انتخاب آرايش بهينه‌ي شبكه‌ي عصبي در تحليل خاك‌هاي مخلوط با خرده لاستيك‌هاي بازيافتي

Optimized Artificial Neural Network for Analyzing Soil-Waste Rubber Shred Mixtures

يكي از پارامترهاي مهم در انتخاب مصالح مقدار مقاومت برشي و رفتار مكانيكي آن‌ است. روش‌هاي مختلفي جهت بهبود و افزايش پارامترهاي مقاومتي خاك‌ها همچون به‌كارگيري خرده‌هاي لاستيك مخلوط با خاك موجود است. جهت بررسي درصد اختلاط و تاثير آن در پارامترهاي مقاومتي تركيب ايجادشده، نياز به انجام آزمون‌هاي سه محوري است؛ كه از معايب اصلي آن مي‌توان به لزوم در اختيارداشتن تجهيزات آزمايشگاهي، زمان‌بر بودن و توانايي انجام يك آزمايش در واحد زمان اشاره كرد. در اين نوشتار با استفاده از 906 داده‌ي ناشي از آزمايش‌هاي سه محوري به طرح شبكه‌ي عصبي مصنوعي جهت پيش‌بيني تنش انحرافي و كرنش حجمي اين مخلوط‌ها پرداخته شده است. درنهايت، شبكه‌ از نوع انتشار برگشتي با آرايش 2-12-4 به‌عنوان شبكه‌ي بهينه انتخاب شده است، كه نتايج دقت مناسب شبكه‌ي مذكور جهت شيبه‌سازي رفتار فوق را نشان مي‌دهد.

Shear strength and mechanical behavior are most important features when construction materials are to be selected. In recent decades, the use of recycled materials has increased dramatically. Rubber produce is a widespread element that can hardly be decomposed. Considering the advances in geotechnical engineering, different approaches, such as adding waste rubber shreds to sandy soils, were proposed for stabilization and bearing characteristic improvements. Rubber shreds and rubber shred – soil mixtures can be used as alternative backfill material in many geotechnical applications. The reuse of rubber shreds may not only address growing environmental and economic concerns, but also help solve geotechnical problems associated with low soil shear strength. An appropriate mixing fraction and its influence on strength parameters and volumetric strain behavior must be obtained through triaxial testing, which is a time consuming process and requires laboratory equipment. Not being able to examine different specimens with a unit is another problem of triaxial testing and is another disadvantage of this experiment. On the other hand, artificial neural networks, namely ANNs, are an artificial intelligence field built from a large number of simple processes. ANN processing elements deal individually with parts of a large problem, and indeed self-learning mechanisms. Low input parameters and multi modeling at the same time in the ANN approach, lead to reduced modeling process duration. This study, based on 906 experimental data acquired from triaxial tests on different sandy soil-rubber mixtures, is qualified to stimulate triaxial tests and is used to design the ANN for predicting deviatoric stress and volumetric strain changes. Back propagation networks with 4-12-2 combinations with sigmoid internal and linear external functions were chosen as the optimized ANN. The results showed that ANN is perfectly capable of modeling soil-rubber mixture mechanical behavior.