Toward Efficient and Deployment-Oriented Data-Driven Fault Diagnosis in Induction Machines: A Deep Learning Approach

Data-driven fault diagnosis has gained attention due to its real-time application potential. Leveraging near-real-time data availability, efficient algorithms for fault diagnosis are developed. Current approaches often fall short, being either too shallow or requiring costly feature engineering. Deep neural networks offer end-to-end frameworks with built-in feature extraction, potentially eliminating the need for separate feature engineering and achieving optimal features. This study considers rotor asymmetry fault diagnosis as a sequential modeling problem and employs a multi-view approach using two deep neural networks specialized in sequential modeling. One network is a CNN-based model, while the other is an RNN-based model. Together, they efficiently model the data from two different perspectives, resulting in an effective ensemble of features. Two motors are used to evaluate the proposed method under healthy and faulty conditions, across various fault severities and load levels. Experimental results demonstrate greater efficiency and a significant improvement of over 5% compared to recent deep learning methods in classifying stator current sensor data based on fault severities and load levels.