Federated Generative Models for Predictive Maintenance in Industrial Environments

Predictive Maintenance (PdM) is a proactive approach to maintenance based on predicting the future evolution of the health state of industrial components and their failure time. It allows for minimizing downtime by planning maintenance interventions in advance and reducing the costs associated with unexpected breakdowns. PdM strongly relies on Machine Learning (ML) and Data Analytics (DA). However, implementing predictive maintenance across multiple, geographically dispersed industrial plants presents challenges, particularly in terms of data privacy, ownership, intellectual property, and regulatory compliance. A possible solution to ensure that data remains on the original source is Federated Learning (FL). In this work, a generative model based on a Variational Autoencoder (VAE) is employed for anomaly detection in time series data. The model is trained via FL is shown to improve the VAE generalization capabilities in an industrial setting. Moreover, we highlight that the partial federation of the model, which involves the transmission of a subset of the model parameters, allows for increasing the overall performance while reducing the total parameter communication overhead. This result is particularly important for applications aiming to maximize communication efficiency by reducing the total transmission latency between the end nodes.