High-voltage cables are crucial for constructing a safe and reliable urban power grid amid rapid urbanization. Damage to these cables can severely influence power transmission, potentially causing safety issues and economic losses. Maintaining high-voltage cables is challenging and costly, highlighting the need for efficient, non-destructive defect detection methods. Traditional methods, such as partial discharge detection, high-order harmonic analysis, and broadband impedance testing, struggle to accurately detect buffer layer ablation defects and locate specific defect positions. In contrast, computed tomography (CT) imaging provides a more intuitive visualization of defects and can quantify buffer layer ablation sizes from certain angles. However, conventional circular CT (RCT) techniques are unsuitable for detecting in-service high-voltage cables in confined spaces. In this study, we address the challenges of in-service cable detection by utilizing L-STCT technology combined with a deep learning-based method, using an improved Cascade R-CNN (region-convolutional neural network) to enhance the recall rate. The proposed method offers an effective solution for the non-destructive detection of internal cable defects.

We utilize L-STCT scanning to detect cable defects, with the SIRT algorithm used for image reconstruction. The resulting images are preprocessed to create an L-STCT dataset. To extract deeper features from the images, the ResNeXt101 with 64 filters is integrated into the Cascade R-CNN as the backbone for feature extraction, mitigating issues such as gradient vanishing and overfitting caused by excessive network depth. An attention mechanism is incorporated to help the network focus on defect-related information, improving its resistance to noise and artifacts. In addition, the EFPN module is introduced to enhance the detection of small targets while preserving other valuable information, enabling multi-scale feature extraction. The original position regression function is replaced with the Focal-EIoU loss function for more accurate localization, forming an optimized Cascade R-CNN. Although RCT cannot be directly applied to in-service cable detection, the similarity between RCT and L-STCT datasets allows for transfer learning; the network is pre-trained on the RCT dataset and then fine-tuned on the L-STCT dataset to further improve the recall rate of the network.

Ablation experiments confirm that the improved Cascade R-CNN network exhibits enhanced noise and artifact resistance with the introduction of the attention mechanism, while the EFPN module effectively identifies small defect structures. Compared to the original network, the optimized version shows significant improvements in accuracy and recall, demonstrating the algorithm’s suitability for cable defect detection (Table 3). The performance of the enhanced algorithm surpasses that of many mainstream target detection networks under the same dataset conditions (Table 4). The approach also offers advantages such as lower dataset and hardware requirements, making it highly practical. Transfer learning results indicate that pre-training the network on the RCT dataset improves its performance on the L-STCT dataset. Following transfer learning, the network achieves higher accuracy and recall rate comparable to those obtained with the original network (Table 5), confirming the effectiveness and applicability of the improved network.

In this study, we propose an enhanced cable defect detection algorithm based on the Cascade R-CNN, tailored to address challenges such as background noise and the detection of small targets. The algorithm performs well on the L-STCT dataset, achieving an accuracy of 0.884 and a recall rate of 0.927. With the RCT dataset pre-training, accuracy improves to 0.901, and recall reaches 0.959. The results demonstrate that while RCT cannot be directly applied for in-service cable defect detection, the similarities between the RCT and L-STCT datasets facilitate transfer learning, guiding the network to more effectively detect defects. The proposed algorithm offers a high defect recognition accuracy and a low miss rate, making it valuable for detecting defects in in-service cable buffer layers.