An Analysis of Semi Supervised Machine Learning in Electrical Machines

B. Prashant,2Metukumelli Durgarani,3Sarika Harika,4Bommineni Gowtham Kishore,5Dadi Sai Konda Naga Venkata Durga Mahesh

doi:10.62643/

Authors

B. Prashant,2Metukumelli Durgarani,3Sarika Harika,4Bommineni Gowtham Kishore,5Dadi Sai Konda Naga Venkata Durga Mahesh Author

DOI:

https://doi.org/10.62643/

Keywords:

Semi-supervised learning, electrical machines, condition monitoring, fault diagnosis, predictive maintenance, feature extraction, data-driven modeling, machine learning, industrial automation, signal processing

Abstract

The accelerating growth of industrial automation and the rising complexity of electrical
machines have intensified the demand for advanced systems capable of intelligent
monitoring, fault detection, and predictive maintenance. Conventional supervised machine
learning approaches have been extensively used to interpret operational data, supporting
accurate fault identification and performance enhancement. However, these techniques rely
heavily on large volumes of labeled data, which are often expensive, labor-intensive, and
difficult to obtain in practical industrial environments. Meanwhile, a substantial portion of
machine-generated data remains unlabeled and underexploited, despite its potential value.
Semi-Supervised Machine Learning (SSML) offers a practical solution by integrating both
labeled and unlabeled data to strengthen model performance, adaptability, and generalization.
This study examines the application of SSML techniques in the monitoring and maintenance
of electrical machines with the objective of improving system reliability, minimizing
unexpected failures, and optimizing maintenance planning. Several prominent SSML
strategies are analyzed, including self-training, co-training, generative modeling, and graphbased
approaches. Self-training iteratively enhances model accuracy by incorporating highconfidence
predictions from unlabeled samples. Co training utilizes multiple independent
feature perspectives to improve learning from partially labeled datasets. Generative
techniques, such as variational autoencoders and generative adversarial networks, create
realistic operational patterns to augment limited fault data. Graph-based methods leverage
structural relationships among sensor readings to propagate labels efficiently and capture
intricate operational interdependencies. The research explores the practical implementation of
these SSML methods for condition monitoring, early fault detection, and remaining useful life
estimation of electrical machines.