Federated Learning for Predictive Maintenance: Model Comparisons and Privacy Advantages

Predictive maintenance has evolved into an essential component of modern industrial systems, enabling early detection of equipment defects and reducing downtime. Traditional methodologies have frequently relied on centralized machine learning models to collect large amounts of sensitive data, presenting significant privacy and security concerns in the Big Data era. This work presents a federated learning architecture developed exclusively for predictive maintenance, allowing multiple clients to train machine learning models simultaneously while keeping their data decentralized. Several machine learning algorithms were used, including Logistic Regression, Random Forest, Ridge Classifier, AdaBoost, Decision Tree, and Gradient Boosting, as well as advanced techniques such as the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalances, Principal Component Analysis (PCA) to reduce dimensionality, and Particle Swarm Optimization (PSO) to tune hyperparameters. Empirical evaluations on the Gearbox Fault Diagnosis and Machine Predictive Maintenance Classification datasets demonstrate that the proposed federated learning framework achieves performance comparable to traditional centralized approaches while preserving data privacy. In the Machine Predictive Maintenance Classification dataset, Random Forest achieved the highest performance, attaining an impressive accuracy of 93.17% in the federated learning approach. For the Gearbox Fault Diagnosis dataset, the accuracy of all models was remarkably similar in both the traditional and federated learning approaches, which is a noteworthy outcome. These results indicate that federated learning can effectively balance the demands of privacy and performance in predictive maintenance applications, making it a viable alternative for industries that prioritize data confidentiality.