Research On Fault Diagnosis Of Centrifugal Pumps Based On Support Vector Machine

As one of the most universal rotating machines,pumps have been widely used in many industries,including petrochemicals,water supply and drainage,agricultural irrigation,etc.The pumps' safe operation is crucial to the whole sysytem.The studying object of this paper is single-stage single-suction centrifugal pump.Based on the analysis of the existing method of fault diagnosis of centrifugal pump,aiming at the key technologies of feature extraction method and centrifugal pump fault mode recognition,the physical essence of the fault mechnism is investigated.The main research contents of this paper are as follows:This paper firstly analyze the common faults of centrifugal pumps,such as rotor faults and bearing faults,given the causes of these faults and the corresponding fault characteristics.Combined with the common fault forms of centrifugal pumps and the structural features of the experimental self-priming pump,the failure parts are designed and fabricated in the laboratory.Then we do the experiment and collect vibration signals in different stations.The key step of centrifugal pump fault diagnosis is feature extraction.In this paper,three kinds of fault feature extraction methods are propose.In the first place,frequency domain of the vibration signals are concidered.Through the spectrum we can find some obvious differences in all the state of the centrifugal pump,so we consider the frequence as the frequency-domain fault feature.In order to find the sensitivity and stability of the characteristics,we do the statistic analysis to the collected vibration data.If the pump works in the failure state,the vibration data would be changed,so we consider the Lempel-Ziv complexity of time series.Traditional Lempel-Ziv complexity method use mean value as the coarse-grained threshold value and would miss a lot of useful information.Here we propose a new coarse-grained method based on probability density function of time series.Four sections are supposed to be the best choice in order to keep enough information and computational efficiency for the fault diagnosis.From the result we can find that different situation has obvious difference.Considering the fact that characteristics of single scale extraction may miss some crucial fault information,we use multi-scale fuzzy entropy(MFE)to measure the multi-scale information of time series.Restrict by the limited experimental data,the support vector machine(SVM)is used for mode recognition.There are many measuring points in the experiment and a single measuring point can extract a lot of fault features,it's necessary to optimize the measurement points and feature vectors in order to improve the efficiency and recognition rate.Depending on the Max-Relevance and Min-Redundancy rule(m RMR)to extract appropriate measuring point and feature vector,we can reconstruct the feature vectors and train the least squares support vector machine(LS-SVM)with the new feature vectors.By the test of trained support vector machine we find it's good performation.Combined with feature vector extract from single method,the method of combined feature selection shows obvious advantages in this paper.