Feature ensemble learning using stacked denoising autoencoders for induction motor fault diagnosis

Fault diagnosis is significant to induction motor which has been widely used as industrial power driving sources. By fault diagnosis, proper maintenance can be arranged to avoid accidents, ensure safety and reduce maintenance costs. However, variable operating conditions and background noise always reduce effectiveness of traditional fault diagnosis methods. Currently the most advanced machine learning technology, that is deep learning, not only has strong ability for hierarchical representations, but also can adaptively extract feature information from considerable measured data without necessary of much prior knowledge. Nevertheless, only the last layer is utilized for feature extraction in current deep learning models, due to its better performance of feature abstraction and compactness than that of the lower layers. Different feature layers hold different identification ability. In order to consider all features in different layers and achieve effective fault diagnosis, an alternative method of feature ensemble learning based on stacked denoising autoencoders (SDAE) is proposed in this paper. Our proposed method can be constructed in three steps: first, obtaining the samples of raw data under different health conditions. A deep neural network with multiple hidden layers is then established layer-by-layer by considering each pair of layers as a denoising autoencoder (DAE), with the strategies of greedy layer-wise pre-training and fine-tuning. Finally, softmax regression is created by the last hidden layer and integrate feature to classify different fault types. Feature ensemble is considered by mutually independent rules with different weights according to corresponding accuracy. Effectiveness of the proposed method is demonstrated though identifying health conditions of an induction motor.