應用研究 2024年 第53卷 第21期
DOI: 10. 7672 / sgjs2024210011
基于 YOLOv5 的鋼結構節點損傷檢測研究
韓銘
作者簡介：
韓銘，碩士，E?mail:3432714399@ qq. com作者單位：
中國地震局工程力學研究所，黑龍江哈爾濱 150086基金項目：
?黑龍江省自然科學杰出青年基金（ JQ2022E006）；中國地震局工程力學研究所科研基金（2021B01,2021EEEVL0308）摘要：
以鋼結構節點損傷檢測為出發點，針對算法在個人困難數據集上的優化問題，使用預訓練權重，通過分析訓練過程中的損失趨勢評估合適的訓練周期。選擇 CBAM 注意力機制提升遷移學習的效率和性能，使用 AdamW優化器加快模型收斂速度，改善數據集的劃分策略以展現模型真實性能，提高模型的魯棒性，防止過擬合。根據先進算法理論優化了模型損失函數，提升模型在個人數據集上的精確率和召回率。針對問題復雜度與算法復雜度匹配性進行試驗，選擇最適合個人數據集的 YOLOv5n6 模型，最終優化出適合在現實場景中應用的鋼結構節點損傷檢測模型權重。
English：
Taking steel structure joints damage detection as the starting point, for the optimization problemof the algorithm on personal difficult datasets, using pre?training weights, evaluating the appropriatetraining period by analyzing the loss trend during the training process, selecting the CBAM attentionmechanism to improve the efficiency and performance of the migration learning, using the AdamWoptimizer to accelerate the convergence speed of the model, improving the dataset partitioning strategy toshow the real performance of the model, and improve the robustness of the model to prevent overfitting.The model loss function is optimized according to the theory of advanced algorithms to improve theaccuracy and recall of the model on the personal dataset. The tests were conducted for the problem of thematching between problem complexity and algorithm complexity, select the YOLOv5n6 model that is mostsuitable for the personal dataset, and ultimately optimize the model weights of steel structure jointsdamage detection,which is suitable to be applied in the real?world scenarios.