Research On Fault Feature Extraction Technology Of Mechanical Equipment

Mechanical equipment is one of the important components in the manufacturing industry.With the rapid development of artificial intelligence and industry,the manufacturing industry is gradually becoming intelligent.The structure of the mechanical equipment is complex,the parts are closely related and cooperate with each other,and the working environment of the mechanical equipment is mostly harsh.Moreover,the rolling bearing has the highest frequency of use in the mechanical equipment and is more prone to wear.The normal operation of the rolling bearing directly determines the operation of the whole mechanical equipment.Therefore,this topic carries on the characteristic extraction to the mechanical equipment rolling bearing each kind of fault type,and designs the classifier model to carry on the fault diagnosis.In this paper,the signal is firstly denoised by wavelet packet.Since there may be multiple fault type signals in the signal collected from a single sensor,ICA algorithm is selected in this paper to extract the characteristics of the signal and separate the fault signals of different types,so that the signal of a single fault type can be processed later.In this paper,the EMD method is selected to extract the characteristics of the signal.Due to the endpoint effect in the EMD decomposition,support vector regression is introduced to improve the EMD decomposition effect.Then,according to the energy distribution,the number of decomposition layers is determined to obtain the eigenvectors.At the same time,this paper also introduces AR model,combines EMD decomposition method with AR model,and takes the autoregressive coefficient and model variance of AR model as the feature vectors.At the same time,a fault diagnosis classification model is designed to test two methods of fault feature extraction.Firstly,the structure of the main classifier of neural network was designed,and elastic network was introduced to suppress the over-fitting phenomenon.Meanwhile,the weights of the neural network were iteratively updated based on the loss gradient based on Adam algorithm,and then the subordinate classifier was constructed based on k-d tree,so as to improve the accuracy of the whole classification model.After training and testing the fault feature vectors extracted by the two methods,the classification model has a higher accuracy rate of 92.17% for the test set composed of the feature values extracted based on AR and EMD,which proves that the feature values extracted based on AR and EMD model can better reflect the state information of the fault.