Abstract:To solve the problem that it is difficult to extract effective fault features from nonlinear and unstable vibration signals,a fault feature extraction method was proposed based on improved adaptive noise complete set empirical mode decomposition (CEEMDAN) and t-distributed stochastic neighbor embedding (t-SNE) algorithm.Cubic Hermite interpolation was used to replace cubic spline interpolation to construct envelope,which could improve the decomposition accuracy of non-stationary signal by traditional CEEMDAN.The original signal was decomposed by the improved CEEMDAN,and the effective intrinsic mode components (IMF) were screened by correlation coefficients,and the time-frequency features,singular values and energy values of the effective IMF components were extracted to construct a high-dimensional mixed domain feature set.Finally,t-SNE algorithm was used to mine high-dimensional mixed domain feature information to obtain low-dimensional sensitive features,which were input into support vector machine for classification,and the classification accuracy rate was used as the evaluation index of feature extraction effect.The experimental verification on the gearbox fault simulation test bed shows that this method can accurately extract fault features,and provide a new idea for fault feature extraction.