پيش بيني رفتار تنش_كرنش مصالح شني با استفاده از شبكه هاي عصبي مصنوعي

Prediction of stress- strain behavior in gravelly material based on Artificial Neural Networks

در اين پژوهش رفتار مكانيكي مصالح درشت دانه شني با استفاده از شبكه عصبي چند لايه پرسپترون، كه از پركاربردترين شبكه هاي عصبي مصنوعي در مسايل ژيوتكنيكي است، شبيه سازي شده است. ابتدا اطلاعات دقيقي از آزمون هاي منابع مختلف در سراسر كشور تهيه و عوامل موثر بر مقاومت برشي خاك هاي درشت دانه بررسي شده است. پس از حذف اطلاعات نادرست، روند يادگيري، آزمايش و پيش بيني شبكه طي شده است. در آموزش شبكه از الگوريتم يادگيري پس انتشار خطا استفاده شده است . پارامترهاي استفاده شده در آموزش شبكه شامل خصوصيات دانه بندي، چگالي خشك، چگالي نسبي، درصد سايش لس آنجلس، فشار همه جانبه، كرنش و تنش انحرافي است. براي تعيين چگونگي و مقدار تاثير ورودي ها بر خروجي مدل، تحليل حساسيت روي آن ها انجام شده و نتايج به دست آمده با قوانين مكانيك خاك مقايسه شده است. بررسي مدل گوياي اين واقعيت است كه شبكه ارايه شده، توانايي لازم براي پيش بيني رفتار تنش_كرنش خاك هاي درشت دانه را دارد.

Prediction of stress-strain behavior of geotechnical material is one of the major efforts of engineers and researchers in the field of geomechanics. Experimental tests like tri-axial shear strength tests are the most effective apparatus to prepare the mechanical characteristics of gravelly material; but due to difficulties in preparing test samples and costs of the tests, only several tests will be done in a new project. Artificial neural network is a kind of method, in which engineer could judge the results based on numerous data from other similar projects, which enable the engineer to have a good judgment on the material properties. In this research, the behavior of gravelly material was simulated by use of multi-layer perceptron neural network, which is the most useful kind of artificial neural networks in the field of geotechnical engineering. For instance, first exact information was provided from laboratory tests of various barrow areas of embankment dams in the country and effective parameters on shear strength of coarse-grained material were studied. After omitting incorrect or weak data, 95, 20 and 23 sets of data were used for learning, testing and evaluating data, respectively. Input parameters for the model were as follows: particle-size distribution curve, dry density, relative density, Los-angles abrasion percent, confining pressure, axial strain; and outputs were selected as deviator stress. In order to reach a steady state in the model and force the model to behave homogenous to the all inputs, data was normalized to the value between .05 and 0.95. In the simulation, back-propagation algorithm was used for learning or error reduction. The aim of the simulations was defined to reduce error between real data and predicted values; for instance root mean square error (RMS) was used to be minimized through simulation and predicted versus real graphs were used to observe the global error of the model. After modeling the data based on some criteria, it was shown that curves of stress-strain from simulation tests were in good agreement with those from laboratory. These close coherencies were observed in all training, testing and evaluation data, in which the RMS errors were 0.038, 0.037 and 0.026, respectively. To reach this ultimate step, a 10*19*1 multilayer perceptron was used via trial and error. In order to determine quality and quantity of the effect of inputs on outputs, and prove that the results were in good agreement with soil mechanic principles, sensitivity analyses were done on the average data of the inputs. Results show that confine pressure, uniformity coefficient and relative density of the material were the most effective parameters on the stress-strain curves; thus the model has enough capability to predict the stress-strain behavior of gravelly soils.