Domain Adaptive Sparse Transformer for Aeroengine Bevel Gear Fault Diagnosis

Recently, intelligent diagnosis of rotating machinery has received more attention due to the advancement of industry. Generally, training and test sets are presumed independent and identically distributed in most researches. Nevertheless, the change of working conditions in actual production will lead to the shift of distribution to deteriorate the model adaptability. At the same time, the commonly used deep convolutional neural network focuses on the local relevance from vibration signals while disregarding their global features in the realm of fault diagnosis. This paper proposes a domain adaptive sparse transformer network (DASTN) for the above problems, which employs Wasserstein distance to measure the distribution differences among data sets of multiple working conditions. The feature extractor used in this method is sparse transformer network (STN), which can capture global features via specific network structure with time embedding. Experiment results conducted on datasets of aeroengine bevel gears with multiple rotational speeds prove the superiority of our proposed method in unsupervised domain adaptive fault diagnosis.