Fault identification of catenary dropper based on improved CapsNet

Abstract

Traditional fault identification algorithms applied to catenary dropper suffer from various problems due to its small contact area. These problems include misidentification and lower recognition rate of the faulty dropper. Compared with the traditional convolutional neural network, the vector is utilized as the input of the capsule network (CapsNet) for the first time, which can well retain the feature information such as the direction and angle of the target, and is more suitable for identifying the dropper under complex background. Therefore, this paper proposes a dropper fault identification algorithm based on improved capsule network. The convolutional layer of traditional 9 ×9 capsule network is simplified through 1 × 1 reduction layer and 3 × 3 convolutional layer, and the optimization algorithm is adopted for parameter optimization to shorten the training weight time. At the same time, the output can retain more information such as direction and angle, which can accurately identify the breakage and falling of current carrying broken. Thus, in order to better improve the accuracy and real-time of detecting the fault dropper from a running train operation, a dropper fault identification algorithm based on an improved CapsNet is proposed in this paper. Experimental results show that the improved CapsNet is well-suited for fault identification of catenary dropper, as it can effectively remove the interference caused by the complex background on the dropper image, and identify the image containing the faulty dropper with a higher recognition rate.