模組式類神經網路於高性能混凝土抗壓強度預測之應用

Abstract

為了增加傳統混凝土的應用範圍，因此發展出高性能混凝土，材料上除了使用傳統混凝土原有材料，更添加其他波索蘭材料，而配比方法也異於傳統，文獻已證明高性能混凝土相較於傳統混凝土更具流動性及高強度，相對地，高性能混凝土的行為也更為複雜，更不易以回歸分析方式建立有效抗壓強度預測公式。而類神經網路具有處理非線性問題之能力，但對於複雜性高之問題，單一類神經網路可能發生學習不易或學習效率低落之狀況，模組式類神經網路為改進此缺點，將原始複雜問題劃分成多個小問題，進行分析以提高學習效率並減低問題之複雜度，因此本研究透過模組式類神經網路建立高性能混凝土之抗壓強度預測模型，並輔以敏感度分析取得類神經網路輸入變數對於網路本身的影響關係。另一方面，為瞭解模組式類神經網路是否可有效應用於高性能混凝土強度預測，並與單一類神經網路所建立之強度預測模型進行比較驗證。

In addition to the four basic ingredients of the conventional| concrete, i.e., Portland cement, fine and coarse aggregates, and water, the making of HPC needs to incorporate the supplementary cementations materials, such as fly ash and blast furnace slag, and chemical admixtures such as superplasticizer. Hence, the characteristics of HPC are much more complex and hard to build an effective model to estimate the strength by mathematical model. For overcoming the disadvantage of standalone neural network learning model, modularized neural networks divide the original problem into several little aspects, which could be learned easily and calculate. This work presents a modularized neural network to predict the strength properties of high-performance concrete (HPC) mixes. About thousand data collected from different labs are used as training instances. For the sake of comparison, the training instances are also trained using two standalone neural networks, one for conventional concrete and the other for HPC. Moreover, the sensitive analysis is employed to find the strength relationship between the input and output data. The simulation results reveals that the modularized neural network can reasonably predict the strength of HPC.