Under conditions of long-term operation of high-power conductive slip rings with high current and strong electromagnetic interference, and rotational contact coupling, the acoustic radiation characteristics of the housing may be altered due to bearing ball/raceway damage, spindle fatigue cracks, dynamic eccentricity, and brush jumping. Given the problems of high dimensionality, strong redundancy and a small sample size in the multi-scale candidate features obtained through empirical mode decomposition of the original audio signal, this paper enhances the Lasso feature selection method based on 60kW fault simulation platform data and introduces a segmented adaptive Lasso feature selection strategy. Lasso has sparse filtering properties, so a corresponding way was proposed to reduce the bias in coefficients resulting from uniform penalization; this method progressively performs filtering on groups of features to reduce the dimension of the single regression solution while retaining the traceability relationship among feature number, selected variables, and classification results. Five types of operational audio data were obtained from a 60kW high-power conductive slip ring fault simulation platform as the objects, and a 45-dimensional feature set of the original signal and IMF1-IMF8 was constructed; ten-fold cross-validation was performed on KNN, MLP and SVM classifiers. The results show that the segmented adaptive Lasso can provide good classification results in all five states, with an average accuracy of 0.873 when n=10; Compared with Lasso, its accuracy increased by 0.034, 0.010, 0.040 and 0.075 under normal, spindle fatigue crack, dynamic eccentricity and brush jumping conditions, respectively, and the average calculation time decreased by 38.61%. Improve the generalisation ability and real-time performance under high-dimensional small-sample conditions while keeping the interpretability of acoustic features.
