EXPLAINABILITY BASED FEDERATED LEARNING FOR THE PREDICTIVE MAINTENANCE IN SMART MANUFACTURING

Abstract

In today’s manufacturing environment, machines are expected to run continuously, which makes early fault detection and maintenance planning very important. At the same time, industries need to ensure that sensitive operational data is not shared across different plants or systems. To handle both these requirements, this paper presents Fed-Fusion-XAI, a framework that combines predictive maintenance with privacy-preserving learning. The system uses a hybrid deep learning model that brings together CNN and BiLSTM techniques to process sensor data and identify unusual patterns in machine behavior. It not only detects possible faults but also estimates the Remaining Useful Life (RUL) of equipment such as turbofan engines and CNC machines. To make the predictions easier to understand, an explainability method based on SHAP is applied, which helps in showing how different factors influence the model’s decisions. Instead of collecting all data in one place, the framework works in a distributed setup where individual nodes train their own models using local data. Only the learned parameters are shared with a central server, which combines them using a federated learning approach. This helps in maintaining data privacy while still benefiting from collective learning. The system also includes a real-time data pipeline for handling continuous data and a simple dashboard for monitoring results. The proposed approach provides a balanced solution that focuses on accurate prediction, data privacy, and better understanding of model outputs, making it suitable for real-world industrial applications.