A semi-supervised transferable LSTM with feature evaluation for fault diagnosis of rotating machinery

摘要

Aiming at the issue of impracticality or costliness of collecting enough labeled signals under all working conditions, the performance of a method usually suffers a significant loss when the model trained on one working condition is directly applied to another working condition. Thus, the entropy gain ratio combined with a semi-supervised transferable LSTM (EGR-STLSTM) network is established for fault diagnosis of rotating machinery under variable working conditions. First, the ERG is devoted to evaluating the multi-domain features extracted based on the characteristics of rotating machinery. Then, the optimal feature subset is fed into the STLSTM to obtain a pre-trained network. Finally, a semi-supervised transfer learning strategy, that is, with the aid of a small number of target labeled samples, is applied to the pre-trained model to achieve competitive performance in target tasks. The proposed EGR-STLSTM network fully uses the vibration characteristics of rotating machinery and retains the specialized knowledge derived from the source domain. The fault diagnosis method is verified with raw vibration signals from a bearing test rig and a gearbox test rig. The experimental results indicate that compared with the well-known methods, the proposed method significantly improves the diagnostic performance by reusing the pre-trained model and reduces the demand for massive labeled samples under variable working conditions.