Abstract: Centrifugal compressors, steam turbines, and flue gas turbines, among other large rotating machinery, are core power equipment in the petrochemical enterprises. Intelligent diagnosis of common typical equipment faults is crucial for carrying out intelligent operation and maintenance. To address challenges such as the difficulty in extracting early-stage weak faults, poor noise interference resistance, confusion of signal features across different fault states, and low learning performance of common features in cross-condition data, the refined composite zoom multi-scale weighted permutation entropy index has been constructed, which is effective for capturing subtle oscillation patterns across the full frequency band. A method for cross-domain fault transfer diagnosis in rotating machinery is also proposed. Firstly, the method is initiated with the decomposition, filtering, and reconstruction of raw vibration signals from a multi-source typical fault database to extract their sensitive common features. Subsequently, the multi-kernel twin ensemble feature learning strategy is employed to iteratively enhance the model's feature classification performance for five types of source domain faults: rotor unbalance, shaft misalignment, static and dynamic rubbing, oil film whirl, and surge. Then, the semi-supervised manifold feature transfer strategy is used to minimize the differences in feature distributions between the target and source domains, with the strong classifier mapping and matching fault category labels. Finally, the effectiveness of the proposed method is validated using real engineering fault case data and compared against five published entropy features and six fault diagnosis methods from the references, demonstrating superior diagnostic performance of the proposed method under multiple operating conditions for different equipment.

Key words: refined composite zoom multiscale weighted permutation entropy, rotating machinery, cross-domain fault diagnosis, multi-kernel twin ensemble feature learning, semi-supervised manifold feature transfer