Study On Fault Detection Method Of Vibration Signals Based On LSTM Auto-encoders

Rotating machinery such as wind turbines and steam turbines are important industrial production equipment,and their stable operations are of great significance to avoid potential accidents,improve the economic efficiency of the enterprise.Fault detection and analysis based on the vibration signal during equipment operation is an important method for equipment maintenance management.Currently,fault detection of mechanical vibration signals is mostly based on intuitive monitoring data.In this paper,in the context of more sophisticated computer technologies such as deep learning,and the accumulation of vibration signal data,the study of vibration signal fault detection based on LSTM auto-encoders is carried out.Feature extraction is an important part of a data-driven approach to vibration signal fault detection.Mainstream signal feature extraction methods all have certain limitations,so this paper proposes a model based on the combination of deep auto-encoder and LSTM(Long Short-Term Memory)cells,which can extract features end-to-end and uses Gradient Boosting Decision Tree(GBDT)algorithm for classification detection of extracted features,thus establishes a framework for time series signal fault detection based on LSTM auto-encoder.In this paper,experiments are designed on the Case Western Reserve University bearing datasets and the detection of faults under different working conditions is achieved using the proposed framework.In the case of slow convergence of LSTM auto-encoders during training and poor performance of feature extractions,a distribution difference based on KL divergence was added to the original training error function to optimize the error between input and output.And the structure of Skip Connection was proposed on the auto-encoder model.Experiments comparing the original model proved that the proposed improved error function has better effect with the improved model with faster convergences.For the common problem of inadequate labeling data,this paper proposes a method based on transfer learning for transferring the knowledge learned from source domain adequate labeled datasets to inadequate labeled datasets.A discrimination method based on Dynamic Time Warping(DTW)distance is proposed to construct the source domain dataset;the parameters of the LSTM auto-encoders model trained on the source domain are migrated to the target domain model to achieve parameter based transfer.And two datasets are used as the source and target domain data for experimental simulation comparison.Experimentally,the proposed method achieves fine detection accuracy,and the effectiveness and feasibility of the proposed method for fault detection in small labeling data sets is demonstrated.