Study On Weak Fault Fusion Diagnosis Methods For Rotating Machinery Transmission Parts

With the development of modern industrial technology,rotating machinery is widely applied in advanced manufacturing,aeronautics and astronautics,naval architecture & ocean engineering,rail transportation,wind power,nuclear power and other significant engineering fields.Bearings,gears and other transmission components are indispensable core infrastructure components of rotating machinery in the process of power transmission and motion transformation.Once there are faults occurring,the whole system will be paralyzed,which will even lead to catastrophic accidents.Due to the harsh operating environment,variable working condition and complex structure of the rotating machinery,the running state signal tends to be weak,nonlinear,complexity,diversity and coupling.Moreover,it is submerged in the strong background noise and interfered by the extraneous signal,which may cause the fault feature extremely weak and difficult to be extracted.Therefore,it is of great significance to carry out the fault diagnosis of core components of the rotating machinery to avoid the occurrence of catastrophic accidents,prevent the casualties and improve the economic efficiency in enterprises.According to solve the key technical problems in rotating machinery,such as reliability and accuracy of the weak fault diagnosis in complex interference environment,the thesis carried out the research about weak fault diagnosis of the rotating machinery transmission parts using vibration signal and current & torque signal.In the aspect of the weak fault diagnosis based on traditional vibration signal,the research is conducted from the signal enhancement,feature extraction and fault recognition.A cascade enhancement method for nonlinear weak faults in the background of strong noise,a texture based method for weak fault feature extraction and a deep learning technique for fault recognition under the complex nonlinear coupling conditions is studied in the thesis.In the aspect of the weak fault diagnosis based on non-vibration signal,the current and torque signal acquired from the motor driving system is employed to characterize the rotating machine running state,and the fution of multi-source sensor information and multi-source decision model for accurate diagnosis and decision of rotating machinery is studied in the thesis.The detailed research content is as follows:(1)To address the problem that the fault signal is weak and the feature is difficult to extract under strong background noise,a cascade enhancement method based on minimum entropy deconvolution and energy operator is proposed to process the fault signal.Firstly,according to shock wave characteristics of the fault signal of rotating machinery,minimum entropy deconvolution is applied to reduce the noise in the signal using the kurtosis value as the objective function.The noise reduction is obtained while the transient impact noise components in the signal are also enhanced.Then,according to the characteristic that is suitable for detecting the instantaneous change of the signal,the energy operator algorithm is employed to effectively extract impact features of fault signal and further enhance fault characteristic of the signal.Then the fault impact characteristics submerged by the noise can be effectively detected through the calculation of the energy spectrum.The proposed method is validated using the rolling bearing hybrid fault data,which can effectively extract the characteristics of mixed fault and judge the corresponding fault type accurately.(2)Aiming to solve the problem that the non-stationary and non-linear feature is difficult to extract because of the influence of the strong background noise,a texture based method using the time-frequency distribution is proposed to characterise the running state of the rotating machinery.Firstly,an adaptive optimal kernel(AOK)time-frequency analysis method with excellent time-frequency resolution,strong adaptability and noise immunity is applied to process the low SNR fault signal,which can automatically adjust the kernel function to track the small change of signal,reduce the interference of background noise,and effectively suppress the cross-interference term and maintain the high time-frequency resolution.The simulation results show the performance of AOK,S-transform and Wigner-Ville distribution under different SNR conditions.After obtaining the time-frequency distribution image of the signal,the uniform local binary patterns(uLBP)is emloyed to extract the histogram of the local binary patterns(LBP)to represent fault information of rotating machinery beause uLBP can reduce the non-uniform features generated by random noise effectively.The performance of the proposed method is comparing with other methods in the references in the terms of the classification accuracy,classification stability and computational complexity under different operating conditions with different SNR noise.(3)To solve the problem that the weak fault signal of rotating machinery presents the complexity & diversity and the traditional shallow neural networks are difficult to recognize,a rotating machinery fault diagnosis method based on based on the optimizated deep belief networks is proposed.The time-domain statistical features,envelope spectrum features,the instantaneous frequency spectrum features and advanced statistical features are extracted to construct the original high-dimensional hybrid domain feature vector to characerize the operating state of the rotating machinery.The maximum relevance minimum redundancy(mRMR)method is used to eliminate error and redundant features,and then the spatial clustering of features was achieved by linear discriminant analysis(LDA).Combining with the advantages of feature selection and feature transformation,we obtained more compact and sensitive low-dimensional feature subset.Taking advantage of the deep nonlinear network structure of the deep belief network,a model describing the complex mapping relation between the fault characteristics and fault category was established.Compared with the traditional classification model,deep belief networks can realize more complex function approximation with the original data,and have the stronger ability of mining essential feature of data.The discrete particle swarm algorithm is applied to optimize the number of hidden layer selection to improve the recognition rate and stability of rotating machinery fault diagnosis.(4)Aiming to address the problem that it is difficult to install the sensors and the single source signal cannot characterize the equipment running states comprehensively,this thesis proposed a diagnosis method with multiple sensors information and multiple decision fusion models based on motor driving system.The motor driving system is used as the sensor to acquire the current signal and torque signal of motor to represent the running state of the transmission parts.The influence of transmission components faults on motor current and torque is also studied in this paper.After obtaining independent local diagnosis results of multiple sensors by SVM and ELM,homologous and heterologous sensors information fusion is realized using D-S evidence theory.Then,the fusion diagnosis results of multiple decision models are successfully obtained.The proposed method achieves the global consistent decision of status of rotating machinery and increases the accuracy,fault tolerance and robustness of fault diagnosis.At the end of the thesis,we summarized the main research contents and put forward the expection and direction for the future study.