Multi-adapter based Federated Learning Framework for Remaining Useful Life Estimation Considering Heterogenous Local Edge Systems

초록

While various Federated Learning (FL) models have been proposed for predictive maintenance considering multiple distributed production environments, the different data formats in each local environment and the heterogeneous architecture of edge systems have been regarded as major issues hindering successful FL. In order to overcome these issues, a new and effective multi-adapter based global model considering local edge systems with Non-Independent and Identically Distributed (Non-IID) data is proposed. The proposed FL model allows local machines with heterogenous structures to reflect their unique characteristics while sharing the common features of the target product family to predict the remaining useful life. To achieve this, a multi-stream deep learning architecture based on multi-adapters is proposed for handling the multiple features of various local machines. To demonstrate the effectiveness of the proposed framework, the remaining useful life of the same type of aircraft engines manufactured in different production environments is predicted. Compared to existing FL methodologies, the proposed framework is experimentally demonstrated to achieve high prediction accuracy for the remaining useful life by handling non-IID data generated from local edge systems with different structures.

키워드