Research On Dimensionality Reduction Method Of Rotor Fault Dataset Based On Hypergraph Model

In recent years,the continuous application of sensors and advanced facilities for monitoring and testing has led to a spurt in the growth of information characterizing the operating status of equipment,and this situation is driving the research of rotating machinery fault diagnosis into the era of "industrial big data".However,how to efficiently solve the problem of scientific accumulation and exploitation of big data resources is an important foundation for the development of machine intelligence decision technology for equipment,especially for the realization of intelligent rotating machinery.In the fault diagnosis technology of rotating machinery,the quantitative features of the collected vibration signals are extracted from multiple domains and angles to describe the operation status of the equipment.With the increase of the number of extracted features,the description of the equipment state becomes more and more comprehensive,but the problem is that the feature set inevitably contains redundant information,which not only increases the complexity of the operation but also is not conducive to the subsequent fault classification work,thus causing the "dimensional disaster".Therefore,how to eliminate the redundant features in the fault data set to achieve dimensional simplification of high-dimensional data has become a key fundamental problem that must be solved first to solve the data management problem of rotating machinery and scientifically accumulate massive data resources with high value density.Based on the above reasons,this thesis uses graph embedding and the hypergraph model as the theoretical basis to investigate the dimensionality reduction method of the rotor fault dataset,and the research progress achieved is summarized as follows.2(1)Aiming at the problem that traditional graph embedding methods cannot take into account the local and global structures of faulty data,a data dimensionality reduction algorithm based on Local-Global Standard Hypergraph Embedding(LGSHE)is proposed.The algorithm constructs a local intra-class and inter-class standard hypergraph and a global intra-class and inter-class standard hypergraph to characterize the data structure by redefining the computation of the hypergraph weight matrix,and then constructs an objective function based on the criterion of compressing intra-class information and separating inter-class information to achieve the purpose of extracting fault-sensitive features.The performance of LGSHE is verified on two rotor experimental benches with different structure types,and the results show that the algorithm can achieve good fault classification accuracy.(2)To better extract the low-dimensional features of the faulty dataset,and to comprehensively and accurately portray the intrinsic structure of the high-dimensional faulty dataset,a dimensionality reduction algorithm based on Multi-Structure Collaborative Discriminative Embedding(MSCDE)is proposed.The algorithm builds intra-and inter-class hypergraph structures to portray the higher-order relationships between data samples,simultaneously builds intra-and inter-class graph structures to enhance the direct connection between two samples,and finally constructs intra-and inter-class regularization structures to consider the global features of the faulty dataset.The intrinsic structure of the three structures can be effectively mined for fault information.The two-span rotor system is selected as the object,and the implementation process of the MSCDE-based fault identification method is sorted out.The experimental results demonstrate that the algorithm can effectively reduce the dimensionality of the rotor fault data and correctly identify the fault types based on the reduced,low-dimensional dataset.(3)Aiming at the problem of difficult fault classification due to high dimensionality and limited labeled samples in rotor fault datasets,an algorithm named Semi-supervised Hyper-Marginal Fisher Analysis(SHMFA)is proposed to improve the value density of fault datasets and make full use of unlabeled samples.Based on the idea of semi-supervised learning,the algorithm first constructs intra-class and inter-class hypergraphs by using the intra-class and inter-class nearest neighbor points of labeled samples,and then builds unsupervised eigenvalue and penalty hypergraphs by using the nearest neighbor and away points of unlabeled samples to achieve the comprehensive use of labeled and unlabeled samples,and then achieves dimensionality reduction of fault data sets.Simulation experimental results using a single-span rotor experimental bench show that the present algorithm has higher fault discrimination capability and good robustness compared with the conventional LDA and MFA.In the data-driven context,the dimensionality reduction algorithm based on the hypergraph model proposed in this thesis can provide a certain scientific basis for the exploitation and scientific accumulation of rotating machinery fault data resources.However,it is also worth exploring the issues such as parameter selection of the algorithm,adaptive feature extraction,and software development of fault diagnosis embedded in the algorithm.Solving the above problems can lay the relevant theoretical foundation for the practical application of intelligent fault identification technology.