Semi-supervised learning and condition fusion for fault diagnosis

• Manifold regularization based semi-supervised learning is introduced to fault diagnosis.
• Unlabeled condition data are also utilized to enhance the multi-fault detection.
• A new single-conditions and all-conditions labeled mode is proposed to feed SSL.
• This SSL approach outperforms supervised learning in both labeled modes.
• The manifold fundamental of single-conditions labeled mode is analyzed with dimensionality reduction.

Supervised learning has been developed to collect condition monitoring (CM) data for fault diagnosis and prognosis. However, labeling the condition monitoring data is expensive due to the use of field knowledge while unlabeled CM data contain significant information of normal conditions or faults, which cannot be explored by supervised learning. Manifold regularization (MR) based semi-supervised learning (SSL) is first introduced to fault detection by utilizing both labeled and unlabeled CM data, and then a new single-conditions labeled mode based on MR is proposed for SSL learning. This approach, leveraged by effectively exploiting the embedded intrinsic geometric manifolds, outperforms supervised learning in both single-conditions labeled and all-conditions labeled modes within the application of two real-life fault detection datasets. The experimental results also suggest that most effective classifier in practical application could be trained by the SSL approach and fault type representation with medium load condition. The improved predictive performance implies that the manifold assumption of MR has its inherent fundamentals. Finally, the manifold fundamental of single-conditions labeled mode is analyzed with dimensionality reduction.