Cascading Time-Frequency Transformer and Spatio-Temporal Graph Attention Network for Rotating Machinery Fault Diagnosis

Abstract

Rotating machinery fault diagnosis is of great importance to guarantee safe and optimal operations of industrial processes. Heavy noise and dynamic behaviors usually make accurate mechanical fault diagnosis impossible while using the standard methodologies, particularly when they disregard specific domain information related to time, frequency, or space. To address these challenges, we propose a novel model, called spatio-temporal-frequency graph attention network (STFGAT), which can integrate time domain, frequency domain, and spatial information. The model leverages the Transformer to encode time and frequency information, then refined complex patterns in the time and frequency domain through self-attention mechanism and frequency domain attention, and finally captures the hidden patterns behind the data through the collaboration of time and frequency information. The encoded information is subsequently fed into the spatio-temporal graph attention network (STGAT) to allow the model to take full use of the spatial relationships between different components of the mechanical system and the temporal relationships across various time lags. This process improvement can learn complex patterns and relationships within the data, thereby facilitating predictions regarding the system’s state. The experimental results show that STFGAT outperforms other standard diagnostic models in the case studies and can achieve better diagnostic accuracy.