A comparison study of centralized and decentralized federated learning approaches utilizing the transformer architecture for estimating remaining useful life

The current prognostics approaches for a network of assets are centralized and reliant on the availability of assets’ sensors, failures, and anomaly data. To address this, the data from similar assets are usually aggregated to make a richer dataset for prognosis. However, if similar assets are located at different enterprises, business owners may not be willing to share their raw data with each other. One solution is decentralized Federated Learning (FL), where local client data and training is preserved on-site rather than being shared with a central server. Since FL theoretically addresses the challenges faced by the traditional centralized learning approaches, its performance needs to be investigated and compared with the centralized methods. The current paper aims to compare the performance of a centralized model with two decentralized FL algorithms to predict the remaining useful life (RUL) of an asset. Two prediction models, a long short-term memory (LSTM) and the Transformer architecture were developed to predict RUL. The comparison has been conducted using NASA C-MAPSS dataset where the results indicated that FedProx performed better than FedAvg generally, and Transformer architecture performed better overall than LSTM across all datasets in the centralized and decentralized scenarios.