面向高效通信聯(lián)邦學習的設備故障診斷方法

打開文本圖片集

Abstract：Federated learning achieves joint trairung modeling of fauh data from various factories while protecting privacy.However， due to the high heterogeneity of factory equipment operation data， traditional federated learning has low communicationefficiency. To address the above /ssues， an equ/pment fault diagnosis method based on high-efficiency communication federatedlearning was proposed. Firstly， a federated dynamic weighted balancing model was proposecl， which dynamically adjusts the trainingfrequency and uploaded parameter amount of the factory sub end， and improves communication efficiency by reducing communicationtime-. Secondly， a dual jump gate， cyclic unit diagnostic ruode，l with an attention mechanism was proposecl， which assigns differentweights to diffe，rent features to quickly extract fault features， effectively shortening communication rounds and improving communicationeffic，iency. Finally， experimental validation was conducted using the bearing failure datasets of Case Western Reserve University andPaderborn University. The experimental results show that compared with the federated average algorithm， this method achieves a faultdiagnosis accuracy of 92. 20% ， while reducing communication time by 56. 88%  and shortening communic，ation rounds by 47. 37% ，effectively improving communication efficiency.

Keywords：fault diagnosis; federated learning; data heterogeneity; corrununication efficiency; dynamic weighting

0  引  言

聯(lián)邦學習作為分布式機器學習新范式，在保護數(shù)據(jù)隱私的基礎上實現(xiàn)聯(lián)合多參與方共同建模，進而有效解決數(shù)據(jù)孤島問題。（剩余14005字）