Research Area: Machine Learning
Induction machines have extensive demand in industries as they are used for large-scale production and, therefore, vulnerable to both electrical and mechanical faults. Automated continuous condition monitoring of industrial machines to identify these faults has become one of the key areas in research for the past decade. Among various faults, early-stage identification of insulation failure in stator winding is of significant demand as it is often occurring and accounts for 37% of the overall machine failures. Also, this fault, if identified at its incipient stage, can predominantly improvise machine downtime and maintenance cost. In the proposed work, stator current signal data in the time domain from the experimental setup of both healthy and faulty induction machines are used to train the artificial neural-network models in order to identify the machines condition. Reducing the time required to train the neural network, features are extracted from the raw current signal data and then fed to the classifiers. Various performance characteristics of eleven neural-network models such as the number of features, number of epoch runs, training time, activation functions, learning rate, model loss function, and accuracy concerning each model are quantified. Only a few neural networks could classify a healthy and a faulty induction machine with 94.73% efficiency on generalization the neural-network model with the raw data, whereas 98.43% efficiency with the statistical featured data.
Keywords:
Author(s) Name: Amar Kumar Verma, Shivika Nagpal, Aditya Desai & Radhika Sudha
Journal name: Neural Computing and Applications
Conferrence name:
Publisher name: Springer
DOI: 10.1007/s00521-020-05033-z
Volume Information: volume 33, pages 1297–1310 (2021)
