| The piston pump has the advantages of high stability,high pressure and high efficiency.It plays an important role in industrial and production applications,and its working state plays a key role in the safe operation of equipment.The piston pump has complex structure and runs continuously at high speed.As a result,parts abrasion,slipper loose and other faults may occur during the operation,resulting in production safety accidents.Therefore,the fault diagnosis of the piston pump is an essential method to ensure the safe operation of the equipment.It is a key step to ensure the accurate and efficient completion of maintenance work.The traditional fault diagnosis method of piston pump has some flaws.Its diagnosis process is complex,the overall fault diagnosis efficiency is low,the adaptability is poor and it relies too much on the fault feature extraction technology and diagnosis experience.Therefore,the accuracy of the diagnosis method needs to be further improved.In order to identify the working state of plunger pump efficiently and accurately,this paper studies the application of machine learning method in the fault diagnosis of piston pump.Taking the axial piston pump as the research object,the fault diagnosis experimental platform is designed and built.The vibration signals of each state are processed and analyzed.Combined with machine learning method,the fault diagnosis of plunger pump is studied.The research contents and main work of this paper are as follows:1.Analyze and study the working principle,structural composition and common failure type mechanism of the piston pump.Four theoretical models of machine learning algorithm are applied.The fault test system of piston pump is designed.According to the requirements of the test,the preparation work of component selection,sensor installation,test parameter determination,data transmission channel setting and fault parts manufacturing are completed.Set up the fault test platform of piston pump.Complete the vibration signal sample collection work under different working conditions.2.Analyze the time-domain characteristics of experimental vibration signals with different faults.In view of the noisy working environment of the piston pump,which leads to the high signal-to-noise ratio of the vibration signal,the cosine neighboring coefficients(CNC)method is selected for preprocessing.The validity of the method is verified by comparing the time-domain characteristics of the simulated and experimental signal before and after de-noising.3.To slove the problem that the fault feature of piston pump is difficult to extract and the accuracy of diagnosis is not high,a typical fault diagnosis method based on support vector machine(SVM)is proposed.Firstly,five kinds of vibration signals collected from the experiment are preprocessed by CNC method.Then,after LMD decomposition,the correlation coefficient method is used to reconstruct the signal and further remove the irrelevant information.Finally,the fault feature is extracted by sample entropy,and the feature data set is constructed.Combined with the SVM algorithm of machine learning method,the typical fault types of plunger pump are classified.The accuracy of diagnosis is above 97.5 %.Compared with the classification results of KNN algorithm and BP neural network algorithm,the proposed method has higher diagnosis speed and accuracy.By inputting the original signal and the feature data set of reconstructed signal into SVM for training and diagnosis,the advantages of feature extraction of reconstructed signal are proved.4.Aiming at the problems of undesirable accuracy,complex diagnosis process and poor adaptability of piston pump wear fault state diagnosis under low load condition,a wear fault state diagnosis method based on random forest is proposed in this paper.The normal and four kinds of vibration signals of sliding slipper abrasion collected in the test are taken as the diagnosis objects.Firstly,the CNC method is used for preprocessing.Then the eight dimensional feature vector is extracted from the time domain feature to construct the feature data set.At last,the random forest algorithm is used for fault diagnosis,and the accuracy of diagnosis is 99.6 %.Compared with SVM,KNN and BP neural network,this method has the best accuracy and validity.The accuracy of slipper loose fault diagnosis is over 99.4 %,which proves that the method is suitable. |
