پيش‌بيني پتانسيل رمبندگي خاك‌هاي رمبنده از طريق شبكه‌هاي عصبي مصنوعي

Prediction of collapse potential of soils using Artificial Neural Network

ر اين پژوهش به منظور بررسي قابليت شبكه‌هاي عصبي مصنوعي در تعيين پتانسيل رمبندگي، نمونه‌هاي متعدد خاك رمبنده از يك منطقه (دشت زاهدان) گردآوري شده است. در آزمايشگاه آزمايش‌هاي معمول رمبندگي روي آن‌ها انجام و تعداد 130 نمونه خاك رمبنده به دست آمده از اعماق و مكان‌هاي مختلف دشت در پايگاه داده‌ ثبت شد. آزمايش رمبندگي انجام شده، تحكيم مضاعف بوده كه براي بررسي بيشتر آزمايش‌هاي دانه‌بندي، وزن مخصوص، حدود اتربرگ و خواص مقاومتي نيز روي نمونه‌ها انجام شد. در مراحل بعد نتايج براي ورود به شبكه‌هاي عصبي مصنوعي آماده شده و مدل‌سازي انجام شد. پس از مرحله آموزش شبكه و يادگيري، مدل‌هاي مختلف شبكه مورد سعي و خطا قرار گرفته و در ادامه مدل بهينه شبكه شامل شش ورودي و يك خروجي انتخاب شده است. با توجه به نتايج پيش‌بيني، مشخص شد كه بين داده‌هاي تجربي و پيش‌بيني شده به وسيله‌ي شبكه‌ عصبي مصنوعي بيشتر از 95 درصد همبستگي مشاهده مي‌شود. همچنين نتايج نشان مي‌دهد كه شبكه‌هاي عصبي مصنوعي مي‌تواند پتانسيل رمبندگي را به طور مناسبي پيش‌بيني كند و به دليل استفاده از آزمايش‌هاي ساده و كم هزينه، باعث كاهش حجم محاسبات و آزمايش‌هاي لازم خواهد شد.

Structure collapse and subsidence represent major geotechnical problems, particularly in areas containing loess, which is a widespread collapsing soil. The necessary open soil structure is formed by aeolian deposition of the constituent particles. Collapsible soils are very sensitive to changes of porosity and moisture content. Collapsibility is a property of some cohesive and cohesionless soils or constructed fills whose volume suddenly decreases with an increase of moisture content under practically unchanged total vertical stress. Although the more porous soils or fills collapse under smaller loads than the denser soils or fills, some official documents or design guidelines accepted by many countries concentrate the attention of designers and engineers only on effect of moisture content on soil collapsibility. In this study, the ability of Artificial Neural Networks (ANN) has been investigated to determine the collapse potential of soils. Therefore, different samples of collapsible soil have been collected from an area (Zahedan plain). General tests were performed on the samples in the laboratory and 130 samples of collapsible soil from different depths and locations were recorded in the database. The collapse potential tests (One-dimensional collapse test) was carried out on the samples and with the aim of further investigations, the grain size distribution, specific gravity, atterberg limits and strength properties of the samples were performed. In the later stages, the collapsible samples data were prepared for the artificial neural networks input. Neural network is a functional abstraction of the biological neural structures of the central nervous system. It can exhibit a surprising number of human brain’s characteristics, e.g., learning from experience and generalizing from previous examples to solve new problems. In the early days of artificial intelligence research, Frank Rosenblatt devised a machine called the perceptron that operated much in the same way as the human mind. A perceptron is a connected network that simulates an associative memory. The most basic perceptron is composed of an input layer and an output layer of nodes, each fully connected to the other. A weight is assigned to each connection, which can be adjusted in such a manner that when a set of inputs is given to the network, the associated connections will produce a desired output. Adjusting of these weights to produce a particular output is called ‘‘training’’ of the network; a mechanism that allows the network to learn. Perceptrons are among the earliest and most basic models of ANNs, yet they are in use in many of today’s complex neural net applications. After the network training process and the subsequent learning, some network models have been selected under experiments, which include six inputs and one output. According to the predicted results, it was indicated that the correlation between experimental and predicted data by the ANN is 95%. Furthermore, the results show that artificial neural networks can predict collapse potential of soils, also the calculations and required tests will be reduced due to their simple use and inexpensive tests