Unsupervised Anomalous Sound Detection Using Loss-Weighted Clustered Federated Pre-Training

Sound anomaly detection in industrial applications has to cope with limited and diverse training data as well as domain shifts. Because of privacy and security considerations, the training data may not be collected on a central server. It is nevertheless of interest to leverage all data in the training process of individual clients for improving robustness. In this work, we therefore explore the use of federated and clustered federated learning techniques for improving the anomaly detection performance of individual clients. An enhanced aggregation weighting that takes the loss of individual clients into account yields a reduction of false positives in conjunction with a standard autoencoder. Our results show that also in diverse client data, applying federated learning for the first iterations of the model learning process enhances the overall anomaly detection capability for most clients.