Computational Framework for Machine Fault Diagnosis with Autoencoder Variants

Abstract

Fault diagnosis is an integral component of Condition Monitoring systems used in industries. Deep learning techniques have recently been proven to be very effective in handling heterogeneous real time monitoring data. This paper presents a computational framework for machine fault diagnosis based on deep learning approaches and autoencoder variants. The framework includes methods for extraction of time domain and frequency domain features from raw data, training of autoencoders, and finally, detecting the presence of fault(s) using classifiers. A total of four different variants of autoencoders have been tested and analyzed for fault diagnosis framework; additionally, the option of reducing the learning complexity in autoencoder with layer wise feature selection strategy with ITER ranking has also been applied. For validation purpose, four standard fault diagnosis datasets were used. To obtain inferences, statistical tests was performed across all tested feature extraction techniques, autoencoder variants, and classifiers. The analysis suggests that stacked denoising sparse encoder in conjunction with frequency domain features and support vector classifiers gives consistent classification accuracy thus builds effective fault diagnosis system.