Fault Diagnosis Method of Rotating Machinery Based on CEEMDAN-LSTM Model

Abstract

Abstract To overcome the shortcomings of traditional feature extraction methods that rely on prior knowledge and expert experience in the field and the problems of shallow machine learning in fault diagnosis, such as inability to extract deep features, low accuracy of model recognition and poor generalization ability, a fault diagnosis method based on complete ensemble empirical mode decomposition with adaptive noise and long short-term memory (CEEMDAN-LSTM) model is proposed. Firstly, the original signal is decomposed by complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN); secondly, the temporal features of selected component signals are extracted by long short-term memory (LSTM); thirdly the temporal features of each component are concatenated into feature vectors according to the order of high and low frequencies; finally, the feature vectors are input into the softmax layer to get the fault classification category. The results of gear box fault diagnosis test show that the model can effectively identify the fault location. Compared with empirical mode decomposition and support vector machine (EMD-SVM) model and LSTM model, CEEMDAN-LSTM model can obtain the highest fault identification accuracy.