Research On Neural Network Optimization And Recognition Method For Intelligent Diagnosis Of Mechanical Faults

In recent years,with the rapid development of science and technology and social economy in China,Mechanical equipment is also constantly updated.Therefore,the traditional fault diagnosis technology is also developing towards intelligent diagnosis.Pump is a common rotating machinery,it works is to convert mechanical energy into hydraulic energy to transport liquids with simply describing.Because of its simple structure and durability,the pump is widely used in irrigation,petrochemical engineering,water conservancy,metallurgy and shipbuilding and so on.Centrifugal pump is the one of many types of pump,centrifugal pump has many advantages,such as wide applicability,small size,light weight,low cost,simple structure,easy operation,fewer faults,long service life,no pulsation of discharged liquid and so on.So the centrifugal pump is the most widely used.The internal parts of centrifugal pump will inevitably fail when it works.The timely diagnosis of the centrifugal pump fault is very important.In this paper,centrifugal pump as the main research object,on the basis of fault diagnosis based on probabilistic neural network,combined with principal component analysis method,the diagnosis of centrifugal pump is studied.First,the centrifugal pump fault diagnosis experiment is designed.Select impeller to set four failure modes for experiment.The vibration data collected in the experiment are used to analyze the failure of the centrifugal pump.However,many noises are mixed in the vibration data,so the original signal must be de-noised.In this paper,wavelet packet threshold denoising is used to process the original vibration signal.Then extraction of characteristic parameters from noise reduction signals.In this paper,time domain analysis is used to extract feature parameters,the energy characteristics of the extracted signals are also reconstructed by wavelet packet decomposition as feature parameters.Secondly,due to the large number of commonly used related characteristic parameters extracted from the signal after noise reduction,in order to reduce the parameters and at the same time ensure the accuracy and reliability,this paper adopts the principal component analysis(PCA)method for processing.PCA can transform a large number of original variables into a small number of independent new variables while retaining enough original information,and is often used in signal processing.Because of its simple concept and convenient calculation,principal component analysis method is suitable for dimensionality reduction of extracted characteristic parameters.Finally,because probabilistic neural network has been widely used in pattern classification,compared with traditional BP neural network,probabilistic neural network(PNN)has simple structure,easy to design algorithms,and can use linear learning algorithm to realize the function of non-linear learning algorithm.Therefore,a method of combining principal component analysis with probabilistic neural network is proposed.The new feature parameters optimized by principal component analysis are brought into probabilistic neural network and verified by Tri-Fold cross-validation.After that,it was compared with the two control groups set in the experiment,and the advantages of probabilistic neural network combined with principal component analysis method for centrifugal pump fault diagnosis were reflected in the accuracy and speed of diagnosis.