Real-Time Pantograph Anomaly Detection Using Unsupervised Deep Learning and K-Nearest Neighbor Classification

An image-based real-time pantograph anomaly detection method is presented by combining unsupervised deep learning and nearest neighbor classification. The proposed method includes the following key steps. First, an improved DeblurGAN-v2 deblurring algorithm is applied to the input pantograph image if there exists motion blur. Next, deep learning semantic segmentation with hybrid coding that combines lightweight convolutional neural network (CNN) and vision transformer (VIT) is employed to accurately segment the pantograph structure within the image. And multiscale feature-dense aggregation network based on an attentional feature fusion (AFF) module has been designed to efficiently integrate information from different feature layers. Finally, a $K$ -nearest neighbor (KNN) classification algorithm with deep pretrained features from the segmented pantograph mask image has been utilized to detect anomalies in the pantograph. Experimental results demonstrate that the proposed pantograph segmentation network outperforms several general segmentation algorithms, achieving a high mean intersection over union (MIoU) of 95.86% with a parameter size of 7 M and FPS of 81.7. And nearest neighbor classification with deep pretrained features achieves excellent pantograph anomaly detection performance with area under the receiver operating characteristic (ROC) curve of 0.987 and area under a precision-recall (PR) curve of 0.998. It is verified that the proposed pantograph anomaly detection method does not rely on abnormal data, and can achieve a high anomaly detection accuracy of 98.75%.