CAJ | 학술논문

为了通过神经网络预测土密实度，搭建了土密实度检测装置，采集了土的激光图像，并提取了土的吸收系数、散射系数、激光图像的纹理特征参量和图像的灰度变化率共13个特征参数，通过 SPSS 降维处理，提取了5个主成分因子作为输入特征；并利用 BP 神经网络预测密实度．结果表明：经过31次测试后达到了误差要求，预测值与环刀法结果比较的平均绝对误差为7．14％和平均相对误差为7．71％，所建立的预测模型可行；最后通过试验对预测模型进行验证，其预测值与环刀法结果比较的平均绝对误差为8．62％和平均相对误差为8．76％，说明用神经网络预测土密实度是可行的．
위료통과신경망락예측토밀실도，탑건료토밀실도검측장치，채집료토적격광도상，병제취료토적흡수계수、산사계수、격광도상적문리특정삼량화도상적회도변화솔공13개특정삼수，통과 SPSS 강유처리，제취료5개주성분인자작위수입특정；병이용 BP 신경망락예측밀실도．결과표명：경과31차측시후체도료오차요구，예측치여배도법결과비교적평균절대오차위7．14％화평균상대오차위7．71％，소건립적예측모형가행；최후통과시험대예측모형진행험증，기예측치여배도법결과비교적평균절대오차위8．62％화평균상대오차위8．76％，설명용신경망락예측토밀실도시가행적．
In order to test soil compactness by the neural network,a laser image measurement system of soil compactness was established.The thirteen parameters were collected from each image,such as the ab-sorption coefficient,scattering coefficient,texture features and image gray change rate.The SPSS was per-formed to select five characteristic parameters,and soil compactness was predicted by BP neural networks. The results showed that BP neural networks reached the required error after 31 loops.Compared with the measurement values by using round knife method,the average absolute error was 7.14%,and average rela-tive error was 7.71%,the BP neural network was feasible.The predictive model was verified,the predictive value with cutting ring results in an average absolute error was 8.62%,and average relative error was 8.76%.It is suggested that soil compactness by the neural network is feasible.