A novel hierarchical hyper-parameter search algorithm based on greedy strategy for wind turbine fault diagnosis

Abstract

With the rise of artificial intelligence, deep learning methods are more and more widely used in the field of intelligent fault diagnosis. However, the actual deep model used in fault diagnosis often exhibits over-fitting or under-fitting. In addition, the training process of these models requires configuration of a large number of hyper-parameters, and the selection of these hyper-parameters relies too much on experience. This makes the process of setting hyper-parameters quite tedious and time-consuming. Therefore, a new hyper-parameter search algorithm for network models is proposed, which is called Hierarchical Grid Scaling Hyper-Parameter Random Search (HGSHRS). The optimized model is verified with the fault simulation data of the wind turbines planetary gear system. First, different modal data are obtained by simulating different fault types and constructing the feature maps corresponding to them. Secondly, the CNN model is reused, and the existing pre-training model is used to accelerate the search for approximate optimal hyper-parameters and models. Finally, the proposed algorithm on the modal data of the planetary gear was verified and discuss the experimental results. Experiments prove that the deep learning model applying HGSHRS algorithm achieves relatively good results in fault diagnosis. The proposed method is of great significance for obtaining better hyper-parameters and models. Moreover, there are also considerable improvements in the reduction of operation and maintenance costs of wind turbine.