Multi-objective Defect Detection of Substation Equipment Based on SA-YOLOv7 Algorithm

To address the challenges associated with multi-object defect detection in substation equipment–including widespread defect distribution, intricate substation backgrounds, and challenges in feature extraction–a new SA-YOLOv7 (Simple and Accurate–You Only Look Once version 7) algorithm has been developed. Initially, the Ghost Convolution technique is introduced to replace the standard convolution in the Spatial Pyramid Pooling Cross Stage Partial Connection Networks (SPPCSPC), thus minimizing computational overhead and streamlining model complexity. Thereafter, an integration of the Weighted Bi-directional Feature Pyramid Network (BiFPN) with the Path Aggregation Network (PANet) is implemented. This integration ensures a balanced representation of feature data across different scales during the feature fusion phase. The bounding box regression loss function in YOLOv7 is also refined, using the Wise-IoU (WIoU) approach. This shift heightens the model’s attention to low-quality anchor boxes and improves its target localization. A Simple Parameter-Free Attention Module (SimAM) is further incorporated into both the Efficient Layer Aggregation Networks (ELAN) and the neck network’s output. This module endows the model with 3-D attention weights, bolstering its feature extraction capacity and focus on target entities. Empirical data show that SAYOLOv7 boasts a mean average precision of 93.82% on the substation equipment’s multi-object test set. This performance reflects a 1.96% improvement over the standard YOLOv7 and a parameter decrease of 9.93%. Notably, SAYOLOv7’s detection capability outperforms several contemporary target detection algorithms, signifying a substantial enhancement in the intelligent monitoring of substation apparatus.