Research On Key Technologies Of Intelligent Fault Diagnosis Based On Multi-sensor Information Fusion For Rotating Machinery

In mining,petrochemical,metallurgy and building materials an d other heavy-large engineering fields,rotating machinery is the main form of various t ypes of machinery and equipment.At the same time,its fault types also occupy the vast majority of mechanical equipment failure.Therefore,carrying out the research on fault diagnosis of rotating machinery plays a significant role in ensuring the continuous,stable,safe and efficient operation of mechanical equipment.For the three key problems existing in the field of research on fault diagnosis of rotating machinery which easily ignored: weak fault signal caused by strong background noise,incomprehensive and inaccurate information acquisition based on single sensor and the urgent needs for intelligent fault diagnosis,this subject takes rotating machinery in the background of noise as the research object and surrouds the main line of multi-sensor fusion intelligent fault diagnosis process,and gradually developes further research on weak signal detection,feature extraction and evaluation,feature reduction and intelligent model classification,which aims to provide a new way to solve some key technical problems of intelligent fault diagnosis of rotating machinery based on multi-sensor fusion in noise background.The main contents of this thesis are as follows.Aming at the problems of non-ideal signal enhancement effect caused by a large number of useful frequency components when multi-frequency weak signal is merely detected by SR method,a novel feature enhancement detection method of multi-frequency weak signal based on Empirical Wavelet Transform(EWT)and adaptive parametric compensation Stochastic Resonance(SR)array is proposed.EWT is utilized to adaptively decompose the multi-frequency signal into a number of Amplitude Modulation-Fequency Modulation(AM-FM)components with tightly supported Fourier spectra,in order to make sure the singleness of of frequency component in each component signal.An adaptive parameter compensation SR array model is constructed to realize the parallel detection of multiple single component weak signals.Each unit of this model utilize the structure of "parameter compensation pre-processing operation + classic bistable stochastic resonance model",in which the parameter compensation link is used to realize the conversion of the middle and high frequency components to the low frequency components to cope with resonance detection of weak periodic signal of arbitrary frequency,at the same time the parameters of the resonant elements are optimized by the proper improvement of the classical quantum genetic algorithm.Finally,the resonant output of each unit is band-pass filtered and then synthesized to obtain the enhanced version of original multi-frequency weak signal.The analysis of the simulated and measured fault signal show that the proposed method can efficiently realize the feature synchronization enhancement of multi-frequency weak signal.Aiming at the one-sidedness of traditional evaluation model based on single measure in fault feature sensitivity learning,a multi-domain and multi-class fault feature weighting method based on multi-measure optimal product mixture model is proposed.Firstly,the time-domain,frequency-domain statistical analysis and the time-frequency domain analysis based on EMD of the enhanced vibration si gnal obtained in the previous chapter are used to obtain a multi-domain and multi-category primitive feature set including time-domain,frequency-domain statistical feature and EMD-based energy and Lemple-Ziv Complexity feature.Based on this,a product mixed feature evaluation model is proposed,which takes into account the measures of distance,information and correlation,and is used to evaluate the original feature set.Based on this,the best multi-measure combination strategy is selected from multiple candidate mixed models based on the principle that the variation coefficient of comprehensive feature sensitivity scores is the largest.Then the mixed evaluation model corresponding to the optimal measure combining strategy is utilized to conduct comprehensive sensitivity learning on the original fault feature set.Finally,the original feature set is weighted by the evaluation score of each feature parameter in the form of the corresponding feature weight.Comparative analysis with the other four weighted methods based on single measure shows that the proposed method can make the weighted fault feature set have better clustering and classification ability.In order to overcome issues that the structure and spatial information of different sensor data are usually ignored and the serial superposition of feature dimensions exacerbates the occurrence of "dimensionality disaster" in the traditional multi-sensor ‘feature-feature' feature-level fusion method of rotary machinery fault diagnosis,a multi-sensor fusion fault feature dimension reduction method for rotating machinery based on Supervised Second-order Tensor Local Preserving Projection algorithm under Assembled Matrix Distance Metrix(SSTLPP-AMDM)is proposed.The second-order tensor is used for time-space fusion of fault feature information from multiple sensors.Compared with the popular vector expression of fault featrues from single sensor or multiple ones,its technical advantages are mainly reflected in the access to more state information while r etaining spatial structure information between fault features from different sensors,and will not cause the dimension surge of features.Compared with the classical Second-order Tensor Local Preserving Projection(STLPP)algorithm,the proposed SSTLPP-AMDM algorithm not only considers synchronously the local neighborhood information and the class label information,but also utilizes the AMDM which with better matching performance between the second order tensor data than the existing Frobenius distance metrix in the process of similarity weighting matrix construction.It is decided that the proposed new algorithm has a better dimensionality reduction effect than the classical STLPP algorithm.Aiming at the problems that the conventional diagnosis results of rotating machinery based on single pattern recognition and classification algorithm often show significant uncertainties,a multi-sensor decision fused fault diagnosis method for rotating machinery based on the intergration of Support Tensor Machine(STM)and k-nearest neighborhood(KNN)classification model with Assembled Matrix Distance Metrix(KNN-AMDM)is proposed.Taking the low-dimensional second-order tensor feature sample data obtained from the dimension reduction of multi-sensor feature-level information fusion in the above chapter as the input,the feasibility of STM classification model in fault diagnosis of rotating machinery is discussed,as well as the technical advantage by compared with the traditional Support Vector Machine(SVM)model.Besides,the traditional KNN classification algorithm is improved by the AMDM,so that it can be applied to the classification of second-order tensor data.Based on the diagnosis results of the aformentioned two classification models,basic probability assignments of the corresponding evidence bodies are constructed.And then the final accurate fault diagnosis conclusion is obtained by the evidence synthesis and diagnostic decision.Finally,a multi-sensor information fusion based fault diagnosis platform of rotating machinery in noise background is realized.Based on the brief requirement analysis of the platform,the overall architecture of the platform and the structure of the functional sub-systems are designed.Through the fault diagnosis example of the main reducer of raw material vertical mill in the new dry cement production line,the application effect of the four parts studied in this thesis(ie,the enhancement detection of weak signal,the multi-domain multi-class fault feature extraction and the multi-measurement mixed evaluation weighting,the fault feature dimension reduction based on multi-sensor tensor fusion,the multi-sensor decision fused fault diagnosis)are analyzed in detail.Finally,the practicality and technical superiority of the new methods among the four key technical problems involved in the multi-sensor fusion intelligent fault diagnosis for rotating machinery in noise background are verified by the comparison analysis results of the application effect.