Design of transfer learning structure for slot wedge tightness inspection robot

Abstract

The tightness inspection for the slot wedges is significant for the safe operation of large generators. One of the traditional methods is analysis of the acoustic signals of knocking on the surface of the slot wedge by inspection experts. Nowadays the slot wedge inspecting robot is an effective way to measure the tightness of the slot wedges and classify the level of the slot wedges into different groups. However, there are many types of generators and the precision cannot be guaranteed if the model of one type of generators is applied to another. Although the machine learning methods such as CNN (Convolutional Neural Networks) and RNN (Recurrent Neural Networks) are widely used for classification, they are not suitable for model transfer between different generators. In this paper, a transfer learning based structure is introduced to solve the problem and also the mixture of RNN and CNN is designed to fulfill the system. The structure is tested to transfer models with the acoustic signal sampled by the inspecting robot between the 500 MW and 600 MW generators. Experiment results show that the transfer learning structure can transfer models from one type of generators to another. Compared with the state-of-the-art methods, the proposed structure can improve the inspection precision by at least 36.7% and obtain the average precision over 79.0%.