پديد آورندگان :
علوي عليرضا دانشگاه سمنان - دانشكده مهندسي عمران , نادرپور حسين دانشگاه سمنان - دانشكده مهندسي عمران , فخاريان پويان دانشگاه سمنان - دانشكده مهندسي عمران , نوغاني سلمان دانشگاه آزاد اسلامي - دانشكده فني و مهندسي
كليدواژه :
ظرفيت دوراني , تير بال پهن , شبكه عصبي مصنوعي بيزين , رفتار پلاستيك
چكيده فارسي :
مدل شبكه عصبي مصنوعي بيزين يكي از جديدترين روش هاي محاسبات نرم به شمار مي رود. استفاده از روش هاي متفاوت مانند المان محدود،روش هاي رگرسيوني و آماري در پژوهش هاي پيشين در سال هاي اخير مورد بحث قرار گرفته است. در اين مقاله براي اولين بار با ساخت يك مدل شبكه عصبي مصنوعي بيزين كه جزو جديدترين روش هاي محاسبات نرم است، ظرفيت دوراني تير بال پهن فولادي تخمين زده شده است. استفاده از روش هاي متفاوت نظير المان محدود، روش هاي رگرسيوني و آماري در پژوهش هاي پيشين در سال هاي اخير مورد بحث قرار گرفته است، لذا در اين پژوهش براي تخمين دقيق تر و سريع تر ظرفيت دوراني از تكنيك شبكه هاي عصبي مصنوعي از نوع شبكه بيزين استفاده شده است. داده هاي لازم براي آموزش و آزمايش شبكه بر اساس نتايج آزمايشگاهي معتبر از تاريخچه تحقيقات به دست آمده است. ورودي هاي مدل ساخته شده براي اين منظور شامل نصف عرض بال تير، ارتفاع جان، ضخامت بال، ضخامت جان، طول، تنش تسليم بال و تنش تسليم جان است و تابع هدف نيز ظرفيت دوراني مي باشد. نتايج به دست آمده از اين مدل با نتايج آزمايشگاهي و ديگر مدل هاي ارايه شده در گذشته مورد مقايسه دقيق قرار گرفت. نتيجه اين مطالعه نشان دهنده آن است كه استفاده از اين رويكرد جديد نسبت ديگر مدل ها داراي دقت بالاتري بوده و كاربردي خواهد بود و مي توان از شبكه عصبي مصنوعي بيزين به عنوان ابزاري قدرتمند در اين گونه از مسايل بهره برد.
چكيده لاتين :
In this paper, for the first time using of Bayesian regularized artificial neural network (BRANN) model, which is a novel method of among soft computing (SC) methods (such as fuzzy logic, genetic programming, neural network) to predict the rotational capacity of wide-flange steel beams. Steel is one of the most commonly used materials in construction industries, mainly in steel structures. There are many researches and studies on the behavior of a structural member of steel structure such as beams under different types of loading. The accurate estimation of rotation capacity (plastic rotation capacity) is of significant importance issue for plastic and seismic analysis and design of steel structures especially for high rise building (nonlinear behavior). Similarly, the moment redistribution in a steel structure also depends on the rotation capacity of the section. So the determination and accurate prediction of rotation capacity of steel structures members such as wide flange beams become an important task. Using different methods such as finite element, regression and statistical methods in previous studies has been used in recent years. Therefore, in order to estimate the more accurate value of the rotational capacity of wide flange beams, Artificial neural networks are used with the Bayesian learning process. The Bayesian regularized network assigns a probabilistic nature to the network weights, allowing the network to automatically and optimally penalize excessively complex models. The proposed technique (BRANN) reduces the potential for overfitting and overtraining, improving the prediction quality and generalization of the network. The proposed model (BRANN) is based on experimental data that collected from previous studies. After a comprehensive review of existing literature, 77 data of wide flange beam were selected which had experienced to determined rotation capacity. For this purpose, Half-length of flange, height of web, thickness of flange, thickness of web, length of beam, yield strength of flange and yield strength of web were consider as input parameters (six inputs) while rotation capacity is treated as target of the Bayesian regularized artificial neural network model. The Bayesian regularized artificial neural network is modeled in MATLAB software and applied to predict the rotation capacity. The results of this model were compared with experimental results and other models and equations that presented in the past (including Genetic programming (GP), Li equation and Kemp Equation. An analysis is carried out to check the performance of the proposed BRANN model based on the common criteria such as Mean Absolute Percentage Error (MAPE). The optimal and best model should have the lowest values of MAPE, this parameter is 20.32% for BRANN, 23.49% for a Genetic Programming model that proposed by Cevik, 47/20% for Li’s Equation and 56.98% for Kemp’s equations. The results of Bayesian regularized artificial neural network approach indicate a good agreement between the predicted and measured data. Furthermore, the Bayesian regularized artificial neural network model shows the most optimized results compared to all the previous model and equations. The result indicated that the Bayesian regularized artificial neural network could be used as a powerful tool for engineers and researcher to solve this kind of problems.
