A Prediction Model Of Concrete Strength Based On Artificial Neural Network

Prediction of 28 day's concrete strength, which is very significant to increase the efficiency of concrete production, is a typical multi-variable, nonlinear system. The traditional methods of the prediction are difficult to be applied widely in practice because its accuracy is lower. So artificial intelligence algorithm is applied to the prediction of concrete strength in recent years, such as neural network algorithm. We can get more ideal results by it. There are two kinds of neural network models to predict concrete strength, which are applied extensively now. One kind of neural network model of prediction is using the neural network directly, and another kind of neural network model of prediction is based on statistics.The first kind of neural network model of prediction is not perfect, because it does not pretreat the experimental data firstly. The neural network prediction model based on statistics has two shortages. First, when relativity between original data is small, we can not reduce the data's dimension efficiently by using principal component analysis approach, and the structure of the neural network model can not be simplified efficiently. Second, principal component analysis is a linear algorithm, and it can only extract the linear characteristic of data. If there is nonlinear relationship between each dataset, principal component analysis can not extract the characteristic of the data efficiently. Some information of the original data will be lost and not be used efficiently. It will influence the precision of the model.This paper will pretreat the sample data by using nonlinear principal component analysis approach based on neural network, and set up a prediction model of concrete strength by using nonlinear principal component analysis approach and neural network. The nonlinear principal component analysis approach is an advanced principal component analysis approach. It can extract the linear and nonlinear characteristic of the data more efficiently. A 3- hidden-layer neural network model is most extensively used to analysis the sample data at present. But the model has some shortages as follows. First, the structure of the model is so complicated, that it will take a lot of time to train the network. Second, it needs a lot of sample data, and thetrained network is unstable when there are not enough sample data and some noisy datasets in the sample data. Third, the principal components that we get are non-unique when the numbers of encoding layer neuron are different. So this paper will propose a simplified neural network model, and the shortages will be alleviated. Owing to the slow convergent speed of the traditional BP algorithm, we propose an advanced BP algorithm, which can converge faster. The neural network in the paper will be trained by the advanced BP algorithm. We should specially point out that, all of the prediction models of concrete strength established in this paper with different numbers of bottleneck layer neuron can converge below twenty seconds, and the average prediction errors are all below five percent by the improvement of original sample data pretreatment method. We can get a neural network model with a more simple structure, faster convergent speed and more accurate predictive results.