- 學位論文
以類神經網路預測炸震夯實之成效
Using Artificial Neural Network to Predict the Effect of Blasting Densification
摘要
台灣位於西太平洋地震帶，每月中發生數以千計次地震，其中九二一地震造成土壤液化並帶來重大災害，因此，如何在建築物施工前將地盤改良將是一項重要的防治工程。 本研究分為兩部份，第一部份使用炸震夯實法，規畫佈孔設計並配置炸藥，藉由炸藥引爆時產生之高壓震波，使爆源附近之砂質土壤(sandy soil)產生土壤液化，待液化現象消退後，鑽取土樣進行土壤性質試驗。炸震夯實工法所使用的炸藥藥量、炮管深度、炮孔間距及土壤顆粒之震動速率為實驗設計參數，而炸震夯實之成效取決於含水量、孔隙率、飽和度及相對密度等土壤性質試驗之結果。 第二部份使用倒傳遞類神經網路，包含輸入層、隱藏層及輸出層，各層人工神經元藉由加權鏈結值與閥值連接，經由多次訓練與學習，將訓練值與實際值的誤差回饋於權值，以調整權值與閥值直到網路收斂為止。最後比較含水量、孔隙率、飽和度及相對密度預測值與實際值之誤差。 實驗結果顯示，施炸後土壤性質參數中之含水量、孔隙率及飽和度會隨時間增加而下降，相對密度則隨時間增加而上升；倒傳遞類神經網路預測之誤差範圍皆在20%以內，其中最大誤差百分比為19%。
並列摘要
Taiwan located on West-Pacific seismic belt, thousands of earthquakes occurred around Taiwan in a month. The most famous 921 earthquake which known as Chi-chi earthquake induced soil liquefaction and caused huge damage. Thurs, how to improve the building ground before construction is an important improvement project. This research departs in to two parts: The first part used blasting densification to improve land site. The blasting holes and explosives were designed and set. The sandy soil around the blasting sources was liquefied by high pressure shock waves which were caused by blasting. Drilling and receiving soil samples to conduct soil character tests. The site experiment factors included the weight of explosives, the depth of blasting holes, the distance of blasting holes and the vibration velocity of soil particles. The effect of blasting densification depended on the results of soil character test including water content, porosity, saturation and relative density. The second part used back-propagation neural network. It contented input layer, hiding layer and output layer. Each neurons of layer were connected by weights and biases. After training and learning repeatedly, the errors between training values and real test values were feedback to weights, and regulated the weights and biases until the network was convergence. The predicting values of water content, porosity, saturation and relative density were compared to real test values and the errors were discussed in the end. The result of experiment shows that the values of water content, porosity and saturation increased when time increased, but the values of relative density decreased when time increased; the range of error which was predicted by back-propagation neural network is under 20%, the largest error percentage is 19%.