Research On Online Fault Detection Based On Convolutional Auto-encoder Network

Mechanical equipment is widely used in all aspects of social production and is closely related to the national economy and people’s livelihood.The loss caused by the environment or software errors may trigger mechanical failures that are difficult to troubleshoot.The timely detection and troubleshooting of early failures can greatly reduce maintenance costs and risks.A more accurate estimation of available time can improve the economics of maintenance.With the rapid development of industrial technology and the rise of the fourth industrial revolution,distributed sensor deployment and intelligent integrated control systems have become a developing trend.In order to detect failures in mechanical operation,data-driven intelligent methods have been one of the research hotspots.However,mainstream supervised learning fault detection methods have the problems of difficulty in obtaining fault data and high labeling costs.For this reason,an unsupervised online fault diagnosis algorithm based on wavelet and fully convolutional variational autoencoding network are proposed as well as a supervised online remaining useful time estimation method.The main important work of this paper:(1)The principles of spectrum methods such as Fourier transform,discrete wavelet and continuous transform were explained.In this paper,it used several methods for spectrum analysis of vibration acceleration signals,and introduced support vector machines,explored the effectiveness of methods based on spectrum analysis and machine learning.(2)Bayesian viewpoints were combined to explain the working principles of autoencoder and variational auto-encoder.In this paper,it used convolutional layers instead of fully connected layers to implement full-convolutional variational auto-encoders,reducing model complexity while improving generalization ability.(3)A fully convolutional variational auto-encoder based on wavelet timefrequency spectrum was designed to compute online data loss values.In this paper,an online diagnosis method based on key model ensemble was proposed.The method uses Haar wavelet to transform time-scale spectrum as network input.The loss value of the supervised online training of the fully convolutional variational auto-encoding network was used as the criterion.The output difference of the reference sample was used as a measure to maintain a historical key model queue.According to the statistics of the current loss of the model in the queue,the abnormal score of the current signal was obtained.Validation on FEMTO and IMS data sets showed that this method solves to a certain extent the problems that a single model is difficult to detect slow-changing faults online and the continuous increase in resource consumption of the timing model ensemble.This method can better monitor and alarm the entire life cycle of equipment.(4)Convolutional layer and the gate recurrent unit were used to construct a convolutional timing prediction network.In this paper,it utilized the features provided by the proposed online detection method as the input of the time series prediction network,and online estimated the remaining useful time of the bearing.It verified the proposed method in the FEMTO data set,and compared with other methods,the results showed that the proposed online prediction method achieves a better level of the offline method,and has a better ability in online remaining useful time estimation.