Research And Implementation Of Typical Fault Diagnosis Method For Vertical Centrifugal Pump Based On Machine Learning

This research is carried out under the support of the National Key Research and Development Program(2016YFF0203301)and the Commissioned Project of China General Nuclear Engineering Co.,Ltd.(20200513).As a key equipment in industrial production,the running state of centrifugal pump will directly affect the safety and economic benefits of industrial production.Therefore,in order to effectively prevent the damage caused by centrifugal pump failure to equipment operation and a series of dynamic effects,it is necessary to carry out the research on the typical fault identification and early warning method of centrifugal pump.In this paper,vertical centrifugal pump is taken as the research object,based on typical faults such as rotor imbalance,misalignment and pedestal looseness fault that are common in vertical centrifugal pumps,the research on signal denoising preprocessing,fault feature parameter extraction,fault feature set dimension reduction and fault pattern recognition is carried out,which are four key links of fault diagnosis based on machine learing,and develop a diagnosis system for typical faults of vertical centrifugal pumps by integrating the algorithms of each link.Which aims to provide new ideas and methods for key technical problems encountered in real-time and accurate diagnosis of typical faults of centrifugal pumps.The main contents and results of this thesis are as follows:1.This paper systematically introduces the main process of intelligent fault diagnosis of rotating machinery based on machine learing,and summarizes and analyzes the current research status and development trend at home and abroad from four aspects,fault signal acquisition technology,fault feature parameter extraction,fault type classification and fault pattern recognition.By comparing and analyzing the advantages and disadvantages of different methods,the diagnosis method suitable for the typical faults of centrifugal pumps is selected,and finally the development of the existing condition monitoring and fault diagnosis system is comprehensively summarized.2.Researched and analyzed the reasons and vibration mechanism of the centrifugal pump rotor imbalance,misalignment and pedestal looseness fault,and summarized the vibration characteristics and frequency spectrum characteristics of each fault.Through the built-up vertical centrifugal pump rotor test bench,the above mentioned faults were simulated and analyzed,and the influence law of different faults on the vibration characteristics of the unit under different operating conditions was mastered.3.In order to reduce the vibration signal of mutations and non-stationary random noise component’s influence on the fault feature extraction and pattern recognition,the Kalman filter method is used to denoise the original signal,and the characteristic frequency of the fault signal after filtering is more obvious,and the characteristic information of fault state can be reflected more accurately.Aiming at the problem of low fault recognition rate due to insufficient fault state information mining in feature parameter extraction based on single measure,based on the statistical characteristics of time domain and frequency domain and the time-frequency energy characteristics of Empirical Mode Decomposition(EMD),a multi-domain and multi-category method for extracting typical fault features of vertical centrifugal pump was proposed.At the same time,considering the reduction in computational efficiency caused by the complexity and dimensionality of the original fault feature set.A method for dimensionality reduction of the feature set according to the different contribution of the fault feature parameters is proposed,the ReliefF algorithm is used to weight the feature parameters based on multi-domain and multi-category extraction.According to the weight of each feature parameter in different faults,the feature parameter is selected based on the contribution threshold and combined with the Weighted Kernel Principal Component Analysis(WKPCA)method to realize the dimensionality reduction analysis of the fault feature set.The results show that the method effectively extracts the core principal component feature points of normal,rotor imbalance,misalignment and pedestal looseness fault,and make the feature points between different faults have better intra-class aggregation and inter-class dispersion.4.In order to solve the limitation of SVM engineering application and subsequent system development caused by the selection of key parameter values of Support Vector Machine(SVM)depending on personal experience,a multi-fault diagnosis method for vertical centrifugal pump is proposed to optimize the parameters of SVM.Particle Swarm Optimization(PSO)and Genetic Algorithm(GA)are used to optimize the key parameters of SVM globally.The two methods and the traditional SVM model are applied to the classification and identification of typical faults of vertical centrifugal pumps.The results show that,the traditional SVM model has a recognition accuracy of only 81.5% for feature sets composed of different types of working conditions,while the recognition accuracy of the two pattern recognition models based on PSO-SVM and GA-SVM reaches 100%.Compared with GA-SVM,in the iterative optimization process of PSO-SVM,the fitness value can reach convergence faster and better,indicating that PSO-SVM has more efficient fault identification capabilities.5.Based on the relevant signal processing and fault identification methods mentioned in this paper,the fault diagnosis system of vertical centrifugal pump is designed and developed by combining automation technology with virtual instrument technology.The software of the system uses LabVIEW as the master computer to achieve the parameters setting of experimental conditions,state monitoring and humancomputer interaction and other functions.Using PLC as the slave computer,to achieve data input and output,frequency conversion speed automatic control and monitoring and alarm functions,acquisition Matlab as the data processing platform,to achieve fault feature extraction,feature parameter evaluation and weight,fault pattern recognition and determination and other functions,MySQL database is used to realize the management and query of monitoring data and other functions.By setting up a sea water pump test bench in Jiangsu Zhenhua Pump Industry,the system has been verified for nearly 2600 hours.The results show that the measurement accuracy of the system meets the national acceptance standards and can identify the typical faults of the monitored objects in the operation process in a timely and effective manner.