| 摘要: | 岩體分類系統是決定隧道支撐類型的重要依據。現行岩體分類系統多源自於國外，應用在台灣地區複雜的地質環境有改善之必要。本研究蒐集北二高福德隧道與北迴鐵路新觀音隧道的輪進資料，經輸入因子的正規化與輸出分類的模糊化處理，再運用倒傳遞類神經網路( Back Propagation Artificial Neural Network, BPN )訓練，將樣本學習下來作為其他隧道的岩體分類依據。 福德隧道案例中， RMR的6項因子資料經去除重複性後，共計有300筆樣本，以此訓練得到各處理單元間之權重，將全部2019個樣本類神經網路經分類後得到系統輸出值，將其與目標輸出值(即專家分類的支撐型式)進行對照，結果第一正確率(分類結果與原始類別完全相同)達74.39%，第二正確率(分類結果與原始類別差一級)達96.19%。利用新觀音隧道地質報告表挑選出17項因子，共2099筆資料經類神經網路輸入訓練並測試，其結果第一正確率高達99.05%。研究顯示類神經網路架構以兩個隱藏層、隱藏層內處理單元數13個以上、訓練次數5000次以上、慣性因子與學習速率0.5左右，可以得到最佳岩體分類結果。 經由福德隧道與新觀音隧道資料分析顯示： (1) 在未考慮覆蓋因子的情況下，去除洞口段資料有其必要性。 (2)考慮更多的分類因子，類神經網路系統可以有效學習，做出良好的支撐建議。 (3)只要樣本的品質夠好，使用愈多的訓練樣本可以得到愈好的結果，但若是訓練樣本品質不佳，越多的樣本只會增加訓練學習上的混亂。 (4)僅以隧道開挖的前三分之一的資料進行類神經網路訓練，並無法完全預測隧道剩餘三分之二區段的地質狀況，而給予適當支撐建議。必須有賴野外露頭調查與鑽探…等方式，增加對隧道剩餘區段未知地質況狀的掌控。 (5)本類神經網路系統的輸出值是一個隸屬函數，有利於使用者根據隸屬函數做彈性的決策，作成最後支撐型式的決定。在未來可蒐集更多的隧道地質與工程資料來建立類神經網路模型，並配合群集分類與因子分析等方法，來建立更為完善可行的隧道岩體分類系統。 Rock mass classification system is an important criterion for proposing a support type in rock tunneling. Currently rock mass classification systems used in Taiwan are all come from abroad, and it is necessary to modify these methods for complex geologic condition in Taiwan. In this study, we take the advantages of efficiency and quick learning character of a non-linear and supervised classification problem, and uses BPN (Back Propagation Artificial Neural Network) method to perform the rock mass classification. We collected tunnel engineering reports of the Fu-de and the New Guan-yin tunnels as examples for study. We adopted geologic factors from these reports, and normalize the factors as input level. In the output level of BPN model, we designed a fuzzy membership, so as uncertainty could be considered. We adopt some good samples for BPN learning and the parameters judging, and then use the BPN and rest data for testing the performance of the system. In the case of Fu-de tunnel, we use 300 good samples for BPN training and learning and get a good BPN model. We test the rest 2019 samples with the BNP model, and result reveals that 74.39% cases output are exact by the same type of support as the target type and 96.19% cases output support type within one neighboring class of the target type. In the case of Guan-yin, we picked up 17 geologic factors from the engineering reports and summarized 2099 samples for learning and testing in BPN model. Result reveals that the accuracy rate is 99.05% with the suggestion is exactly as the target type. After these two case studies, the best BPN models are two hind layers, the neural units of hind layers are above 13, training more than 5000 times, and moment factor and learning rate are almost closed to 0.5. Results from the BPN model of the Fu-de and the New Guan-yin tunnels may conclude: (a) If the overburden factor is exclude for the analysis, it is necessary to remove the test data at portal section of the tunnel. (b) The result in BPN training and testing could be better if we consider more factors for analyses. (c) When the quality of data is good enough, we may use as much data as we have to get the best result. Whereas more data produce more chaotic BPN model, when data quality is bad. (d) BPN model trained from the first 1/3 portion of the tunnel is not good enough to predict the support type and geologic condition of the rest of the tunnel. Field investigation and drilling are necessary for determining the supporting type for the rest potion of tunnel. (e) The fuzzy membership function output from the BPN model can help us making decisions. In the future, we may collect a large number of tunnel geologic and construction data, and establish a more general BPN model. With the assistance of factor analysis and cluster analysis, we may construct a more complete and friendly procedure for tunnel rock mass classification and support prediction. |
