Generating feature sets for fault diagnosis using denoising stacked auto-encoder

Recent advancements in sensor technologies and data driven model based techniques have made intelligent diagnostic systems prominent in machine maintenance frameworks of industries. The performance of such systems immensely relies upon the quality of features extracted and the classifier model learned. Traditionally features were handcrafted, where engineers would manually design them with statistical parameters and signal transforms based energy distribution analysis. Recently, deep learning techniques have shown new ways of obtaining useful feature representation that provide state of the art results in image and speech processing applications. This paper first presents a brief survey of traditional handcrafted features and later presents a short analysis of handcrafted features v/s features learned by deep neural networks (DNN), for doing fault diagnosis. The DNN based features in this paper were generated in 3 phases: 1) extracted handcrafted features using traditional techniques 2) initialized the weights of DNN by learning de-noising sparse auto-encoders with the handcrafted features in unsupervised fashion and 3) applied two generic fine tuning heuristics that tailor DNN's weights to give good classification performance. The experimentation and analysis were performed on 5 datasets: one each on Air compressor monitoring, Drill bit monitoring and Steel plate monitoring, and two on bearing fault monitoring data. The results clearly show the prospects of DNN obtaining good feature representations and good classification performance. Further, it also finds that Fast Fourier Transform based features with DNN are more suited for Support Vector Machine as classifier than Random Forest.