Collaborative Intelligent Prediction Method for Remaining Useful Life of Hard Disks Based on Heterogeneous Federated Transfer

Deep learning-based methods for predicting the remaining useful life (RUL) of storage hard drives have become crucial for ensuring data center storage security. To meet the training demands, it is often necessary to obtain monitoring data from different clients. However, users are generally reluctant to disclose local private data, and the heterogeneity of data across clients poses challenges for collaborative training. Based on federated transfer learning (FedTL), this article realizes collaborative prediction while ensuring data privacy. This method completes model training without exposing users' private data. To address the difficulties in collaborative modeling caused by data heterogeneity, a domain separation-based heterogeneous federated transfer (DSHFT) scheme is introduced. This scheme extracts shared and private degradation features from different clients. The global prediction model is constructed using collaboratively shared features, while local private features are used for personalized fine-tuning. Finally, a global aggregation model is constructed based on the feature similarity between local and central clients, and it is applied to unknown target clients. Experiments involving collaborative prediction across multiple data centers demonstrate the effectiveness of this method. The shared and private features extracted through domain separation can significantly enhance the prediction performance of the global model.