基于Keras框架的油气管道缺陷分类应用研究
Application Research on Defect Classification of Oil and Gas Pipelines Based on Keras Framework

针对油气管道缺陷现代化检测问题，本文提出了一基于Keras框架的深度学习检测算法，采用Adam-S优化器算法应用到卷积神经网络，并融合了Environment-Cognition-Action (ECA)-Effi注意力机制，增强了卷积神经网络的特征表达能力，提高了有效验证数据的利用效率，并通过对比实验，分析比较了影响深度学习过拟合现象高度相关的学习率、Dropout两个参数，提取最优权重参数矩阵，建立验证集并导入到预测模型中，得到油气管道6种缺陷特征的预测结果，且预测结果与真实标签高度接近，验证了所搭建神经网络的表现及其有效性，最后，量化对比了目前应用较为普遍的ResNet18网络以及改进前的Adam-CNN算法，有效验证了Adam-S结合改进ECA-CNN算法的优越性能。
This paper proposes a deep learning detection algorithm based on the Keras framework to address the issue of modern detection of oil and gas pipeline defects. The Adam-S optimizer algorithm is applied to convolutional neural networks, and the Environment Recognition Action (ECA) Effi attention mechanism is integrated to enhance the feature expression ability of convolutional neural networks and improve the utilization efficiency of effective validation data. Finally, through comparative experiments, We analyzed and compared the learning rate and Dropout parameters that are highly correlated with the overfitting phenomenon in deep learning, extracted the optimal weight parameter matrix, established a validation set, and imported it into the prediction model to obtain the prediction results of six types of defect features in oil and gas pipelines. The predicted results were highly close to the real labels, verifying the performance and effectiveness of the constructed neural network.