Rotating Machinery Fault Diagnosis Under Multiple Working Conditions via a Time-Series Transformer Enhanced by Convolutional Neural Network

Fault diagnosis plays a vital role in the prognostics and health management (PHM) of industrial equipment and systems, but accurate diagnosis is always challenging in practical applications with varying working conditions. In this paper, we propose a new novel deep learning model named C-Trans, tailored for fault diagnosis of multiple complex working conditions in rotating machinery, aiming to addresse the demand for robust generalized algorithms in real-world applications. Specifically, the proposed method, which integrates the advantages of both CNN and Transformer structures, employs a deep multi-scale CNN for extracting multi-scale features from the measured raw vibration signals. The model enhances its capacity through convolutional operations to learn and generalize across a range of working conditions. Subsequently, a Transformer block structure with multi-head attention mechanism is further introduced, facilitating the establishment of connections between the extracted fault information and corresponding fault categories. Extensive experimental results obtained from five distinct cases and comparative analyses are used to evaluate the effectiveness of our method. Through accuracy comparison and visualization analysis across diverse cases, our findings reveal the average diagnosis accuracy rate of up to 97.57%, showcasing robust capability in discerning a spectrum of fault characteristics.