Research On Mechanical Properties And Early Prediction Of Fly Ash Concrete

With the development of the socio-economic and people's requirement of green environment, a lot of mineral additive and admixture were added in the concrete, especially the large-scale utility of fly ash became the focus. The early quality control of fly ash concrete became more and more important; the traditional methods of the prediction were difficult to be applied widely in practice because its accuracy was lower. So artificial intelligence algorithm was applied to the prediction of concrete strength in recent years, such as neural network algorithm, better ideal results could be got by us. In view of this situation in this article, the paper made a research on early estimating of fly ash concrete strength.Firstly, the paper through a pilot studies of the mechanical properties of fly ash concrete, experiment results provided good support for the theory and data reference of the latter part of the research work, The main work to do was the following: The performances of workability, compressive strength, bending strength and slump were studied in this paper, with different water binder ratio and different fly ash proportion, learned the general change in the law. And then, analysised the influence for fly ash concrete strength by water-cement ratio, the amount of cement, fly ash and so on, with the tool of grading principle. In order to make sure the main factors that influence fly ash concrete strength.Secondly, the paper expounded the theory of neural networks and the Program, got the date from experiment, established BP network prediction model with MATLAB, and compared its test results to actual results. Furthermore, established the presumption mathematical model with application of multiple linear regression analysis, and compared the model's results to the results of the neural network. The results showed that: the date that predicted by BP network prediction is more accurate and less discrete. But there are still some flaws in BP network model: 1. Training needs a large number of samples, the network can not be well trained when the samples'number is less, the power of network generalization is weak; 2. The learning speed of BP algorithm is slowly, due to the nature of BP algorithm for gradient descent, but the objective function that it will optimize is very complex, that is the reason of why BP algorithm's efficiency is low; 3. From the mathematical point of view, BPalgorithm is a local search optimization method, However, it must be resolved to solve the complex problem of non-linear function of the overall situation of extreme value, Therefore, the algorithm is likely to fall into local maximum, so that the failure of training.