Clustering Federated Learning for Bearing Fault Diagnosis in Aerospace Applications with a Self-Attention Mechanism

Abstract

Bearings, as the key mechanical components of rotary machinery, are widely used in modern aerospace equipment, such as helicopters and aero-engines. Intelligent fault diagnosis, as the main function of prognostic health management systems, plays a critical role in maintaining equipment safety in aerospace applications. Recently, data-driven intelligent diagnosis approaches have achieved great success due to the availability of large-scale, high-quality, and complete labeled data. However, in a real application, labeled data is often scarce because it requires manual labeling, which is time-consuming and labor-intensive. Meanwhile, health monitoring data are usually scattered in different regions or equipment in the form of data islands. Traditional fault diagnosis techniques fail to gather enough data for model training due to data security, economic conflict, relative laws, and other reasons. Therefore, it is a challenge to effectively combine the data advantages of different equipment to develop an intelligent diagnosis model with better performance. To address this issue, a novel clustering federated learning (CFL) method with a self-attention mechanism is proposed for bearing fault diagnosis. Firstly, a deep neural network with a self-attention mechanism is developed in a convolutional pipe for feature extraction, which can capture local and global information from raw input. Then, the CFL is further constructed to gather the data from different equipment with similar data distribution in an unsupervised manner. Finally, the CFL-based diagnosis model can be well trained by fully utilizing the distributed data, while ensuring data privacy safety. Experiments are carried out with three different bearing datasets in aerospace applications. The effectiveness and the superiority of the proposed method have been validated compared with other popular fault diagnosis schemes.