Bearing Fault Diagnosis Using Fully-Connected Winner-Take-All Autoencoder

Intelligent fault diagnosis of bearings has been a heated research topic in the prognosis and health management of rotary machinery systems, due to the increasing amount of available data collected by sensors. This has given rise to more and more business desire to apply data-driven methods for health monitoring of machines. In recent years, various deep learning algorithms have been adapted to this field, including multi-layer perceptrons, autoencoders, convolutional neural networks, and so on. Among these methods, autoencoder is of particular interest for us because of its simple structure and its ability to learn useful features from data in an unsupervised fashion. Previous studies have exploited the use of autoencoders, such as denoising autoencoder, sparsity aotoencoder, and so on, either with one layer or with several layers stacked together, and they have achieved success to certain extent. In this paper, a bearing fault diagnosis method based on fully-connected winner-take-all autoencoder is proposed. The model explicitly imposes lifetime sparsity on the encoded features by keeping only k% largest activations of each neuron across all samples in a mini-batch. A soft voting method is implemented to aggregate prediction results of signal segments sliced by a sliding window to increase accuracy and stability. A simulated data set is generated by adding white Gaussian noise to original signals to test the diagnosis performance under noisy environment. To evaluate the performance of the proposed method, we compare our methods with some state-of-the-art bearing fault diagnosis methods. The experiments result show that, with a simple two-layer network, the proposed method is not only capable of diagnosing with high precision under normal conditions, but also has better robustness to noise than some deeper and more complex models.