Fault Diagnosis Based On Improved Information Entropy And 1dcnn For Marine Turbocharger Rotor With Variable Speed

Abstract Due to the lack of a standardised method for determining representative locations for measuring points, it is difficult to select sensitive data on turbocharger rotor faults. In addition, the uncertainty in the feature parameters used for diagnosis under variable rotational speeds leads to low accuracy in fault identification. To address these issues, vibration signals from a turbocharger rotor under various conditions are obtained in this study via a fault simulation test, and a fault diagnosis method for rotor faults under variable speed conditions is proposed based on a sensitivity and multi-dimensional feature analysis of measurement points. The sensitivity of the improved and traditional information entropy is evaluated using a variance analysis across the measurement points, and the most effective vibration measurement points are selected based on the improved information entropy. The effective characteristic parameters of the vibration signals at multiple measurement points are analysed and extracted, and the number of dimensions of the feature parameters is reduced using the t-distributed stochastic neighbour embedding (t-SNE) method. Faults in the turbocharger rotor at the different speeds are classified using a one-dimensional convolutional neural network (1DCNN), and the arithmetic ability of the diagnostic algorithm is evaluated. The results demonstrate that the proposed methods of selecting measurement points and fault diagnosis can effectively identify rotor faults at different degrees and various speeds: the accuracy of fault diagnosis is 99.85%, and the arithmetic ability is markedly enhanced compared with that of traditional methods.