Research on wind turbine gearbox fault diagnosis with deep transfer learning method

        As the most complex component in the transmission system, the operating state of the wind turbine gearbox has a tremendous impact on the monitoring of the health status and operation control of the wind turbine equipment. Abnormalities in wind turbines that lead to downtime not only result in a loss of electrical energy, but also a significant increase in maintenance costs. Therefore, with the wind turbine gearbox as the main object of study, the following studies were carried out:
        For microscopic local conditions in gearbox gear systems, a method for obtaining  modal data using finite element simulation analysis of single tooth faults is proposed. Using a combination of deep auto-encoder structures and BP structures for secondary training strategies, a linear and non-linear performance evaluation method is proposed, which takes into account the relationship between performance and efficiency.
       Hyper-parameter configuration in deep transfer structures is often arbitrary, so a hierarchical transfer network structure hyper-parameter searching method is proposed to address the gearbox planetary system fault classification problem. The algorithm is validated using the classical LeNet-5 reconfiguration transfer application on a modal dataset of the planetary system. Finally, a stability validation and results analysis of the algorithm performance is carried out.
         A compressed sensing-based sparse signal decomposition method is proposed, and the structure of the transfer network is redesigned to achieve deep migration learning from rolling bearing faults to gear faults. A new network architecture was designed using a plug-and-play attention module. Pre-training models were designed and produced for fault data to improve the accuracy and recognition speed of fault diagnosis model classification. Finally, the effects of the same number of samples in the source and target domains and different distributions of sample features on the performance of the transfer learning method and the effects of hyper-parameters on the final performance of the network structure are verified.