Identification of Abnormal Electrical Phenomena in Train-Grid System in Open Set Environment

Rapid and accurate identification of abnormal electrical phenomena (AEPs) in electrified railway plays a vital role in the safe and smooth operation of electric locomotives. In the actual electrified railway operation scenario, due to the randomness and uncertainty of the occurrence of AEP, there are also unknown AEP, which also endanger the safe operation of the electrified railway and the power supply quality of the traction power supply system. Most of the traditional methods for the identification of AEP are based on the closed-set assumption, i.e., the classification results can only be selected from the given known categories and cannot identify unknown electrical abnormalities. Therefore, this article proposes an openmax-based method for identifying AEP, which can identify unknown AEP along with normal data and typical known AEP correctly. The proposed method mainly includes data preprocessing, 1-D convolutional neural network (CNN), and openmax. In addition, to further improve the recognition accuracy (Acc) of the proposed algorithm, different hyperparameters are set for different known classes in the tail fitting process and an activation function is applied to the activation vector (AV) used to calculate openmax. The experimental results show that the method can identify typical AEP with 98.5% Acc and unknown AEP with 91.7% Acc.