A Convolutional Neural Network Approach to the Semi-Supervised Acoustic Monitoring of Industrial Facilities

摘要

Industrial manufacturing facilities require autonomous damage detection and monitoring to maintain functionality and reduce maintenance costs, particularly when human intervention is costly or dangerous. And as robotic manufacturing expands, the need for autonomous process monitoring has consequently expanded as well. In an industrial scenario, damage is typically incurred through machine wear from power transmission sources (e.g. bearing, gear, motor, belt, rail/track, and material wear), and performance degradations may occur nearly instantaneously, or slowly over time. The consequences of such damage can be localized to a single mechanical system, or it can cascade into catastrophic damage across a facility. In this work, a nondestructive method for identify and tracking processes and events within a manufacturing facility is presented, based on analysis of an autonomous acoustic monitoring system. The approach employs a deep convolutional neural network in combination with unsupervised similarity analysis to identify and track industrial processes based on their acoustic signatures within an image-like spectrogram. The approach is designed for flexibility and extensibility to a range of industrial scenarios, and requires only limited labeling of training data. The results of experimental testing indicate that the approach is capable of properly segmenting and tracking manufacturing processes manifested in acoustic signals across a range of spatial and frequency scales, and is capable of handling temporal distortions. Future work on fully unsupervised approaches are discussed as well.