SE-Inception-ResNet Model With Focal Loss for Transmission Line Fault Classification Under Class Imbalance

Accurately detecting and classifying faults are essential for fault clearance and recovery of the system. One of the biggest challenges in fault diagnosis is dealing with fault class imbalances, where some fault types occur less frequently than others. The class imbalance makes the classifier prefer the majority class, which will cause the deterioration of classification performance. To solve the problem of serious class imbalances in fault natures, fault causes, and fault phases for the actual power system, this article presents a transmission line fault classification method based on the squeeze and excitation (SE)-Inception-residual network (ResNet) model with focal loss (FL). In this method, the FL is introduced to focus more on hard-to-classify faults by reducing the weight of easy-to-classify faults automatically. Also, the proposed model integrates the SE-Inception module and the SE-ResNet module, which can exploit the advantages of each to achieve better classification accuracy along with good computational efficiency due to the lighter network structure. To verify the effectiveness of the proposed method, various types of faults with unequal numbers are generated based on the 735-kV three-phase transmission line model, and three-phase signals are converted into 2-D time-frequency maps using continuous wavelet transform (CWT) for extraction of the deeper and higher level features. Experimental results show that the proposed method achieves 98.5% classification accuracy and above 92% F1-score, which indicates the improvement of classification performance in the case of serious class imbalance. Also, the effectiveness of this proposed method is verified by the real recording fault data. In addition, the comparisons with other convolutional neural networks (CNNs) and the cross-entropy (CE) demonstrate its superiority. The interpretability of the proposed networks for fault recognition is provided based on the gradient-weighted class activation mapping (Grad-CAM) visualization, and it reveals the inherent mechanism of its good classification performance from time-frequency features.