Research On Rotor Fault Data Dimension Reduction Method Based On Multi-Manifold Learning

As the core equipment of petrochemical,electric power,metallurgy,and other industrial production fields,condition monitoring and fault diagnosis of rotating machinery are the keys to ensuring safe and efficient industrial production.Industrial big data is becoming a core element in the era of intelligent manufacturing,so the development of data-driven intelligent fault diagnosis and decision-making technology is of strategic importance for the intelligent operation and maintenance of large industrial production equipment in China.However,as modern industrial equipment automation,intelligence,and integration levels continue to rise,generating a large number of equipment operating status monitoring data showing high-dimensional,nonlinear,redundant,and other characteristics,the direct result is a dramatic increase in the demand for data storage space and computer computing power,as well as a serious lag in the ability to make real-time diagnostic decisions about the operating status of equipment.Therefore,how to reduce the size of raw data while extracting sensitive feature information that accurately characterizes the operation status of equipment as much as possible becomes the key to improving the efficiency of equipment fault diagnosis analysis.Multiple manifold learning theory is believed to be able to extract eigeninformation from the complex and high-dimensional data.It aims at reconstructing sensitive feature data in low-dimensional space and reducing data storage size by constructing an embedded graph structure model that fuses prior knowledge and label information,and it can facilitate data analysis.It has been widely used in many fields,including machinery fault diagnosis,face recognition,and text analysis.For these reasons,this thesis investigates the problem of dimensionality reduction of rotor high-dimensional fault datasets based on multiple manifold learning theory,and the main work is basically outlined as follows:(1)A dimensionality reduction algorithm based on strengthened intrinsic local preserving discriminant analysis(SILPDA)is proposed to address the problem of difficult fault classification due to the presence of a large number of high-dimensional non-linear mapping relations in fault datasets.The algorithm maximizes the retention of multiple manifold distributions and local manifold structure information of the data by fusing a strengthened intrinsic structure model with a local similarity matrix to extract a sensitive subset of low-dimensional features that are easy to implement for classification operations.The effectiveness of the algorithm was verified using an experimental dataset of rotor failures.The results show that the algorithm can reduce the difficulty of fault classification while still improving identification accuracy.(2)A dimensionality reduction algorithm based on median feature line multi-graph embedding(MFLME)is proposed to address the problem of limited identifiable information extraction from traditional point-to-point neighborhood graphs,which affects fault identification accuracy.The algorithm firstly improves the projection metric from sample points to feature lines to a median metric,which is used to weaken the extrapolation error of the algorithm;secondly,by defining the near-neighbour feature line graph and the far-neighbour feature line graph,the confusion of different types of samples is reduced to expand the class spacing,which reduces the difficulty for subsequent fault identification decisions.Finally,the feasibility of the algorithm was verified using two different rotor failure simulation experiments.The results show that the algorithm is able to extract a subset of sensitive features from the original fault dataset that are conducive to the implementation of classification operations.To address the problem of local inseparability of faults after dimensionality reduction of a single manifold embedding structure,a multiple feature space collaborative discriminant projection(MFSCDP)algorithm is proposed to reduce the dimensionality of rotor fault data sets.The algorithm improves the efficiency of constructing feature space embedding graphs by establishing a sample point-to-point guided near neighbor feature space selection method and indirectly constructing a reduced-dimensional projection matrix after multi-feature space collaboration using Relief F.Finally,the performance of the algorithm is verified and analyzed using experimental information from two rotor failure simulations of different structural types.The results show that the feature subset extracted by the algorithm will make the boundary between different fault categories clearer,increase the separability between categories,and finally achieve better fault identification.Multiple manifold learning can extract sensitive features characterizing the operating state of machinery,reduce the burden of classifiers,and have broad application prospects in the field of data-driven intelligent fault identification.In addition,the subsequent key exploration should include the direction of optimal selection of the nearest neighbor parameters of the manifold,optimal update of the projection matrix,etc.,and develop the system software of intelligent fault identification on this basis to promote the implementation and practical application of intelligent fault diagnosis technology.