| 摘要: | 由於現有的電阻率成像流程，每次成像皆需要耗費較長時間，並且容易受到人為操作影響，所以不易施行在自動化系統上。上述問題前人嘗試使用神經網路來解決。然而，前人所使用的神經網路相對於現今的深度神經網路，較難以完成困難的任務，且前人亦沒有釋出可供參考的套件，令往後的研究均要重複編寫相對應的程式碼。有鑑於此，本研究的重點在於應用深度學習技術，開發一套電阻率成像套件。本套件的正演模擬部分，是使用Pidlisecky & Knight(2008)開發的電位模擬程式；神經網路部分則建立在以Tensorflow為後端的Keras之上。
我們分別使用合成資料及野外資料，對本研究建立之成像流程進行測試。合成資料部分，我們設計五種不同配置的訓練資料，分別訓練出五個神經網路，並採用層狀、雙塊體及多塊體的電阻率合成模型，評估神經網路的預測能力，藉此觀察不同操作所帶來的效應。此外，我們將EarthImager 2D反演之電阻率剖面與神經網路預測之電阻率剖面進行比較，以找出這些剖面之間的差異；野外資料部分，我們採用2015年10月10日至10月21日，永康場址南北向測線所量測的資料。永康場址於2015年10月13日至10月18日，曾進行整治藥劑的灌注試驗。灌注試驗造成的電阻率變化能否表現在電阻率剖面上，或可作為評價本研究之成像流程的指標。對於野外資料的測試，我們設計兩組訓練資料，一組基於常見地質材料進行設計，另一組則參考岩芯資料進行設計。最終結果顯示基於先驗資訊所訓練出的神經網路，對於本場址具有較良好的預測能力。然而，野外資料的品質不佳，致使運用傳統反演也難以獲得良好的電阻率剖面，自然也難以觀察到整治藥劑的影響。
現階段本研究建構的套件，嵌入現行的資料處理系統相對簡易、快速，且訓練好的神經網路可以大幅縮減實際成像時間。此外，對於具有相似場址條件的相同陣列，可以套用已訓練的模型。更重要的是，此套件的工作流程可加入不同觀測資料，亦可供不同反演問題作為參考。除上述優點外，此套件亦有缺點。例如：訓練階段需花費大量算力、預測的電阻率剖面難以看到細微構造、資料品質及缺失值影響預測結果等。因此後續仍需改善套件。
;Today’s resistivity imaging processes not only take a long time to get one result, but also sensitive to human manipulation, making it difficult to apply to an automatic system. Some studies using neural networks have attempted to solve this problem, but neural networks in these studies have less inference power than modern deep neural networks. Besides, these studies haven’t released a toolbox for reference, so those who study the same research field need to reproduce the corresponding workflow. In the light of this, the goal of my study is to develop a resistivity imaging toolbox based on deep learning. The forward simulation part of this toolbox is a potential simulation program developed by Pidlisecky & Knight (2008); the neural network part is built upon Keras with Tensorflow backbone.
We used synthetic data and field data to test the electrical resistivity imaging process established in this study. In the synthetic data section, we designed five configurations of training data for training neural networks. Then, we used layered, double-block and multi-block resistivity models to evaluate the predictive ability of neural networks. From the result of resistivity profiles, we can observe the effects of different operations on training data. In addition, we compared the resistivity profiles produced by EarthImager 2D and neural networks to observe the differences between these profiles. In the field data section, we used the data measured by the north-south array at the Yongkang site from October 10 to October 21, 2015. The remediation reagent injection was implemented at the Yongkang site from October 13 to October 18, 2015. Whether the change in resistivity caused by the injection can be observed in the resistivity profile can be used as an indicator for evaluating the imaging procedure of this study. For field data testing, we designed two training dataset, one based on common geological materials and the other based on core data. The final result shows that the neural network trained based on prior information has a good predictive power for YongKang site. However, the quality of the field data is not good, making it difficult to obtain a well resistivity profile even with conventional inversion. Therefore, it is difficult to observe the influence of the remediation reagent.
The toolbox constructed in this study is easily embedded in the present data processing system. The trained model can greatly reduce the actual imaging time, and it may be applied to the same array with similar site conditions. More importantly, the workflow of this toolbox can add different observations data and can be used as a reference for different inversion problems. However, this toolbox has both advantages and disadvantages. The training phase requires a lot of computational resource, the predicted resistivity profile is difficult to see the fine structure, and the data quality and missing values affect the prediction results. Thus, we will continuously improve this toolbox in the future. |
