Machine Health State Recognition Through Images Classification with Neural Network for Condition-Based Maintenance

Abstract

A robust machine health state recognition tool is a pillar for condition-based maintenance and Deep Learning approach finds its natural application in such a context. This paper investigates the recognition of machine failures by image classification through a convolutional neural network in a condition-based maintenance environment. The case study involves a refrigerator for large retail establishments. Experimental measures, while the machine is approaching failure, are difficult to be collected, especially in the quantity needed for training and testing the neural network. For this reason, a digital twin of the asset has been created to simulate the behavior of the machine and generate as many data as needed: physically-based models of the machine and failure modes have been included and the simulated behavior has been tuned by using experimental data. Finally, it has been employed to generate signals that, translated into images, test the suitability of the neural network.