Integrated Fault Diagnosis Model Using Tiny Residual Network

With the huge development of the manufacturing industry,the demand for safety protection systems of machinery has been continuously increased by related enterprises.At the same time,the intelligent fault diagnosis technology of mechanical equipment is a key process towards intelligent manufacturing,and it is also the trend of future industrial development.However,due to the insufficiency of the existing fault diagnosis technology,a large labor cost is required to assist in the monitoring and diagnosis of mechanical equipment faults.In response to this problem,the advent of the era of deep learning has put the field of mechanical equipment fault diagnosis on a rapid development track and has attracted widespread attention from academia and industry.Traditional fault diagnosis methods need to go through the process of feature extraction,feature selection,and feature classification.The processing process is complicated and the actual application cost is high.In addition,traditional fault diagnosis methods may also need to rely on the assistance and analysis of professional knowledge in related fields,otherwise,it will lead to greater uncertainty in the final diagnosis.Therefore,this paper proposes a new mechanical fault diagnosis model based on the idea of deep learning technology,combined with a tiny residual network,time-frequency analysis method,and specific activation function.The main contents of this paper are as follows:1.Because of the complicated fault diagnosis process of traditional methods and the large uncertainty of recognition accuracy,this paper proposes a deep convolutional neural network fault diagnosis model based on the residual network,referred to as ITResNet.The model uses the design of residual networks to overcome the vanishing gradient problem.ITResNet model can automatically identify mechanical equipment faults by operating directly over the original fault data during the training process.2.Aiming at the problem that the ITResNet model is difficult to deal with the serious noise pollution of the original fault data,this paper proposes to use three different time-frequency domain analysis methods to preprocess the original fault.Rather than viewing the fault data as a one-dimensional signal,time-frequency analysis studies a two-dimensional signal obtained from the signal via a time-frequency transformation.The results show that by combining timefrequency analysis techniques with the proposed model,the ITResNet model is improved and can accurately classify and detect mechanical fault perception data and diagnose faults.3.Aiming at the problem that the classification and fitting ability of the ITResNet model is difficult to reach the ideal state,this paper further improves the ITResNet model based on the new Mish activation function,Nadam optimizer technology,spatial pyramid pooling,and compressed excitation structure,which is called ITResNet-SPP-SE.By combining these technologies with our model,the total number of parameters is reduced.Experiments show that the fault classification accuracy and stability of the ITResNet-SPP-SE model is better than that of the ITResNet model,and the running time is lower.This paper uses the bearing failure data set to verify and analyze the model proposed in this paper.This paper proposes three models,which are the ITResNet model,the improved ITResNet model by time-frequency domain analysis techniques,and the ITResNet-SPP-SE model.The experimental results of these models show that the average accuracy of each model is 99.8%,99.99%,and 99.999% respectively.Moreover,the fitting ability of the ITResNetSPP-SE model is better than the fitting ability of the improved ITResNet model,which is better than the ITResNet model.In contrast,experiments using other methods and models,traditional methods such as SVM,KNN,and MLP have prediction accuracy rates of 71%,33%,and 87% respectively;methods based on deep learning technology,such as deep full CNN(DFCNN),stack down Noise AE(SDAE)and adaptive CNN,which operate on specific types of signals,have an accuracy of 99.22%,91.79%,and 97.90%,respectively.Therefore,it can be seen that the ensemble models we proposed are more effective in mechanical fault classification and prediction.