Graph-Based Semi-Supervised Learning for Induction Motors Single- and Multi-Fault Diagnosis Using Stator Current Signal

Supervised learning has been commonly used for induction motor fault diagnosis, and requires large amount of labeled samples. However, labeling recorded data is expensive and challenging, while unlabeled samples are available abundantly and contain significant information about motor conditions. In this paper, a graph-based semi-supervised learning (GSSL) approach using both labeled and unlabeled data is proposed. Experimental data for two 0.25 HP induction motors under healthy and faulty conditions are used. Discrete Wavelet Transform (DWT) is employed to extract features from recorded stator current signals. Three GSSL algorithms (local and global consistency (LGC), Gaussian field and harmonic function (GFHF), and greedy-gradient max cut (GGMC)) are evaluated in this study, and GGMC shows superior performance over other two. They are also compared with a supervised learning algorithm, support vector machine (SVM). As induction motors often operate under variable loadings, curve fitting equations are developed based on experimental data to generate training data for untested motor loadings.