Research On Fault Diagnosis And Prognosis Of A Class Of Nonlinear Systems Based On Intelligent Intergrated Methods

It has become an unavoidable major problem that how to improve the safety and reliability of the system in the process of operation,as well as reduce the potential safety hazards,with the development of modern industrial system.Fault diagnosis and prognosis technology is an important method and necessary means to improve the system operation reliability and reduce the system operation risk.Most of modern nonlinear industrial systems have the characteristics of strong interference,complex structure,uncertain parameters,dynamic time-varying,coupling and uncertainty of faults,which makes it difficult in fault diagnosis and prognosis.Therefore,it is a very meaningful research topic to explore how to make use of advanced sciences and technologies,purposefully,methodologically and pertinently to carry out effective fault diagnosis and prognosis of the systems.In this dissertation,taking the real nonlinear systems as the research objects,intelligent technologies based fault diagnosis and prognosis methods are studied,including filtering technology,signal processing,optimization algorithm,neural network and deep learning.The main innovative research outcomes are as follows:(1)Two fault diagnosis methods of nonlinear system based on intelligent optimized particle filter are proposed.In the application of particle filter based fault diagnosis,particle degradation and lack of diversity will lead to the decrease of state estimation accuracy,which then affects the effect of diagnosis accuracy and the robustness of diagnosis system.In the sight of this kind of problems,taking the improvement of resampling strategy as the starting point,combining with the variable frequency based mutation(VFM)strategy and the beetle swarm antennae search(BSAS)algorithm,two improved algorithms of VFM-PF and BSAS-PF are proposed respectively in this dissertation.VFM-PF is inspired by combination of the mutation operation in the immune theory and frequency conversion energy-saving strategy in industrial production process.It adjusts the number of mutation particles in real time through frequency conversion operator,and adopts different forms of mutation operation for particles with different weights,and finally increases the diversity of particles and the computation efficiency,which improves the overall performance of the algorithm.BSAS-PF takes the advantages of the optimization characteristics of the BSAS algorithm to guide low weighted particles to move to high likelihood area.BSAS-PF overcomes the problems of particle degeneration and lack of diversity,which makes the state estimation accuracy better than the traditional PF.The researches on the fault diagnosis of the complex nonlinear system based on these two proposed methods are carried out by taken the doubly fed induction generator(DFIG)in the wind turbine system and the dissolved oxygen process in the aeration tank of the wastewater treatment process as the objects.Results indicate that the improved strategy can not only guarantee the robustness and computation efficiency of the diagnostic system,but also achieve the high performance of fault detection and isolation under the state jumping condition of the systems.(2)An improved ADCS-ELM-based fault diagnosis method is proposed.Problems of the high complexity of nonlinear system,the high degree of fault coupling,and the difficulty of establishing accurate mathematical model would make the model-based fault diagnosis methods hard to get the accurate diagnosis results or poor diagnosis accuracy.Therefore,a fault diagnosis method based on the improved extreme learning machine(ADSC-ELM)is proposed from the perspective of data-driven in this dissertation.Firstly,the nonlinear and nonstationary vibration(fault)signals of bearings are decomposed by the ensemble empirical mode decomposition(EEMD)to obtain the feature representation of the IMF energy.Secondly,the dynamic adaptive step size adjustment strategy is added to the improved cuckoo search(CS)algorithm,which is used to optimize the random assigned parameters of input weights and hidden layer thresholds in ELM to improve the stability,robustness and classification accuracy of the ELM network.Finally,train the ADCS-ELM network and test the fault diagnosis performance of it.An average accuracy of 99.51% of the simulation results achieved in diagnosising the fault categories of rolling ball,inner race and outer race of the bearing.