Multi-layer and multi-scale feature fusion deep image segmentation network: an advanced approach for hot-spot defect segmentation in photovoltaic modules

With the rapid development of the photovoltaic (PV) power generation industry, the operation and maintenance (O&M) issues of PV modules have attracted increasing attention. Among these, hot-spot faults stand out as a critical concern due to their adverse effects on power generation efficiency and safety. Traditional object detection methods can identify hotspots in PV modules. However, they exhibit limitations in detecting small target edges and accurately segmenting multi-scale hotspots. To address this issue, this paper proposes a deep image segmentation network based on multi-layer and multi-scale feature fusion (M−MDNet). The model incorporates a multi-layer and multi-scale window attention feature extraction network (M−MWA), introduces an enhanced atrous spatial pyramid pooling module (ASPP*), and integrates a mixed-dimensional Spatial and Channel Squeeze-and-Excitation attention mechanism (scSE). Experimental results demonstrate that the proposed framework significantly enhances both boundary delineation capability and segmentation precision for photovoltaic module hotspots. Experiments were conducted on self-constructed datasets from an agrivoltaic power station in Hubei Province (HB_Data) and a floating photovoltaic power station in Jiangsu Province (JS_Data). The results indicate that the proposed M−MDNet outperforms mainstream segmentation models, achieving a mean Intersection over Union (mIoU) of 90.35% and an overall accuracy of 97.61%. Furthermore, the proposed model exhibits a reduced parameter count and computational complexity, thereby improving operational efficiency while maintaining high segmentation accuracy.