Research On Piston Pump Fault Diagnosis Based On Improved LCD And SVM

With the rapid development of the railway industry,China’s high-speed railway has reached the world’s leading position in terms of operation scale and overall technology,which also puts forward higher requirements on the safety and reliability of train operation.The piston pump converts the mechanical energy driven by the motor into the pressure energy of the liquid,which is then fed into the electro-hydraulic rutting machine system in the form of pressure,and its working condition is crucial to ensure the safe operation of trains and improve transportation efficiency.In this thesis,taking the piston pump of electro-hydraulic rutting machine as the research object,a fault diagnosis model is studied by combining signal analysis technology and machine learning technology,which effectively solves the problems of modal mixing effect in the decomposition process of piston pump vibration signal under small sample conditions,as well as insensitivity in feature information extraction and the low accuracy in fault type identification.A decorrelation masking local characteristic-scale decomposition(DMLCD)method is proposed to overcome the modal mixing of the local characteristic-scale decomposition(LCD).The traditional method of suppressing modal aliasing is to add Gaussian white noise to the signal to be decomposed to assist in decomposition,and then integrate and average the decomposition results to cancel the added white noise,but the decomposition results are affected by human experience and the decomposition efficiency is low.The DMLCD method improves the decomposition process of LCD by adding a known mask signal to each decomposed signal and embedding correlation coefficient processing.The masking signal present in the components is computationally eliminated and decorrelation is performed between each two components,which effectively suppresses the modal aliasing in the decomposed components and improves the decomposition accuracy.A feature extraction method combining generalized refined composite multiscale dispersion entropy(GRCMDE)with feature overlap index is proposed for the shortcomings of inefficiency of scale factor selection and poor noise robustness in refined composite multiscale dispersion entropy(RCMDE)feature extraction.The method selects the scales with higher feature discrimination by calculating and ranking the feature overlap under each scale factor,and constructs the feature vector set by this method.After decomposing the vibration signal of the piston pump using the DMLCD algorithm,the ISC components with higher correlation with the original signal are screened for reconstruction based on the correlation coefficient principle.Subsequently,the GRCMDE values of the reconstructed signal under each scale factor are calculated,and the feature overlap index is introduced to select the GRCMDE values under the top five scale factors to construct the optimal feature vector set,which makes the feature information extraction more sensitive.To solve the problem of low classification accuracy of the traditional support vector machine(SVM)model,the weighted mean of vectors algorithm(INFO)is proposed to optimize the fault diagnosis model of the support vector machine.The model takes the extracted optimal feature vector set of piston pump as input for parameter optimization.To verify the diagnostic effect of the model,the optimal feature vector set is constructed using four kinds of piston pump fault data for testing,and the average of 10 test results is taken as the final recognition accuracy,and the accuracy reaches 99.29%.And the effectiveness of the proposed model is proved by various comparison experiments.