تخمین تزریق‌پذیری خاک‌های دانه‌ای با به‌کارگیری داده‌های آزمایشگاهی و چند روش طبقه‌بندی هوشمند

تزریق‌پذیری یک پارامتر بااهمیت در عملیات تزریق است و پیش‌بینی صحیح آن منجر به انتخاب مناسب مواد سیال تزریق شونده می‌شود. این پارامتر در اکثر مواقع با روش‌های تجربی تخمین زده می‌شود و پیش‌بینی را با خطا همراه می‌کند. در این تحقیق سعی شد به‌منظور ساخت و صحت سنجی چند مدل داده‌کاوی در حوضه‌ی طبقه‌بندی، مجموعه‌ای از داده‌های آزمایشگاهی در عملیات تزریق موجود در چندین منبع به کار گرفته شود. مدل‌های طبقه‌بندی بکار گرفته‌شده در نرم‌افزار orange شامل روش‌های ماشین بردار پشتیبان، شبکه عصبی مصنوعی، نزدیک‌ترین همسایگی، جنگل تصادفی و بیزین ساده می‌باشند. در این مدل‌ها، متغیرهای ورودی عبارت است از: نسبت آب به سیمان در دوغاب تزریق شونده، دانسیته نسبی خاک، فشار تزریق، درصد ریزدانه خاک، نسبت قطر ذرات خاک که 15 درصد وزنی نمونه از آن کوچک‌تر است به قطر ذرات سیال تزریقی که 85 درصد وزنی نمونه از آن کوچک‌تر است (n1=d15 soil/d85 grout و n2=d10 soil/d95 grout). پس از مدل‌سازی، نتایج نشان می‌دهد که مدل‌های بکار گرفته‌شده به‌خوبی رابطه‌ی بین تزریق‌پذیری و عوامل موثر آن را تعریف می‌کنند و از دقت بالایی در تخمین تزریق‌پذیری خاک‌های دانه‌ای برخوردار هستند. با توجه به ماتریس کارایی مدل‌ها، مدل شبکه عصبی مصنوعی با دقت 0/86 درصد و مدل نزدیک‌ترین همسایگی با دقت 0/85 درصد عملکرد بهتری نسبت به سایر روش‌ها دارند. بعلاوه در بررسی اهمیت متغیرهای ورودی بر اساس شاخص‌های امتیازدهی، متغیرهای n2 و n1 تاثیرگذارترین متغیرها در روند پیش‌بینی صحیح تزریق‌پذیری هستند.

Estimation of Groutability of granular soils using laboratory data and several intelligent classification methods

SummaryIn this study, in order to construct and validate several classification models, a set of laboratory data was used in the grouting operations in several literature. Classification models include Artificial Neural Network (ANN), Support Vector Machine (SVM), KNearest Neighbor (KNN), Random Forest (RF) and Naive Bayes (NB). Orange software has been used in this regard. The results showed that the models have a high accuracy in estimating groutability, and among them the artificial neural network method with 0.86% precision has better performance than other methods. In addition, in examining the importance of input variables based on scoring indices, the N2 and N1 variables are the most influential variables in the process of correctly predicting groutability. IntroductionThe purpose of grouting is to strengthen and improve the mechanical and hydraulic properties of the rock and soil. The fluid that is injected into the cavities and fissures of the environment is like a viscous liquid consisting of grains whose size is important in the grouting operation. Therefore, determining the groutability ratio in grouting operation is considered as an important parameter. Today, studies using data mining science show that the groutability of granular soils, in addition to grain size, is affected by various factors of the soil and the material of grout, which predicts groutability more accurately. Throughout history, many researchers have predicted groutability through experimental relationships. However, today, the capability of data mining methods in accurate predictions has shown that one approach in predicting groutability is to use a variety of data mining models and inferential systems. Methodology and ApproachesThe purpose of this study is to evaluate several models of data mining methods, including ANN, SVM, KNN, RF and NB. For this purpose, a set of laboratory information related to groutability has been used in four literatures that include 87 data in order to develop efficient models for predicting groutability. Classification models are created in Orange software. Results and ConclusionsThe output variable is a property of groutability, which as a binary variable has two states of zero meaning nongroutable and 1 meaning groutable. Input variables also include the ratio of water cement in the grout or viscosity (W/C), the relative density of the soil (Dr), grouting pressure (P), the percentage of the soil particles passing through a 0.6 mm sieve (FC), N1 = D15soil / D85 grout and N2 = D10 soil / D95 grout. The values of the evaluation criteria for the methods are almost close to each other. Based on the AUC index, the random forest is the best model and the knearest neighbor method has the lowest value of this index. However, in terms of other criteria, the artificial neural network is higher than other methods and the knearest neighbor method is very close to it. On the other hand, the random forest model has the lowest value of criteria. Ignoring the AUC criteria, ANN and KNN methods are the best methods.One of the capabilities of Orange software is to study the effect and importance of input variables on the prediction of the target variable, in other words, the sensitivity of the output variable to input variables. The results show that variable N2 is in the first level based on the three criteria of information gain, relative information gain and Gini index, and variable N1 is in the second level with a very small difference in the values of the criteria. In addition, in the last row, W/C has the lowest value of the criteria and shows a small role in the correct prediction of groutability.