Unsupervised Feature Learning Using Domain Knowledge Based Autoencoder

Abstract

Autoencoder is an unsupervised feature engineering technique, which is an emerging technique in the prognostics and health management (PHM) domain. However, since neural network based techniques such as auto-encoder have many hyper-parameters belonging to the number of hiddenunits, the number of layers and activation functions, substantial efforts are required in a heuristic way to make the autoencoder learn proper features. In this paper, we propose a novel method to regularize an auto-encoder by exploiting domain knowledge, such as mechanical engineeringexpertise. The proposed autoencoder learns robust features in a fast and efficient manner, therefore resulting in minimal consideration for the hyper-parameters. In the proposed method, some of the hidden units of the autoencoder are forcibly pre-trained by back-bone signals, such as 1Xsinusoidal wave of vibration, and the remaining hidden units efficiently learn the features of fault signals by minimizing the redundancy of learned features. The domain knowledge based regularization reduces the degree of freedom (DOF) of the autoencoder model as well as guide the model to learn the more physically reasonable features. Various fault data measured from a journal bearing rotor testbed are used for demonstration of the proposed method.