Incipient Stator Winding Fault Detection and Severity Estimation in Induction Motors With Unsupervised Machine Learning Algorithms

Induction motors are considered as the power-houses of the industry. These motors amount to approximately $60-70 \%$ of global industrial energy consumption which drives multiple fields such as medical, oil and gas, marine, and transportation. Extensive research has been conducted to enhance the reliability of induction machines by implementing fault diagnosis techniques at early fault inception stages. The reason for its challenge is due to the incipient nature of the fault, which may degrade over time unnoticed due to its minimal effect. Machine learning has improved drastically in the past few years and is known for its large community and easy accessibility. This paper aims to apply an unsupervised machine learning algorithm to detect the severity level of stator winding faults that occur in the induction motor, which allows for predictive maintenance, enhancing the lifetime of the motor. The K-means, Gaussian mixture model and hierarchical clustering, all of which use the concept of clustering, referring to the grouping of similar data that have likable characteristics, are used in this paper. The metrics used to evaluate the aforementioned models show that the clusters are well-separated and distinct from each other but do not correlate well with ground truth. K-means achieved a score of 0.85 for Davis Bouldin, 499908.73 for Calinski Harabasz while the adjusted rand index was 0.00014 and normalized mutual information score was 0.0002. Similarly, GMM achieved $1.63,342100.39,0.005$ and 0.006 for Davis Bouldin, Calinski Harabasz, adjusted rand index, and normalized mutual information respectively. The hierarchal clustering was able to classify six classes at a threshold distance betweem 1600-2000.