基于神经网络算法的滚刀磨损量预测方法-《郑州大学学报(工学版)》

文章信息/Info

摘要:: 为了给盾构施工开仓换刀提供参考，本文将滚刀磨损简化为多元非线性拟合问题，构建了数据分析框架，综合考虑机械、地质、管理三大类因素，对滚刀磨损量进行预测。以广州地铁18号线番禺广场到南村万博站区间盾构区间为工程背景，选取共14种参数，结合公式获得标定后数据共2386条，包含34把正面滚刀，共81次滚刀磨损量。通过SMBO方法和LM算法改进BPNN算法训练过程，充分发挥神经网络优势，所得模型对83.3%的测试样本的预测值决定系数(R2)高于0.86，相比标记样本时参考的公式，准确度有较大提高，表明该方法所训练模型对滚刀磨损量的发展趋势预测更加准确。

Abstract:: To provide a reference for manual cutter inspection in shield tunneling, our study simplifies the wear of disc cutters as a multivariate nonlinear regression problem and construct a data analysis fr＜x＞amework to predict the cutter wear quantitatively by combining the effect of three kinds of factors, which are machinery, geology, and management. Taking the shield tunnel section from Panyu Square to Nancun Wanbo station of Guangzhou Metro Line 18 as the engineering background, we select 4 parameters and obtained 2386 labeled data derived from 34 face cutters and 81 manual inspections. SMBO method. We expedite the training of BPNN with the LM algorithm, which fully exploits the regression ability of the neural network. The prediction got coefficients of determination (R2) over 0.86 for 83.3% of the test samples, and the accuracy was greatly improved compared with the reference formula used for data labeling. It shows that the model trained by this method has higher accuracy in the prediction of disc cutter wear.