Accurately extracting deviation features in frequency response curves, which reflect winding deformation states, and selecting appropriate machine learning algorithms are critical for achieving a precise quantitative diagnosis of winding deformation based on frequency response analysis (FRA). To address the existing challenges in transformer winding fault diagnosis, including the absence of a systematic feature evaluation framework for frequency response data and the limited recognition accuracy of machine learning models, a novel hybrid feature selection and diagnostic framework was developed. First, a high-dimensional feature pool comprising 25 numerical indices was extracted from experimental FRA curves. To eliminate feature redundancy and arbitrary selection, a hybrid mechanism integrating maximum-relevance, minimum-redundancy (mRMR) with random forest (RF) was developed to dynamically construct task-specific optimal feature subsets. Furthermore, an inverse-distance-weighted K-nearest neighbors (IKNN) model was introduced to enhance diagnostic sensitivity by accounting for feature-space distance variations. Experimental results obtained from a laboratory winding model demonstrate that the proposed mRMR-RF-IKNN model significantly outperforms traditional and optimized benchmarks across multiple macro-evaluation metrics. This study provides a systematic, intelligent screening mechanism that ensures high-precision identification of both the types and severity of faults in power transformers.
Visible-light photovoltaic (PV) inspection is affected by periodic grid-line backgrounds, low-contrast cracks, irregular crack topology, and the extreme aspect ratios of slender defects. To address these difficulties, a frequency-domain hybrid deformation and geometry-regression YOLO network, termed FHDG-YOLO, is proposed in this study. The method introduces frequency-domain dynamic decoupled convolution to attenuate periodic background responses, incorporates a high-resolution P2 detection head and efficient multi-scale attention to retain and recalibrate shallow spatial details, embeds DCNv2 to adapt convolutional sampling to irregular defect boundaries, and replaces the original regression loss with MicroShape-IoU for geometry-sensitive localization. Experiments are conducted on a reorganized two-class visible-light PV dataset containing 6493 images, comprising 6262 screened public images and 231 field images collected by the authors. On the 1300-image validation split, FHDG-YOLO obtains a Precision of 0.954, Recall of 0.943, mAP@0.5 of 0.971, and mAP@0.5:0.95 of 0.861. Compared with YOLOv8n, mAP@0.5 and mAP@0.5:0.95 increase by 3.6 and 6.9 percentage points, respectively. On the held-out 649-image test split, the corresponding mAP values are 0.970 and 0.860, compared with 0.931 and 0.785 for YOLOv8n. Under the original four-class Panel Solar validation protocol, FHDG-YOLO obtains mAP@0.5 and mAP@0.5:0.95 values of 0.954 and 0.843, compared with 0.929 and 0.780 for YOLOv8n.