TinyML Anomaly Detection and Fault Prediction for Industrial Applications

Tiny Machine Learning (TinyML) for industrial applications, in the context of this project, involves the deployment of machine learning (ML) models on the edge for structural health monitoring, to allow low- latency decision-making and proactive preventive maintenance of industrial equipment. Conventional maintenance practices are largely based on routine inspections, such as work-by-inspection (WBI) and scheduled maintenance, which are labor intensive, costly, and susceptible to diagnostic errors. This proactive approach bridges these gaps by detecting and diagnosing potential failures in structures such as bearings, bridges, pipelines, and manufacturing equipment before they result in a catastrophic failure. Recent developments in condition monitoring employ cloud-based ML algorithms; however, these ap- proaches are constrained by high latency, high operational cost, significant energy consumption, and network limitations, particularly in remote industrial environments. In this project, using cement mill bearings as a case study, TinyML was investigated as an alternative solution. A lightweight Temporal Convolutional Network (TCN) was compared to other light weight models and was implemented to predict the bearing temperature of the slide shoe 45 minutes in advance. With a mean absolute error of 0.02048, the approach demonstrated early fault prediction and automated preventive actions through the water sprayer and alarm activation.