Study On Bearing Fault Diagnosis Algorithm Based On Convolutional Neural Network

Mechanical equipment is becoming much larger,faster,more precise,more systematic and more autonomous in current industrial society.The key to ensure the safety of machine is to set up reliable health monitoring system.The mechanical health condition monitoring is entering the age of ‘big data' for much more monitoring points and sampling rate.Traditional intelligent fault diagnosis methods based on signal processing and classifier cannot meet the requirements of mechanical big data,because those methods need proficient experience of experts,much time to design and cannot ensure their general ability.This paper proposed a series of intelligent bearing fault diagnosis methods based on convolutional neural network to address the problem above,the proposed methods can automatically extract features and detect the faults without any expert experience.First,this paper presents a convolutional neural network with two convolutional layers to diagnose the temporal vibration signals.With the help of data augment,this model can achieve more than 99% accuracy on CWRU bearing dataset.The proposed model was the first convolutional model which works directly on the temporal vibration signal for bearing fault diagnosis.Second,a convolutional neural network framework named WDCNN for bearing fault diagnosis is presented.This model works directly on vibration signals and is featured as follows: it has large first-layer convolutional kernels and very small convolutional kernels in the following layers.The model is easy to train because of the batch normalization layers.Besides,it is easy to design the model with the given design rules.WDCNN can obtain 100% accuracy on CWRU bearing dataset.Third,the noisy environment and various working load problems which can degrade the performance of diagnosis system are addressed.This paper treated the problems as domain adaptation problems in machine learning,and introduced adaptive batch normalization algorithm to improve the performance of WDCNN model.As result,this method helps WDCNN model hold high accuracy in noisy and various working load environment.In addition,visualization technologies are employed to show the signal processing of WDCNN.Last,TICNN model is proposed to cover the shortage of adaptive batch normalization method which needs statistics of all the testing samples.First layer convolutional kernels with changing dropout rate and small mini-batch training are employed as training interference to train the neural network,which enhance the general ability of the model.Ensemble learning is used to improve the accuracy and stability of the model.Although no information of testing samples and any other denoising preprocessing methods are used in TICNN model,TICNN can perform well in noisy and various working load environment.