A Fault Diagnosis Method Based On Deep Belief Networks

Fault diagnosis,as an important part of the condition based maintenance,can extract fault-related information and establish nonlinear relationship between extracted features and system conditions by monitoring,processing and learning the sensor data during the operation of the equipment.Fault diagnosis can distinguish the type,position and degree of fault by achieving three objectives,which are fault detection,fault isolation,and fault identification,and thus detect early fault in time to make reasonable maintenance decisions,improve system reliability and reduce maintenance costs.With the development of sensor technology and condition monitoring technology,a large amount of real-time condition monitoring data with rich feature information is stored.The deep learning,which has good ability of big data processing and feature learning,can be used in the problem of fault diagnosis with lots of monitoring data.This paper will research on fault diagnosis based on deep belief network,which a deep learning method that use condition monitoring data of mechanical components to train the model.This thesis will take rolling bearing as the research object to explore the application of deep belief network in the fault diagnosis of mechanical components.Firstly,the condition monitoring method and fault types of bearing are analyzed,and the bearing vibration data provided by Case Western Reserve University is used as the original signal for fault diagnosis research and analysis.Secondly,the length of input data is determined according to the signal sampling frequency and sample length,and cut off the original signal to input data which have the specified length.At the same time,the fault categories are determined according to the purpose of fault diagnosis and fault types and fault diameters of sample,and the standard output data is set corresponding to the fault category of each sample.Thirdly,the process of frequency domain transformation and normalization are presented.Then,the corresponding DBN model is established and trained according to the length of input data and fault category,and the model is evaluated through multiple performance indicators.Fourthly,multiple combinations of hidden layers are set according to the different number of neurons in the hidden layer and the best combination of hidden layers is determined in this paper by comparison and analysis.Finally,the sensitive analysis for the position of sensor and outer race fault are made to provide guidance for the acquisition of condition monitoring data,and the method is used to further identify the different position of outer race fault.