Heterogeneous Transfer Learning Based on Stack Sparse Auto-Encoders for Fault Diagnosis

摘要

Fault diagnosis can reduce the risk of accidental failure and play a vital role in ensuring the reliability and safety of industrial systems. The traditional fault diagnosis algorithms mostly require enough training samples. However, in many cases it can be difficult and expensive in some scenarios. In this paper, the auxiliary domain data are used to train the learner and a novel heterogeneous transfer learning method is proposed for fault diagnosis. Data from source domain and target domain are represented by heterogeneous characteristics of different dimensions in heterogeneous transfer learning. We project the source domain and the target domain into the same feature space through two different auto-encoders. Then the similarity of distribution between source domain and target domain could be evaluated. The concept of distance to the center of the domain is introduced to evaluate the similarity of distribution between source domain and target domain. Firstly, it is introduced into the projection process using a small number of target domain labeled samples supervised training sparse auto-encoders (SAEs). Then, the second encoder is used to extract further features. Finally, the source domain data was used to train SVM, and use it to diagnose the target domain data. The experiment result shows that classifier trained by different auxiliary domain data have different performance for target data. The proposed approach performs better than the traditional machine learning approach when there is little labelled data in the target domain.