Research On Monitoring Method Of Centrifugal Pump Cavitation Based On Support Vector Machine

This thesis is carried out under the support of the National Natural Science Foundation(51976079)and the National Key Research and Development Project(2018YFC0810500).Centrifugal pump has been widely used in urban water supply,chemical industry,car,military industry and so on.Therefore,the stable operation of centrifugal pumps is of great significance,this is not only related to whether production and life can be stabilized,but also closely related to people’s daily safety.Cavitation is a common multi-phase,complex and unstable flow phenomenon in pumps.The cavitation of the centrifugal pump will cause performance degradation,cause large vibration noise and even cause material damage.Therefore,monitoring the operating status of centrifugal pumps can help technicians make judgments about the operating status of centrifugal pumps,stop operating or return to factory for maintenance to reduce economic loss,avoid major accidents,and increase the service life of centrifugal pumps.In this thesis,a miniature high-speed centrifugal pump is taken as the research object,and the vibration signals of the three operating states of normal operation,slight cavitation and severe cavitation of centrifugal pump under fixed flow and four rotational speeds are collected.The time domain analysis method,frequency domain analysis method and timefrequency domain analysis method are used to extract the vibration signal respectively,and then uses support vector machines and BP neural networks to classify the feature vectors to achieve the purpose of monitoring.The main research contents and innovations of this thesis are as follows:1.The mechanism and influence of cavitation are introduced,the current monitoring methods of centrifugal pump cavitation operation status domestic and overseas are summarized.It is believed that choosing the appropriate monitoring method during cavitation monitoring can improve the monitoring accuracy and efficiency.A miniature high-speed centrifugal automobile electronic pump was selected as the research object,a centrifugal pump closed test rig was built.The program written in LabVIEW was used to collect the external characteristics and vibration signal data of the centrifugal pump,and the cavitation performance curve of the centrifugal pump was drawn.The experiment found that the external characteristics at different speeds conformed to the similar law,and the cavitation curve also had a hump.At the same time,the uncertainty of the test data in this thesis is analyzed to verify the accuracy of the test bench.2.Choose three feature extraction methods to extract vibration signals in different states.The time domain analysis method selects mean,standard deviation,skewness,and kurtosis.In the frequency domain analysis method,the sensitivity index of the frequency domain is analyzed first,then choose to observe frequency multiplication and frequency band amplitude as feature vector.The time-frequency domain analysis method selects the Hilbert-Huang transform(HHT)for the feature extraction,this method gets rid of the limitations of Fourier analysis,which can analyze non-steady-state nonlinear data more accurately.3.A centrifugal pump cavitation monitoring method based on support vector machine(SVM)is proposed.The grid parameter optimization and K-CV cross-validation are used to optimize the support vector machine parameter group,which improves the accuracy and promotion of the support vector machine,the study found:(1)During normal operation and slight cavitation identification and classification of centrifugal pumps,it is found that the recognition rate of standard deviation and kurtosis in time domain features is high.In the frequency domain characteristics,the 1~2kHz amplitude recognition accuracy of frequency conversion characteristics in three directions of tongue and inlet pipe and three directions of the inlet pipe is more than 90%.Among them,the amplitude perform of 1~2kHz in the three measuring points directions of the inlet pipe is the best.In the time-frequency domain characteristics,the IMF energy ratio at the tongue of the pump volute as the feature vector has the highest recognition accuracy.(2)When the three operating states of the centrifugal pump are identified and classified into normal operation,slight cavitation and severe cavitation,the classification recognition rate of standard deviation as feature vector is the highest in the time domain characteristics.In the frequency domain characteristics,the 1~2kHz amplitude recognition accuracy of the three frequency conversion directions and the three measurement directions points of the inlet pipe are more than 90%.In the time-frequency domain characteristics,only the signal recognition rate at the tongue of the pump volute is above 90%.In order to further improve the accuracy of the support vector machine for monitoring the three operating states of the centrifugal pump,the combined features are also used for the training of the support vector machine.The recognition accuracy of the model for the three states at four rotational speeds are all higher than 95%.4.This thesis compares and analyzes BP neural network recognition accuracy and SVM recognition accuracy,and the study found:(1)When BP neural network classifies and recognizes time-domain features,except that the average recognition accuracy of the average feature vector is slightly higher than that of SVM,the average recognition accuracy of the remaining three feature vectors is not as high as that of SVM.(2)When BP neural network classifies and recognizes frequency domain features,it is found that the SVM has two types of feature average classification recognition rate which is more than 90%,while the BP neural network has only one type of features average classification recognition rate which is more than 90%.(3)When BP neural network classifies and recognizes time-frequency domain features,the features average recognition rate of HHT at the volute tongue and the X direction of the inlet pipe is lower than that of SVM,although the HHT feature recognition rate of the Y and Z direction of the inlet pipe is higher than SVM,but the average recognition accuracy rate is only 70%.In addition,when training models with timefrequency features as feature vectors,SVM has a type of features with an average classification recognition rate above 90%,while BP neural networks have no feature recognition rate above 90%.The thesis shows that the overall recognition accuracy of SVM is higher than that of BP neural network,which proves the feasibility of using support vector machines to monitor the cavitation state of centrifugal pumps.