An intelligent unsupervised fault diagnosis method based on subspace distribution alignment

Recently, marvelous success has been obtained for machine learning approaches in solving the mechanical defect detection problems. However, the traditional machine learning algorithms usually fail to generalize to new input distributions, which may cause the classification accuracy drops dramatically. To overcome this deficiency, domain adaptation techniques can be employed for transferring and adapting source domain to target domain. In this paper, an effective unsupervised domain adaptation approach called subspace distribution alignment (SDA) is applied to the defect diagnosis of gearbox. Firstly, the Fast Fourier Transform (FFT) is adopted for preprocessing the raw vibration signals into the frequency domain. Then, SDA algorithm is employed for mapping the source and target domain into the same subspace. In addition, the z-score normalization is applied to the source and target subspace. At last, the different fault types of gearbox are classified by the softmax regression classifier. The detection results show that our approach outperforms the other intelligent fault detection approaches listed in this paper.