(3)A fault diagnosis method based on optimized stacked denoising auto-encoder deep neural network is proposed.The nonlinear fault has the characteristics of transmissibility,coupling,secondary,uncertainty and diversity,which leads to the problems of difficult to trace the cause of fault effectively and low accuracy of fault diagnosis.Compared with the traditional fault diagnosis methods,stacked denoising auto-encoder(SDAE)can deal with a large number of unlabeled bearing fault vibration data,and adaptively extract deeper fault features to achieve the classification of bearing fault status,avoiding the tedious process of manually designed fault features extraction,which is conducive to improving the accuracy and efficiency of fault diagnosis classification.However,the hyperparameters of SDAE network obtained by experience enumeration performance a weak generalization ability in different fields of fault classification problem.This kind of diagnosis process is also related to the designer's experience,and the efficiency is low.Therefore,a newly designed optimization algorithm,called artificial transgender longicorn algorithm(ATLA)in this dissertation,is employed to adaptively determine the hyperparameters of the SDAE network to improve the accuracy of diagnosis and the generalization performance of the network model.The simulation results,which are conducted under varying operation conditions on the bearing fault data set,show that the proposed ATLA-SDAE method is superior to back-propagation(BP)neural network,support vector machine(SVM)and convolutional neural network(CNN)in generalization performance and fault classification accuracy.Moreover,ATLA-SDAE has a better diagnositic performance of efficiency and real-time than the CNN,which is a kind of deep learning model with deeper structure.(4)A real-time fault prognostic method of key process parameters of complex systems based on an optimized extreme learning machine is proposed.It is an effective fault prognostic method to realize fault prognosis through the prediction of the system key parameters in real-time.In this dissertation,considering the indirect measurement idea of soft sensor,a novel ICS-ELM neural network is constructed for system key parameter prediction purpose.First of all,an improved cuckoo search(ICS)algorithm based on the dynamic adaptive search step size and dynamic discovered probability adjustment strategies is employed to optimize the parameters such as the connection weights and the hidden layer biases of ELM regression network,which improves the stability and prediction accuracy of the ELM network.Secondly,the instrumental variables are selected for target parameter prediction through the principal components analysis(PCA)by reducing the dimension and the attribute of process data.Finally,an ICS-ELM prediction model of biochemical oxygen demand(BOD),which is a hard-tomeasure key parameter of wastewater treatment process,is established.The results show that the proposed method can not only achieve the accurate prediction of BOD,but also provide a feasible solution for the fault prognosis of the other key parameters of wastewater treatment process,as well as in other process industries.(5)A method of remaining useful life(RUL)prediction based on deep learning convolutional neural network is proposed.RUL prediction of equipment is an important research content in system fault prognostic.However,the complex nonlinear system has a large number of parameters and a large amount of process data with high dimension,which makes it difficult to select feature variables and establish accurate prediction model to realize effectively high performance fault prognosis.To solve this problem,an image features based RUL predicting method is proposed in this dissertation.Firstly,preprocessing the data sample set to obtain the graphical data set used for prediction work.Secondly,optimized CNN network is employed to extract the feature map from the graphical data set to construct high performance CNN-HI health indicator for representing the state of each degradation stage of the equipment.Finally,a Gaussian process regression(GPR)-based analysis is used for RUL prediction,which is verified by the PRONOSTIA ball bearing data set.Results illustrate that the proposed method has high prediction accuracy of health indicator to map the state of degration of a bearing effectively,which is helpful to realize the RUL prediction accurately in this study.It provides an important theoretical reference for RUL prediction of bearing and the other machineries,and it has a certain practical value.In this dissertation,an in-depth study on model-based and data-driven-based fault diagnosis and prognosis methods in the background of strong disturbance and multi disturbance is performed,and a variety of novel fault diagnosis and prognosis methods are proposed.To a certain extent,the proposed methods can solve the problems of low accuracy,poor real-time performance,and weak robustness of fault diagnosis,and lack of prognostic methods,brought by the system complexity,fault coupling and uncertainty characteristics of the system under various complex conditions such as state jumping and multi operation conditions.The research results have important reference value and practical significance for the development of intelligent technologies-based fault diagnosis and prognosis,and assurance of the operational security and reliability of nonlinear system.