Federated Anomaly Detection for Fault Diagnosis Using Autoencoders

This paper introduces a federated learning approach for anomaly detection in machine fault diagnosis, utilizing Autoencoders (AEs) and Variational Autoencoders (VAEs) on multimodal sensor data. Employing privacy-preserving distributed training, we compare Federated Averaging (FedAvg) and Federated Proximal (FedProx) for model aggregation. Results demonstrate the AE’s superior performance over the VAE. Notably, FedProx achieved an AUPR of 0.9454 and an F1 score of 0.9236, highlighting its effectiveness in handling data heterogeneity with imbalanced datasets. This study validates the potential of AE with FedProx for robust, real-world anomaly detection in predictive maintenance.