Disrupting Downtime: Different Deep Learning Journeys into Predictive Maintenance Anomaly Detection

This article discusses the innovative application of artificial intelligence (AI) to develop a predictive model that aims to evaluate the condition of the machine by focusing on the probability of failure. The research uses a synthetic dataset prepared to simulate real-world situations where machines are equipped with sensors that monitor various health indicators and record the occurrence of faults. This data set consists of 10,000 inputs, each containing five numerical measurements: air temperature, process temperature, rotation speed, torque, and machine wear, in addition to the type of product produced, for a total of six input variables. The output of the model is the fault state of the machine, displayed as true or false. A hybrid artificial neural network integrating a GRU (Gated Recurrent Unit)-based model with the Transformer Encoder block was used for prediction. This combination highlights the superior predictive capabilities of the model. This approach represents a shift from traditional maintenance programs, which are often time-based and often result in unnecessary resource use, to a more efficient, condition-based maintenance strategy. This new strategy aims to ensure that maintenance activities are carried out only when necessary, thus optimizing resource use and minimizing downtime.