Fusion of Low-level Features with Stacked Autoencoder for Condition based Monitoring of Machines

Abstract

Effective condition based monitoring has received much attention in the area of prognosis and health management of machines due to its benefit such as improvements in reliability and security, economical efficiency and decreases equipment damage. Feature extraction is one of the essential step in condition based monitoring which determines the performance of diagnosis model. This paper proposes a novel discriminative feature extraction technique based on fusion of low-level features or hand-crafted features with high-level features to detect every inserted faults as well as latent abnormalities in the machine. Low level features are extracted using signal processing techniques while high-level features are obtained from stacked autoencoder based deep neural network. Thus, both features represent raw data but generation methods for both are entirely different. Discriminating features build an effective classifier that categorizes test data into the respective classes. In the proposed methodology multi-class SVM has been used as classifier. The effectiveness of proposed methodology is validated on acoustic datasets collected from air compressor. These acoustic datasets have been recorded from most sensitive positions of air compressor under various health conditions.