基于精细复合多尺度分数阶注意熵的汽轮机振动故障智能诊断方法OA

[Objective]To address the issues of frequent vibration faults in steam turbines under deep peak shaving and frequent variable operating conditions,as well as the difficulty of traditional single-scale entropy feature extraction methods in effectively processing nonlinear and non-stationary signals,an intelligent reduced-order fault diagnosis method based on refined composite multiscale fractional attention entropy is proposed.[Methods]First,fractional calculus is introduced into the refined composite multiscale attention entropy framework to construct the RCMFAE algorithm,which extracts multiscale fractional attention entropy features from vibration signals.Subsequently,t-distributed stochastic neighbor embedding(t-SNE)is employed for dimensionality reduction and visualization of the high-dimensional feature set.Finally,a support vector machine(SVM)model optimized by a particle swarm optimization and genetic algorithm joint optimization(GAPSO)algorithm is used for fault classification and identification based on the reduced-dimensional feature set.Vibration data are collected from the ZT-3 rotor test rig to simulate four typical states-normal operation,unbalance,misalignment,and rub-impact-and the proposed method is experimentally validated.[Results]Experimental results show that RCMFAE maintains clear inter-class separation across the full scale,achieving a Fisher score of 5.74-up to 13.66％higher than comparative methods.After t-SNE dimensionality reduction,various fault states exhibit distinct clustering characteristics in three-dimensional space.In the fault classification stage,the RCMFAE-GAPSO-SVM model achieves an average diagnostic accuracy of 98.94％,reaching up to 100％,outperforming comparison methods such as AE,MAE,CMAE,AMDE,and MFSDE.In noise robustness tests,RCMFAE maintains accuracy above 95％under signal-to-noise ratios ranging from-5 dB to 10 dB,demonstrating strong robustness.Additionally,the average computational time of the RCMFAE algorithm is approximately 5.59 seconds,showing superior efficiency compared to CMAE and AMDE.[Conclusion]The proposed method integrates fractional attention entropy with multiscale refined composite analysis,enabling simultaneous capture of both local and global features of vibration signals,effectively characterizing long-range dependencies,and significantly improving fault feature discrimination and diagnostic accuracy.