Construction of Health Indicators for Rotating Machinery Using Deep Transfer Learning With Multiscale Feature Representation

In many applications, it is not easy to generate enough whole-life data for training a deep neural network, which may reduce the performance of a health indicator (HI). To solve this problem, a new HI construction method based on deep transfer learning is proposed in this article. First, a new multiscale domain-adversarial neural network is proposed to extract representative features from the data collected under different working conditions. By introducing the maximum mean discrepancy regularizer and the Laplace regularizer, this model can enhance features' discriminant ability for incipient fault and exploit overall degradation information simultaneously. Then, a new HI is obtained via dimensionality reduction based on the extracted features. Moreover, a new HI assessment metric is proposed to effectively estimate the performance of the obtained HI. This new index considers not only the nonlinear correlation but also the geometric similarity of the degradation processes under different working conditions. Comparative experiments are conducted on the IEEE PHM Challenge 2012 bearing data set and the XJTU-SY bearing data set. The results show that the proposed method outperforms several typical transfer learning methods and nontransferable deep learning methods in terms of monotonicity and correlation.