基于原型联邦学习的工业故障诊断方法

A Prototype Federated Learning Based Approach to Industrial Fault Diagnosis

摘要: 故障诊断是保障工业系统安全运行的关键。当前，数据异构导致的非独立同分布问题使得传统的中心化数据处理方法在工业故障诊断领域面临着巨大的挑战。基于此本文提出一种基于原型联邦学习的时序数据处理方法AP-FED，该方法分为基础阶段和原型学习阶段。基础阶段进行数据增强、特征提取与全局参数的传递；原型学习阶段进行边缘设备端的原型提取与云端的全局原型聚合，确保在全局类原型的指导下，联邦学习的局部网络能够更有效地学习到特征表示。为验证模型有效性，使用真实工业数据集与多个联邦学习基线方法进行对比。经仿真实验，所提方法在FD、CWRU与CNC三个数据集中分别至少提升15.37%、21.30%与1.81%，证明该方法具有较高的精度和泛化能力。

Abstract: Fault diagnosis is the key to ensuring the safe operation of industrial systems. Currently, the non-independent homogeneous distribution problem caused by data heterogeneity makes the traditional centralized data processing method face great challenges in industrial fault diagnosis. This paper proposes a temporal data processing method AP-FED based on prototype federated learning, divided into a foundation phase and a prototype learning phase. The foundation phase carries out data enhancement, feature extraction, and global parameter transfer; the prototype learning phase carries out prototype extraction at the edge device end and global prototype aggregation in the cloud, ensuring that the local network of federated learning learns the feature representations more efficiently under the guidance of the global class prototypes. To validate the model’s effectiveness, real industrial datasets are used for comparison with multiple federated learning baseline methods. After simulation experiments, the proposed method improves at least 15.37%, 21.30%, and 1.81% in the three datasets of FD, CWRU, and CNC, respectively, which proves that the method has high accuracy and generalization ability.