A Semi-Supervised Learning Approach for Fault Detection and Diagnosis in Complex Mechanical Systems

The integration of artificial intelligence in mechanical fault detection and diagnosis (FDD) helps to increase reliability, reduce costs, and improve the overall performance of mechanical systems in Industry 4.0 applications. Most interesting industrial applications nowadays come from dynamic environments where data are generated continuously over time and where the labeled data are scarce and expensive. Therefore, semi-supervised learning (SSL) can be particularly useful in FDD because faults may be rare or difficult to identify, and may not be fully represented in the labeled data. By using a combination of labeled and unlabeled data, SSL can help to identify these rare or difficult-to-detect faults, leading to more effective FDD. In this paper, graph-based SSL relying on label propagation is combined with conventional classification algorithms to detect potential failures in complex mechanical systems. Experimental results on realistic pneumatic and hydraulic systems from the related literature show that the proposed method can effectively enlarge the labeled datasets and interestingly identify different types of non-nominal conditions with higher accuracy compared to baseline methodologies.