Fault Diagnosis Of Rotating System Based On Difference Scalogram And Convolutional Neural Networks

Rotating system is widely used in petroleum and petrochemical industry,mechanical equipment.Once the rotating system breaks down it may cause casualties,economic losses,and other serious consequences.There are two problems in rotor system fault diagnosis:(1)Rotating system fault diagnosis is mainly based on vibration signal which is under the high noise background and complex operating condition,that will lead to a low rate of fault diagnosis.(2)In the process of intelligent fault identification,fault feature selection is not affected by human factors,and the fault diagnosis rate is low.In view of the above two questions,taking the rotating equipment as the research object,the main research contents of this paper are as follows:(1)Analyzing main fault types,fault feature in time domain,fault feature in frequency domain and time-frequency domain from three aspects including theory and simulation and experimental signals.The results show that scalogram in the noise free case can clearly distinguish mass unbalance,shaft misalignment,rotor radial rubbing and base looseness fault,but under high noise background it cannot effectively identifies the fault type.(2)Proposing a difference scalogram method which is based on a combination of maximum correlation kurtosis degree solution of convolution filtering,Hilbert envelope spectrum and rearrangement scalogram.We can obtain a high amplitude in multiple harmonic component with this method,it means we can distinguish the fault types more clearly than the traditional scalogram.(3)To solve two problems one is the conflict between the convolutional neural network and the one dimensional vibration signal the other is low fault diagnosis rate,proposing a fault diagnosis method which is combined by the improved convolutional neural network and the difference scalogram.The implementation steps of the method are designed,and the fault diagnosis accuracy of this method is verified by the experimental analysis of the rotor system.The accuracy of the method is 96%.(4)Researching two main method restrict neuronal and dropout those were eliminating the overfitting,improving the sparsity of neural network and increasing the capacity of network generalization effect.And according to the determined optimal retention probability p equal to 0.4,convolutional neural network optimization for the deep sparse rectifier convolutional neural network.Experiment results show that the model we proposed raise classification accuracy by 8 percentage points than the traditional CNN.