基于端到端深度学习的大坝渗漏电性异常识别研究

Research on the identification of electrical anomalies of dam leakage based on end-to-end deep learning

【目的】电法勘探中传统的电性异常判识方法计算复杂、效率低,且成果解译对初始模型依赖性强,难以满足海量数据快速处理的应用需求。【方法】对此,结合深度学习思想,提出了一种基于端到端的大坝视电阻率识别网络(Apparent Resistivity Network, ARNet)模型,将传统的渗漏判识方法转换为从输入视电阻率数据到输出异常体分布的非线性映射问题。利用pyGIMLi有限元工具建立2.2×104个不同形状、位置及电阻率值的大坝渗漏地电模型,计算视电阻率形态分布;通过对模拟数据进行批量提取、插值网格化处理,构建了网络训练数据集;采用相应的前端、后端及后处理方法完成网络模型搭建,并以“损失值”和“交并比”作为模型性能指标,采用随机梯度下降算法实现模型权重参数的迭代优化。【结果】结果表明:ARNet模型经过1 000次训练后,损失值降至0.068,交并比达到94.60%;在不同电阻率、位置、形状的单、双异常体对比测试中,模型匹配度达到97.66%,误差低于0.257 m。【结论】结合水库大坝的实测数据试验,ARNet模型具备良好的泛化性能,较传统反演可实现对异常体的高精度智能识别,研究成果拓宽了深度学习技术在大坝安全领域的应用研究。

[Objective]The traditional method for identifying electrical anomalies in electrical prospecting are computationally complex,inefficient,and highly reliant on the initial model for result interpretation,which makes it challenging to meet the application demands of rapid processing of massive detection data. [Methods]In this regard, combined with the idea of deep learn-ing, an end-to-end apparent resistivity identification network (Apparent Resistivity Network, ARNet) model for dams was pro-posed, converting the traditional leakage identification method into a problem of nonlinear mapping from the input apparent resis-tivity data to the output distribution of anomalous bodies. The pyGIMLi finite element tool was used to establish 2. 2×104 dam leakage geoelectric models with different shapes, locations and resistivity values, allowing for the computation of the distribution of apparent resistivity forms. A network training dataset was constructed by performing batch extraction and interpolation grid pro-cessing on simulated data. The network model construction was completed using network front-end, back-end, and post-process-ing techniques. The loss value and intersection over union were employed as model performance indicators, and iterative optimi-zation of model weight parameters was carried out using a stochastic gradient descent algorithm. [Results]The result showed that after 1000 epochs of training, the ARNet model achieved a loss value reduction to 0. 068 and an intersection over union of 94. 60%. During comparative testing involving single and double anomalous bodies with varying resistivity, positions, and shapes, the model demonstrated a matching accuracy of 97. 66% and an error below 0. 257 m. [Conclusion]Combining experi-mental data from reservoir dam measurements, the ARNet model demonstrates excellent generalization performance. In compari-son to traditional inversion method , it achieves highly accurate intelligent identification of anomalies. The research outcomes ex-pand the application of deep learning technology in dam safety.