Machinery health monitoring system

The present paper introduces the concept of Machinery health monitoring system AI: predictive maintenance software- defined digital twin hardware-independent framework on industrial machinery. Specifically, the machine model consists of a discrete-time Markov Chain with three machine operating states: healthy, degrading, failure that allows realistic generation of synthetic data without the use of physical sensors. Multivariate time-based series data, including temperature, vibration and load, is also enhanced by stochastic perturbation in order to simulate the real-life operating environment. A sequence-to-sequence LSTM autoencoder using a temporal attention mechanism on which an unsupervised anomaly detecting model is trained is used to learn normal operational patterns. The model calculates reconstruction error as score of an anomaly which can be used to detect deviation to healthy behavior reliably. An automatic thresholding method allows the machine states to be classified in real-time. This system can run at a 1 Hz streaming rate and sub-milliseconds inference latency on a GPU system. The architecture proposed will consist of FastAPI-based back-end as a server, WebSocket layer as a means of communication, and a React- based front-end as a visualization and monitoring tool. This study shows that it can be scaled and repeated to predictive maintenance with synthetic data, and this provides the basis to make the transition to actual industrial applications.