Study On Neural Network Method For Predicting Deformation Of Foundation Pit

For the importance of prediction in pit deformation, after the analysis of existing prediction methods and pit deformation, the prediction using neural network is focused on. And based on the relevant research work of prediction with neural network, the feasibility that combined neural network is introduced to predict pit deformation is proposed.With neural network tool of MATLAB, the bootstrap and BP-RBF combined neural networks are realized by self-compiled program. To bootstrap combined neural network, its specific content is that a weight item with coefficient of ? is added to the half of performance function named mse. This content makes every sub-network have different performance function and realize the prediction difference of internal sub-networks, by which the generalization ability of the composite structure is ultimately enhanced. To BP-RBF combined neural network, different prediction characteristics of BP and RBF network are treated as starting point. Firstly, the prediction of pit deformation with the two networks is carried out separately. Secondly, the predictive values of the two networks are entered into a new RBF network as a sample. Finally, the deformation prediction is carried out again, and a good result is achieved.By the specific training process of the two combined neural networks mentioned above, it's found that there is a very close relationship between sample quality and specific training process. When it comes to the representative ability of sample firstly, it's needed that both training samples and test samples contain the similar mapping relationship. If the training samples don't contain certain information in the test samples, the prediction then there will be a great deviation. The second problem is noise of samples. When the samples contain a large noise, a good training effect will be gotten by the trainlm training function, which learning ability is very excellent, however, the prediction to test samples is not very good. That means the noise is also stored in the network weights, which, in other words, is "over generalization". Then, compared to the trainlm function, a better result could be gotten by the traingdx training function which training process is somewhat moderate.Finally, in the field of deformation forecast of foundation pit, the further application prospect of combined neural network is discussed.