Research On Fault Diagnosis Method Of Centrifugal Pump Rotor Based On Deep Learning

Centrifugal pumps are widely used in important fields such as national defense and security,industrial production,household water supply and drainage and agricultural irrigation.Rotor unbalance and rotor misalignment is the most common fault of centrifugal pump rotor.When the rotor unbalance and misalignment fault occurs,it will not only affect the hydraulic performance of the centrifugal pump itself,but also threaten the safe and stable operation of the unit.Therefore,in order to ensure the safety of industrial production and improve production efficiency,it is necessary to carry out the research of centrifugal pump rotor fault diagnosis method.This project focuses on horizontal centrifugal pumps and conducts research in three areas: fault signal processing,fault diagnosis models,and fault diagnosis methods for balanced and unbalanced data,based on two typical rotor faults: rotor unbalance and rotor misalignment.The aim is to provide new ideas and methods for the key issues encountered in centrifugal pump rotor fault diagnosis and to provide support for the safe and stable operation of centrifugal pumps.The main research contents and achievements of this paper are as follows:1.A summary and analysis of the current trends and status of fault signal processing techniques,fault diagnosis methods based on convolutional neural networks,and fault diagnosis techniques under non-uniform data conditions have been conducted.By comparing the advantages and disadvantages of different methods,a suitable approach for diagnosing centrifugal pump rotor faults has been determined.2.To address the problem of non-smooth and non-linear vibration displacement signal which makes it difficult to extract rotor fault features,an IPSO-VMD-KLD-based rotor fault feature extraction method for centrifugal pumps is proposed.Firstly,to improve the convergence speed and convergence accuracy of PSO algorithm,IPSO algorithm is proposed by introducing the strategies of improved Tent chaotic mapping,adaptive inertia weights and adaptive learning factors for improvement;secondly,the parameters of VMD are automatically searched for by IPSO algorithm to realize the adaptive decomposition of vibration displacement signal;again,the effective IMF components are established based on KLD Secondly,the IPSO algorithm is used to automatically find the optimal VMD parameters and realize the adaptive decomposition of the vibration displacement signal.Finally,the proposed method is validated by simulated signals and experimental data.The results show that the proposed IPSO algorithm has faster convergence speed and higher convergence accuracy compared with the PSO and GA optimization algorithms.In terms of signal reconstruction,the IPSO-VMD-KLD-based rotor fault feature extraction method can realize the adaptive decomposition and reconstruction of vibration displacement signals and accurately extract the rotor fault features.3.To address the issues of traditional fault diagnosis methods being time-consuming and existing deep learning models only being able to extract features from a single dimension,an improved 1D Le Net-5 model and 2D Le Net-5 model were proposed,and a centrifugal pump rotor fault intelligent diagnosis method based on a dual-stream CNN model was developed on this basis.The dual-stream CNN model was optimized through manual parameter tuning,and the effectiveness of the optimized dual-stream CNN model in diagnosing faults in centrifugal pump rotors was verified.The results show that when the optimizer is Adam,the learning rate is0.0001,the training batch size is 8,and the Dropout value is 0.4,the model performs the best.Moreover,the average validation accuracy of the model on the signal reconstruction dataset is 9.24% higher than that on the unreconstructed signal dataset.In terms of visualization analysis,as the depth of the model increases,the various types of fault data in the final validation set are classified into three independent clusters.In the comparative analysis of different models,the performance of the dual-stream CNN model is superior to that of other models,and the fault diagnosis accuracy remains at 100%.4.To address the problem of unbalanced number of normal state samples and fault state samples in the data set due to limited samples and difficult access to centrifugal pump rotor fault data,a CWGAN-GP model is proposed for generating rotor fault data based on CGAN and WGAN-GP models,and combined with a dual-stream CNN model to realize fault diagnosis of centrifugal pump rotor under data unbalanced state.The quality and performance of the data generated by the proposed method are evaluated and validated in terms of visual analysis,statistical metrics and comparison with different data generation models.The results show that the CWGAN-GP model is able to generate high-quality data similar to the original real data;on data sets with different degrees of unbalance,the dual-stream CNN model has better fault diagnosis results on the expanded data set of CWGAN-GP model compared with the rest of the generation models,and the improvement of fault diagnosis accuracy ranges from 1.40% to 13.33%.