تعيين ظرفيت باربري شمع هاي كوبيده‌شده در خاك هاي غيرچسبنده با رويكرد شبكه‌ي عصبي موجكي

LOAD CARRYING CAPACITY DETERMINATION OF PILE DRIVEN IN NONCOHESIVE SOIL WITH WAVELET NEURAL NETWORK APPROACH

با وجود منابع علمي زياد، كماكان تعيين ظرفيت باربري شمع هاي كوبيده‌شده در خاك هاي غير چسبنده‌ توام با پيچيدگي است. يك شبكه‌ي عصبي موجك، توابع موجك را به‌منزله‌ي توابع فعال‌ساز نرون هاي لايه‌ي پنهان از شبكه‌ي عصبي پيشخورد به‌كار مي گيرد. در اين شبكه ها هر دو پارامتر انتقال و مقياس موجك ها در كنار وزن هايشان بهينه مي‌شوند. در رويكردي خاص از ساخت اين نوع شبكه ها، با عنوان ويونت، پارامترهاي انتقال و مقياس ثابت مي‌ماند و فقط وزن ها بهينه مي شوند. در اين نوشتار، با درنظرگرفتن رويكرد فوق در مورد داده هاي صحرايي موجود، به پيش بيني ظرفيت باربري شمع هاي كوبيده‌شده در خاك هاي غير چسبنده پرداخته شده است. مقايسه‌ي مقادير ظرفيت باربري نهايي حاصل از آزمايش‌هاي صحرايي با مقادير پيش بيني‌شده از شبكه ها، حاكي از عملكرد بهتر آن ها بالاخص شبكه‌ي ويونت پيشنهادي است.

Despite numerous investigations regarding the bearing capacity of a driven pile in noncohesive soil, its calculation is a complicated trend. Any prediction from numerical analysis is highly dependent on the model adopted for modeling the soil behavior. However, setting up a realistic model to calculate the load carrying capacity of a pile is rather difficult. Most research shows that the capability (i.e. pattern recognition and memorization) of an ANN is suitable for inherent uncertainties and imperfections found in geotechnical engineering problems, considering its successful application without any restrictions. The combination of the wavelet transform theory with the basic concept of neural networks leads to a new mapping network, called neural network adaptive wavelets, or wavenets, which is proposed as an alternative to feed-forward neural networks for approximating arbitrary nonlinear functions. A wavelet network is a feed-forward neural network using wavelets as activation functions of its hidden layer neurons. In this network, both the position and dilation of the wavelets are optimized besides the weights. In one special approach of this network construction, so called the wavenet, the position and dilation of the wavelets are fixed and the weights are optimized. In this research, considering the late mentioned procedure for available experimental data, the potential for applying a neural network and its adaptive wavelets (wavenets) has been shown for predicting the load carrying capacity of a pile driven in noncohesive soil. The validation tests show that the artificial intelligence solutions clearly outperform in predictive accuracy under varying training and testing conditions. These methods can be employed for predicting the load carrying capacity of a pile driven in noncohesive soil, in comparison with other computational and time consuming methods, considering the complexity of the soil characteristics. Numerical results indicate that substituting the wavelet function as a feed-forward neural network transfer function can enhance network performance and efficiency. Therefore, the proposed wavenet with a feedforward neural network structure (wavenet), which uses the SLOG1 wavelet function as its hidden layer activation function, is much better in comparison to standard feedforward, in terms of performance generality.