Hyperparameter optimization results.

Inspection and diagnosis of construction projects involves health monitoring of buildings and related facilities, and the utilization of renewable energy sources, such as solar energy, is critical to the smooth operation of modern construction projects. The detection of solar cell defects is related to the reliability and efficiency of building photovoltaics and has become an area of interest. Existing deep learning-based solar cell defect detection models significantly improve the accuracy of solar cell defect detection, however, deep learning-based solar cell defect detection models ignore the effect of network hyperparameters on their model performance. In this study, the hybrid model CMNS-YOLO, which combines the crawfish optimization algorithm with the MNS-YOLO model, is proposed to achieve the ultimate detection accuracy. First, Mamba-Like Linear Attention is introduced to design the C2f-MLLA module to improve the target feature representation capability of solar cell sheet defects; second, Bidirectional feature pyramid frequency aware feature fusion network is designed to enhance the recovery ability of target detail features as well as the fusion ability of image features; then ShapeIoU is used to solve the target aspect ratio misalignment problem and construct the improved MNS-YOLO network; finally, COA is utilized to adjust the parameters of the MNS-YOLO network. Experimental results on the PV-Multi-Defect and PVELAD datasets show that compared with the baseline model, the detection accuracy of the proposed model on the two datasets is improved by 6.3% and 2.3% while maintaining the lightweight characteristics of the model. Therefore, the proposed method has considerable potential in the field of solar cell defect detection.

建筑工程巡检与诊断涵盖建筑及附属设施的健康监测，而太阳能等可再生能源的利用对于现代建筑工程的平稳运行至关重要。太阳能电池缺陷检测与建筑光伏系统的可靠性及运行效率息息相关，已然成为研究热点。现有基于深度学习的太阳能电池缺陷检测模型虽显著提升了检测精度，但这类模型往往忽略了网络超参数对模型性能的影响。本研究提出了融合小龙虾优化算法（crawfish optimization algorithm, COA）与MNS-YOLO模型的混合模型CMNS-YOLO，以实现最优检测精度。首先，引入类Mamba线性注意力（Mamba-Like Linear Attention）机制设计C2f-MLLA模块，以提升太阳能电池片缺陷的目标特征表征能力；其次，设计双向特征金字塔频率感知特征融合网络（Bidirectional feature pyramid frequency aware feature fusion network），以增强目标细节特征的恢复能力与图像特征的融合能力；随后，采用ShapeIoU解决目标长宽比失配问题，构建改进后的MNS-YOLO网络；最后，利用COA对MNS-YOLO网络的参数进行调优。在PV-Multi-Defect与PVELAD数据集上的实验结果表明，相较于基线模型，所提模型在这两个数据集上的检测精度分别提升了6.3%与2.3%，同时保留了模型的轻量特性。综上，所提方法在太阳能电池缺陷检测领域具备可观的应用潜力。