针对当前变电站电力设备缺陷类型复杂、目标尺度多变以及背景干扰的问题，提出一种面向变电站设备缺陷检测的深度特征感知方法。首先设计了一种基于滑动窗口机制的C2f+模块，使得主干网络能够有效地处理跨窗口特征信息，并扩大网络的感受野，加强网络对局部特征的深度关注和全局特征处理能力。然后，整合SimAM到网络的特征融合阶段，赋予网络自主学习和动态微调注意力权重的能力，以提高对多尺度缺陷的感知性能。其次，为避免不同尺寸和形状的设备缺陷引入的检测误差，设计了一种融合NWD和CIOU的损失函数，通过对预测框和真实框的二维高斯分布建模，提高模型对缺陷位置的学习能力。在变电站设备缺陷数据集上的实验结果表明，所提方法对11种缺陷类别的平均检测精度mAP达到了98.27 %，且在9种缺陷检测中的实现了最佳精度，优于其他6种主流的目标检测方法和电力设备缺陷检测方法。同时，FPS达到了78.4，具有较高的实时性能。

To address the complexity of defect types, variability in target scales, and background interference in power equipment within substations, a deep feature perception method for substation equipment defect detection is proposed. Firstly, a C2f+ module based on a sliding window mechanism is designed to enable the backbone network to effectively handle cross-window feature information, expand the network′s receptive field, and strengthen the network′s deep attention to local features and global feature processing capabilities. Next, SimAM is integrated into the feature fusion stage of the network, empowering the network with the ability to autonomously learn and dynamically fine-tune attention weights to improve perceptual performance for multi-scale defects. Additionally, to mitigate detection errors introduced by defects of varying sizes and shapes, a fusion loss function combining NWD and CIOU is designed. By modeling the predicted and actual bounding boxes as two-dimensional Gaussian distributions, this enhances the model′s learning ability for defect localization. Experimental results on a dataset of substation equipment defects demonstrate that the proposed model achieves an average detection accuracy （mAP） of 98.27 % across 11 defect categories. It also outperforms six mainstream target detection methods and power equipment defect detection methods with the best accuracy in 9 out of 11 defect detections. At the same time, the model achieves a frame rate （FPS） of 78.4, demonstrating high real-time performance.