No abstract

At present, the trend of complex and intelligent rotating machinery and equipment is becoming more and more obvious, which generates a large amount of high-dimensional and nonlinear fault monitoring data that is difficult to handle. This makes the traditional dimensionality reduction algorithms based on point-to-point metrics or a small number of graph embedding structures lose their utility. To solve this problem, a multiple feature-spaces collaborative discriminative projection (MFSCDP) algorithm for rotor fault dataset dimensionality reduction is proposed. The algorithm first improves the projection metric from sample point to feature space into the median metric in order to achieve the effect of weakening the extrapolation error of the algorithm, and based on this, we propose a sample point-to-point guided nearest-neighbor feature space selection method to improve the construction efficiency of the feature space embedding graph. Then, by using Relief F to indirectly construct the reduced dimensional projection matrix with multiple feature spaces of collaboration. Finally, the proposed MFSCDP algorithm is used for the dimensionality reduction process of the rotor fault dataset. The algorithm's performance was verified using experimental information from rotor failure simulations of two different structural types. The result shows that the algorithm can reduce the difficulty of fault classification and improve the accuracy of identification.

No abstract

The vibration signal contains a lot of state information, and its time domain, frequency domain, and time-frequency domain features are extracted to describe the rotor’s operational state in detail. However, multi-domain and multi-angle feature extraction can lead to information redundancy, causing the "dimensionality catastrophe" problem, which also makes fault classification more difficult. Dimensionality reduction (DR) is a technique that can help with this issue, but the majority of current DR algorithms are based on a simple intrinsic structure. To accurately characterize the intrinsic structure of high-dimensional fault dataset, this study proposed a novel DR algorithm that considers local information, global information, and the hypergraph model, named Local-Global Standard Hypergraph Embedding (LGSHE). LGSHE redefines the hypergraph weight matrix’s calculation formula and constructs the local intra-class standard hypergraph, local inter-class standard hypergraph, global intra-class standard hypergraph, and global inter-class standard hypergraph to characterize the fault data structure. LGSHE can accurately characterize the intrinsic structure of high-dimensional fault dataset and increase fault classification accuracy. The performance of LGSHE is validated on two different structures of double-span rotor experimental benches, and the results show that the algorithm can achieve good fault classification accuracy.

No abstract