土壤液化潛能之類神經網路分析

Abstract

地表基礎地層受到地震力激烈擾動，導致土壤發生液化現象，降低基礎之承載能力，使結構物產生不均勻之沉陷，造成結構之破壞及人員之傷亡，為地震引起之主要災害。土壤液化現象起因於飽和低凝聚性土壤受地震力載重作用，導致孔隙水壓急速上升，而降低土壤原有強度。土壤液化現象為一相當複雜之土壤行為，工程實務中，為了解一工址在遭遇地震時發生土壤液化之可能性，需經由土壤性質、環境條件及地震特性等影響因素評估土壤之液化潛能。目前大部分土壤液化潛能評估多是以地震液化地區的現地觀資料所建立之半理論方法。經由土壤液化潛能評估過程，可了解一地區之液化可能性及其風險，以提供工程實務參考之用。 類神經網路為一模仿生物神經網路運作架構之資訊處理系統，其具有高維度的學習功能，適合應用於分類及預測等問題。近年來已有學者以類神經網路分析土壤液化問題，並提出相關之類神經網路液化潛能評估模式。由過去相關研究顯示，類神經網路在土壤液化評估之應用有良好的表現，為一可行的方法。 本研究利用三層倒傳遞類神經網路架構，以地震特性及土壤性質為網路模式輸入參數，地震後土壤中之超額孔隙水壓為網路輸出值，建立類神經網路液化潛能評估模式。採用實際液化案例為網路訓練範例進行網路訓練。並以實際孔隙水壓記錄做為網路測試範例，檢驗網路模式之準確性。

Ground failure induced by liquefaction is a major cause of damage in post earthquake and poses considerable hazard to structures and their occupants. Liquefaction is one of the most complex phenomena resulting from earthquakes. The basic cause of liquefaction in saturated cohesionless soil during earthquakes is the buildup of excess pore pressure due to cyclic shear stress induced by the ground motion. The liquefaction potential of a particular site can be assessed based on information such as soil properties, environmental factors and earthquake characteristics. Most current techniques used for the assessment of liquefaction are semitheoretical methods. By these methods we can classify a site as liquefiable or non-liquefiable. Artificial neural networks are information-processing systems whose architectures essentially mimic the biological system of the brain. Neural networks are capable of mapping and capturing many features and subfeatures embedded in a large set of data that yield a certain output. The feasibility of using neural network to model the complex relationship between the seismic and soil parameters, and the liquefaction potential has been investigated. This paper presents the development of neural network models for the prediction of the excess pore pressure after earthquakes. backpropagation neural networks with one hidden layer were used to establish the models. The neural networks were trained using actual fields records, and tested using actual pore pressure records.