CNN-based Autoencoder and Post-Training Quantization for On-Device Anomaly Detection of Cartesian Coordinate Robots

Robots are usually conducting difficult and laborious work on behalf of humans in various industrial fields. Specially, a Cartesian coordinate robot operates in a Cartesian coordinate system and is used in many industrial sites with little error and very high precision in work. However, due to the low-price requirements of Cartesian coordinate robots, there is a tendency not to apply fault diagnosis technology in industrial sites using Cartesian coordinate robots. In the case of a Cartesian coordinate robot having a precise and low error, it is more sensitive to failure than other robot systems because a slight failure may cause malfunction or product defects. Therefore, a low-cost technology is needed to diagnose the failure of the Cartesian coordinate robot. This paper deals with deep learning technology based on convolutional neural network (CNN) that recognizes the anomalous sound of sliding rails of Cartesian coordinate robots, to diagnose failures, and then post-training quantization to apply them on small and cheap embedded devices. This paper also describes how to manipulate the raw sound data for using as the input of deep learning. In order to apply the deep learning technology to small systems, I also evaluate several post-training quantization methods after converting the model into a lightweight model.