FeDMus: Federated Dynamic Example Mining for Unlabeled Sensor Data in Industrial Automation

Machine learning-based time series analysis has numerous applications in industrial automation, particularly in identifying sensor signals collected from distributed machines. However, large-scale repetitive operations generate massive amounts of redundant data, posing challenges to centralized learning, which requires collecting all data. Federated active learning addresses these challenges by selecting the most informative samples from multiple local data distributions for training the local model, aggregating and broadcasting model weights without transferring raw data, thus preserving data privacy and reducing labeling costs. We propose a dynamic federated active learning strategy, FeDMus, Federated Dynamic Example Mining for Unlabeled Sensor Data. By monitoring the training states of local and global models, FeDMus optimizes the timing of sample selection and dynamically allocates the query budget. FeDMus includes a dynamic selector and a dynamic filter: the proposed selector dynamically allocates the query budget based on training loss and weight dispersion, selecting informative data based on uncertainty and diversity across both models; the filter dynamically adjusts the training set by reducing simple samples to prevent the model from getting stuck in redundant data. On the typical industrial dataset CNC machine, FeDMus achieves an accuracy of 91.96% using only 10% of the data. We validated our method on four datasets, and the experimental results show that FeDMus outperforms state-of-the-art methods in tasks such as anomaly detection, fault diagnosis, action and trajectory recognition.