0
0
基于DLS-Net深度學習框架的太陽能板表面灰塵檢測方法
首發(fā)時間:2025-12-16
朱佳鑫,男,研究方向為深度學習,圖像處理
劉婧 1劉婧,女,講師
摘要:太陽能電池板表面灰塵積累會嚴重影響光伏發(fā)電效率,準確檢測灰塵污染程度對于維護運營至關(guān)重要。針對傳統(tǒng)卷積神經(jīng)網(wǎng)絡(luò)在灰塵檢測中存在的梯度衰減、參數(shù)冗余以及細微特征提取能力不足等問題,本研究提出了一種基于密集連接網(wǎng)絡(luò)的多層次注意力增強模型DLS-Net(DenseNet-based LPA-SSA Network)。該模型以DenseNet-121為基礎(chǔ)特征提取框架,通過密集連接機制實現(xiàn)特征復(fù)用和梯度有效傳播;引入局部金字塔注意力模塊(LPA)實現(xiàn)多尺度空間特征自適應(yīng)融合,增強對不同面積灰塵覆蓋區(qū)域的感知能力;進一步集成序列洗牌注意力模塊(SSA)實現(xiàn)跨組通道信息交互,深度挖掘特征間的互補關(guān)系。在包含1187幅太陽能電池板圖像的數(shù)據(jù)集上進行的實驗表明,DLS-Net的準確率達到98.90%,相比基準DenseNet-121模型提升了2.68個百分點,F1分數(shù)達到97.62%。消融實驗驗證了各模塊的有效性,Grad-CAM可視化結(jié)果表明模型能夠準確定位灰塵特征關(guān)鍵區(qū)域。本研究為太陽能電池板智能運維提供了高效可靠的檢測方案。
關(guān)鍵詞: 神經(jīng)網(wǎng)絡(luò) 深度學習 太陽能電池板 灰塵檢測
For information in English, please click here
Solar panel surface dust detection method based on DLS-Net deep learning framework
朱佳鑫,男,研究方向為深度學習,圖像處理
Liu Jing 1劉婧,女,講師
Abstract:The accumulation of dust on solar panel surfaces severely affects photovoltaic power generation efficiency, making accurate detection of dust contamination levels crucial for maintenance and operation. To address issues in traditional Convolutional Neural Networks (CNNs) for dust detection, such as gradient degradation, parameter redundancy, and insufficient fine-grained feature extraction capability, this study proposes a multi-level attention-enhanced model based on dense convolutional networks, termed DLS-Net (DenseNet-based LPA-SSA Network). The model employs DenseNet-121 as the foundational feature extraction framework, achieving feature reuse and effective gradient propagation through dense connectivity mechanisms. It incorporates a Local Pyramid Attention (LPA) module to enable adaptive fusion of multi-scale spatial features, enhancing the perception capability for dust coverage areas of varying sizes. Furthermore, it integrates a Sequence Shuffle Attention (SSA) module to facilitate cross-group channel information interaction and deeply exploit complementary relationships among features. Experiments conducted on a dataset containing 1,187 solar panel images demonstrate that DLS-Net achieves an accuracy of 98.90%, representing a 2.68 percentage point improvement over the baseline DenseNet-121 model, with an F1 score of 97.62%. Ablation experiments validate the effectiveness of each module, and Grad-CAM visualization results indicate that the model can accurately localize key regions of dust features. This research provides an efficient and reliable detection solution for intelligent operation and maintenance of solar panels.
Keywords: Neural Network Deep learning Dust detection Solar panels
基金:
引用

No.****
動態(tài)公開評議
共計0人參與
勘誤表
基于DLS-Net深度學習框架的太陽能板表面灰塵檢測方法
評論
全部評論