This review provides a foundational understanding and roadmap for future SSL research in FDD.