Open Access Open Access  Restricted Access Subscription or Fee Access

Autoencoder Networks for Unsupervised Feature Extraction of Acoustic Monitoring Signals

JEFFREY BYNUM, PARASTOO KAMRANFAR, DAVID LATTANZI

Abstract


Manufacturing facilities require autonomous damage detection and monitoring to maintain functionality and reduce maintenance costs, particularly when human intervention is costly or dangerous. In an industrial scenario, damage is typically incurred through the wear or faulting of power transmission systems such as bearings, belts, or motors. The consequences of such damage can be localized, or it can cascade into catastrophic failure across a range of processes and systems, with outsized economic impact. Acoustic monitoring systems can help to prevent such problems through near real-time assessment of process quality and degradations, however robust data analytics play an integral role. In this work, a nondestructive method for identifying and tracking processes and events within a manufacturing facility is presented, based on a high-dimensional analysis of features extracted from acoustic signals after a preliminary segmentation by a convolutional neural network. Most approaches to feature extraction utilize a combination of features manually extracted from time and frequency domain signal representations, often in conjunction with principal component analysis. The process presented here considers the applicability of autoassociative neural networks, or autoencoders, for performing feature extracting in an unsupervised regime. Using experimental data collected from mechanical equipment, features extracted via a conventional approach are compared against features extracted using an autoencoder approach, both for the ability to distinguish between recorded mechanical actuations, and for their ability to track changes in these processes over time. The results of this evaluation indicate that an autoencoder is capable of results comparable to conventional methods. The major limitation of the approach is the need for larger data sets in order to ensure network convergence. Future work includes the incorporation of these autoencoders into a deep learning framework, as well as applications to nonlinear manifold learning.


DOI
10.12783/shm2019/32486

Full Text:

PDF