Convolutional Neural Networks for Electrical Endurance Prediction of Alternating Current Contactors

• Published in: IEEE transactions on components, packaging, and manufacturing technology (2011) Vol. 9; no. 9; pp. 1785 - 1793
• Main Authors: Cui, Hechen, Wu, Ziran, Wu, Guichu, Xu, Xiaofeng, You, Yingmin, Fang, Yandong
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.09.2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: AC contactors; Accuracy; Alternating current; Artificial neural networks; Classification; Contactors; Contacts; Convolution; Convolutional neural networks; convolutional neural networks (CNN); Data acquisition; Electric contacts; electrical endurance prediction; Fatigue tests; Loss measurement; Machine learning; Neural networks; Regression models; Signal processing; Switching circuits; Waveforms
• ISSN: 2156-3950; 2156-3985
• DOI: 10.1109/TCPMT.2019.2930741

This article proposes a deep-learning-based approach that predicts electrical endurance of alternating current (ac) contactors by using measured electric waveforms. We execute AC-4 tests and acquire voltage and current waveforms during breaking processes. For every waveform, we extract signals within a period in which the breaking arc occurs. Extracted three-phase signals of every breaking operation are reformed to a 32 × 32 × 6 example. We also measure the accumulated mass loss of the contacts to represent the electrical endurance condition and label the examples according to the mass loss proportion. Convolutional neural networks (CNNs) are employed to build prediction models. Both a LeNet5-like CNN structure (CNN-A) and an improved CNN structure (CNN-B) with 3-D convolution layers are implemented to train classification and regression models. Experimental results show that CNN-B for classification achieves the best prediction accuracy. Then, a voting mechanism is utilized for accuracy improvement and achieves the prediction accuracy of 82.14% for ten classes and 90.98% for five classes. Especially, the accuracy is over 90% for the last classes, which represent the ending condition of electrical endurance. Therefore, it is proven that the method can be utilized to predict electrical endurance for the AC-4 condition.