Abstract
During aero-engine production and design, optimizing engine performance frequently involves minimizing the gap between dynamic and static rotors, thereby elevating the likelihood of friction. Rubbing amplifies equipment vibration and, in severe instances, jeopardizes the entire shaft system, rendering it incapable of normal operation and incurring substantial economic losses. This paper presents a diagnostic method for identifying engine rotor rubbing faults using wavelet time-frequency images with a deep convolutional autoencoder neural network. The method converts original vibration signals into two-dimensional time-frequency images via wavelet decomposition, serving as input for a deep neural network. Leveraging the autoencoder’s characteristics with minimal dimensionality reduction parameters and the convolutional network’s robust feature extraction capability for two-dimensional data, the network is trained to extract crucial image features. Subsequently, these features undergo classification an d diagnosis of faults by training a Softmax classifier. An experimental system was established to validate the method’s feasibility, simulating fault phenomena such as misalignment, looseness, cracks, and rubbing by adjusting the fixing degree of the rotor system support seat. Test data are collected using the data acquisition system, which also preprocesses the raw data by introducing noise and conducting wavelet transforms. Subsequently, the data undergo transformation into time-frequency images, serving as inputs to the autoencoder neural network for extracting essential features. These features are inputted into this classification convolutional network, enabling fault diagnosis. The results could demonstrate the validity of the deep convolutional autoencoder neural network for detecting engine rotor rubbing faults.
