From: Deep learning for component fault detection in electricity transmission lines
Study | Data source | Dataset size | ML approach (classifier) | Detection component classes | Pros and cons per our study |
---|---|---|---|---|---|
Zhao et al. [30] | Cropped image from Videos from manned Helicopter | > 1200 (256 × 256) | Pre-trained AlexNet using SVM with RBF kernel. (Caffe framework) | 2 | Required more feature engineering their model is limited as it does not include object detection of the faults |
Liu et al. [31] | UAV images | > 1500 (500 × 500) | Pre-trained faster RCNN (Caffe framework) | 1 | Focused only on a single type of insulators and its fault |
Jiang et al. [32] | UAV aerial images (China Power Grid) | 485 (300 × 300) | Ensemble learning using SSD meta-arch (TF OD API framework) | 1 | Again, only a single-type faults was focused on using ensemble learning model |
Tao et al. [33] | UAV captured images | 1956; 996 synthethic (4608 × 3456) | Pre-trained ILN + DDN (similar mechanism with faster RCNN) (MXNet framework) | 1 | Single type fault and small-scale problem using a two-phase object detection (OD) method. Hence, it requires greater pre-processing of the input images |
Bai et al. [35] | Picture manually from the industrial field | 139 (227 × 227) | AlexNet + SPP-Net | 1 | A pre-trained model was utilized to transfer information and save time and computational resources. However, the model is limited to the image classification problem |
Wang et al. [10] | Pictures achieved by industrial camera and mobile phone camera | 510 (280 × 328) | Pre-trained CNN + object detection | 1 | The unique method described in this work successfully detects the power lines and its faults. However, the study explores just a unary EPTN component with |
Guo et al. [37] | Helicopter captured images | 250 (1024 × 1024) | RCF + FRAC | 1 | The dataset utilized in the model had low spatial, spectral, and geometric resolution. Consequently, the predictive model was not developed to characteristically identify broken strand faults at varying perception /distance-depth level |
Chen et al. [38] | Images | 300 (400 × 600) | SOFCN | 1 | This research proposes a fault detection approach employing a second-order full convolutional network (SOFCN) to overcome low-level visual feature extraction, background complexity, and classifier design methodology. This has resulted in the model complexity. A drawback to this method is the SOFCN approach requires segmented ground truth data, which are hard to annotate and limits the amount of EPTN component fault types that was considered |
Liao et al. [39] | Aerial images | > 2000 (600 × 800) | Improved faster RCNN (Soft-NMS) keras | 1 | The proposed method achieves great result for insulator fault types and suitable for detecting insulator defect in complex backgrounds. As a two/phase OD method, major demerit to this method is that it is quite slower than the normal Faster R-CNN |
Li et al. [40] | UAV images | 800 | ROI mining Faster RCNN | 2 | The paper focuses on a unique fault type—bird’s nest obstruction detection. Also, the focus loss function is implemented in the RPN classification stage to balance the quantity of unbalanced data, foreground, and background data |