Noncontact machinery operation status monitoring system with gated recurrent unit model.

In manufacturing industry, assembly line monitoring provides statistical information about overall perfor-mance and reliability of the legacy machines, ensuring that the machines give maximum yield output. However, most legacy machines lack internet connectivity and advanced functionality, increasing the difficulty for tracking task. There-fore, this work seeks to introduce a noncontact acoustic method to track machines rather than the mainstream vibrational approach. In order to provide accurate tracking of the daily machine operation for our machine tracking system, we consider scenario of background noises such as environmental sounds from multiple sources as well as neighbouring machine's sound. Thus, several neural networks are employed to recognize the machine status accurately. The objective of our work is to investigate the effect of machine types and states on recognition performance of neural network models under extremely noisy environments as well as to demonstrate the possibility of recognizing the sound on edge device. The main contribution of this article is the proposal of lightweight recurrent and convolutional-based models for ma-chine sound recognition. The experimental results of our extensive testing included with multiple types of machines and background noises show that the proposed system with gated recurrent unit model has the best recognition accuracy of F1 score 0.913 with standard uncertainty of 0.026 with decent inference speed on edge device.