小样本离散时序数据的集成学习特征提取方法OA

Accurately extracting features from high-dimensional,small-sample,time-series data re-mains a critical challenge in the health assessment of high-precision equipment,which typically exhibits multi-component configurations,multi-mode degradation behaviors,and long inspection intervals.To ad-dress this problem,this paper proposes a feature extraction method that integrates an improved Mann-Kendall(MK)test with kernel principal component analysis(KPCA).After standardizing the raw time-series measurements,we first construct a sliding-window MK trend analysis model coupled with a weigh-ted Euclidean distance deviation analysis module.This design overcomes the inherent limitations of the conventional MK algorithm,namely,its inability to recognize multi-mode degradation patterns and its lack of redundancy filtering capability—thereby enabling precise identification of both monotonic trend-type and variance-increment-type degradation features while efficiently eliminating redundant variables.Sub-sequently,to circumvent the arbitrary selection of kernel parameters and the loss of coupling information inherent in direct dimensionality reduction using standard KPCA,we develop an improved KPCA frame-work with adaptive kernel parameter optimization and a two-stage dimensionality reduction strategy.This framework builds a nonlinear coupling model based on the selected core degradation features,and finally fuses them into a one-dimensional health indicator(HI).Experimental results demonstrate that the pro-posed method significantly outperforms traditional methods in terms of adaptability to small-sample sce-narios,feature selection accuracy,and nonlinear coupling modeling capability,providing effective technical support for health condition assessment of complex electromechanical equipment.