Intelligent Fault Diagnosis Method Of Hydraulic Axial Piston Pump Based On Time-frequency Analysis And Convolutional Neural Network

Hydraulic axial piston pump is known as the power heart of hydraulic transmission systems.It has been widely used in national defense equipment,industrial equipment,emergency rescue equipment,exploration and exploitation equipment for deep-sea and deep earth resources and scientific research equipment.The work reliability of a hydraulic axial piston pump is critical to ensure the high precision,the high speed,the continuous and stable operation of equipment.Its failure can lead to great impact and severe losses,such as downtime,breakdown of production line,and even catastrophic crash.Therefore,intelligent fault diagnosis of a hydraulic axial piston pump is investigated to enhance its operational reliability and reduce the frequency of failure.In response to the development needs of national intelligent manufacturing,this dissertation explores new ideas and methods of artificial intelligence theory by integrating data mining,machine learning and other intelligent technologies.Theoretical basis is provided for intelligent fault diagnosis and health management of a hydraulic axial piston pump,and the intelligence and reliability are improved.The main research contents of this dissertation are as follows:1.The mechanism of typical faults of a hydraulic axial piston pump is analyzed comprehensively.The basic principles of classical time-frequency analysis methods are briefly introduced.The basic structure and training process of convolutional neural network(CNN)are studied.The principle and implementation method of the Bayesian optimization algorithm are explored.An intelligent fault diagnosis method is proposed combining time-frequency analysis and CNN.This method provides a theoretical basis for intelligent fault diagnosis of a hydraulic piston pump.2.The fault simulation experiment system of a hydraulic axial piston pump is established for the simulation experiment.Then data acquisition of typical fault states is completed for the training and verification of intelligent diagnosis model.Owing to the superiority of continuous wavelet transform(CWT),synchronous compression wavelet transform(SWT)and S transform(ST)in analyzing non-stationary signals,the time-frequency domain conversion of original signals is accomplished including multi-condition and multi-source information.The fault sample dataset is constructed by sample division and label configuration,which provides input for subsequent training and verification of the diagnosis model.3.Two fusion methods of CWT with the improved Le Net 5 and Alex Net are proposed based on traditional classical models.The diagnosis performance of the methods is studied by analyzing the time-frequency characteristics of vibration signal,sound signal and pressure signal.Many key hyperparameters are analyzed to explore the effects on the diagnosis results,such as learning rate,epoch,batch size,dropout ratio,number and size of convolutional kernels.The training and structural parameters of the improved Le Net 5 and Alex Net are preliminary obtained.The Bayesian optimization algorithm is introduced to realize the adaptive learning of the key hyperparameters.The average diagnostic accuracy of the intelligent method combining improved Le Net 5with CWT reaches over 96.94%.The average diagnostic accuracy of the method combining Alex Net with CWT reaches above 98.44%.4.By combining SWT with the improved Le Net 5 and Alex Net model,two intelligent fault diagnosis methods are proposed.The diagnosis effect of the integrated methods is researched by analyzing vibration signal,sound signal and pressure signal.The Bayesian optimization algorithm is used to complete the automatic optimization of the key hyperparameters.The average diagnostic accuracy of the intelligent method combining improved Le Net 5 with SWT reaches more than 94.31%.The average diagnostic accuracy of the method combining Alex Net with SWT achieves above96.69%.5.An adaptive normalized CNN model is constructed by further improving the Le Net 5.The Bayesian optimization algorithm is employed to optimize the hyperparameters,which improves the comprehensive performance of the diagnosis model.An intelligent method is proposed based on the combination of CWT,SWT and ST with the normalized CNN model for fault diagnosis of a hydraulic axial piston pump.The average classification accuracies of five typical states reach above 97.46%,96.47%and 85.96% respectively by the intelligent fault diagnosis methods combining the adaptive normalized CNN model with CWT,SWT and ST time-frequency features.Compared with the diagnostic method combining traditional Alex Net model and time-frequency feature,the average computation time of the proposed method in one epoch is reduced by about 2 times.In this dissertation,an intelligent fault diagnosis method is proposed based on the integration of multiple time-frequency transform and improved CNN.The diagnosis effects of the fusion method are innovatively explored based on multi-source heterogeneous signal of a hydraulic axial piston pump.The classification and recognition performance of the diagnostic models are improved.And the fault diagnosis accuracy and generalization ability based on multi-source heterogeneous signals are enhanced,which lays a theoretical foundation for the development of new intelligent fault diagnosis method and engineering application of hydraulic axial piston pump.