Fault diagnosis method for OLTC based on improved semi-supervised ladder networks

On-load tap changers (OLTC) have complex mechanical and electrical structures, which are the key component for the on-load voltage regulation of transformers. Currently, due to the sample data which are not easy to be labeled, it is difficult to effectively train the OLTC mechanical fault diagnosis models based on vibration signals. To improve the fault diagnostic accuracy for OLTC with limited labeled data, a fault diagnosis method based on Bayesian optimization-convolutional ladder networks (BO-ConvLN) is proposed in this paper. Firstly, the ladder networks are used as a semi-supervised learning method for the feature extraction of vibration signals, which is guided by a large amount of unlabeled data, leading to the enhanced diagnostic ability of ladder networks only with a small amount of labeled data. Then, the fully-connected layers are replaced by convolutional operators in the ladder networks to better extract the features of non-stationary vibration signals. Furthermore, Bayesian optimization is used to optimize the high-dimensional hyperparameters of ladder networks, witch significantly improves the diagnostic accuracy of the model within limited time cost. The experiment results show that the diagnostic accuracy for the three types of faults, namely transmission shaft jams, poor switch lubrication, and top cover looseness, is 91.67% with a label count of only 40, which demonstrates the effectiveness of BO-ConvLN in the fault diagnosis.