关键词: 故障诊断, 分布式, Gap度量, 联邦学习, 航空发动机

Abstract: The safe operation of aviation equipment is essential for flight safety, and aero-engine fault diagnosis is of great importance because the engine is the core power unit of aircraft. In practical engineering scenarios, monitoring data of aero-engines are usually held by different institutions or test platforms. Due to confidentiality and privacy constraints, such data are difficult to collect centrally, resulting in a typical data-island problem. Federated learning provides a feasible way to train a global diagnostic model through local model updating and server-side parameter aggregation without sharing raw data. However, multi-source aero-engine data often follow non-independent and non-identically distributed patterns, which may reduce the performance and stability of conventional federated models. To address this problem, this paper proposes a Gap-metric-based federated aggregation method for distributed aero-engine fault diagnosis. The proposed method evaluates client contributions from the geometric structure of the feature space by considering intra-class compactness and inter-class separability, and then adaptively adjusts aggregation weights. Experiments on an aero-engine dataset, compared with Direct CNN, FedAvg, and FedProx, demonstrate that the proposed method achieves better diagnostic accuracy and stable training performance in distributed non-IID scenarios.

Key words: fault diagnosis, distributed, Gap metric, federated learning, aero-engine