Abstract: The droppers are the main components of the high-speed railway catenary system. They are prone to fracture and relaxation, which directly threatens the safety of driving. Under the action of fluctuating wind and pantograph simultaneously, the acceleration signals obtained by the acceleration sensors installed on the messenger wire and contact line are relatively strong. Using the LSTM network model, it is easy to predict the fracture and relaxation of the droppers. In this paper, the strong feature extraction ability of the convolutional neural network and the time sequence expression ability of recurrent neural network are used to solve the problem of the detection of dropper fault under the action of fluctuating wind only. At the same time, the attention mechanism is introduced to establish the fusion network model of CNN-LSTM-Attention. In the process of network training, the Bayesian optimisation method is used to select hyperparameter. Compared with the LSTM model, the fusion network model improves the detection accuracy of dropper fault and has strong practicability.

Key words: catenary dropper, fault detection, convolutional neural networks, LSTM, attention mechanism