Study on intelligent method of prediction by small samples for ground settlement in shield tunnelling

Abstract

The shield tunnelling in shallow soft ground would bring a risk of damage to the structures or pipelines in dense populated urban areas. The prediction of the ground settlement could reduce the risk and maintain normal tunnelling by effectively controlling parameters in the construction. The criteria need to be studied for all kinds of buildings and underground pipelines, one of that is normally limited to be from -3 to 1 cm in the ground settlement in Shanghai. The neural networks have the strong ability of non-linear mapping, which make it possible to establish a model directly from the observations. It is also convenient in practice to consider synthetic effects of both soil indexes and construction parameters. For the ground settlement could be regarded as a synthetic parameter reflecting the construction process and strata information, the inputs and outputs could only simplify to be sequential settlement. The inputs are detailed as S(N-2), S(N-1) and S(N), those are the observations of settlement in two days before, one day before and current day. The output parameter is only the settlement in the next day (S(N+1)). Currently, the structure design of existing neural networks for predicting settlement has not been solved well. In this paper, an intelligent evolutionary neural network (ENN) is proposed for predicting the ground settlement. The mechanism allows all aspects of the network structure, including the number of hidden nodes (N1 and N2) and learning parameters, to be evolved through genetic algorithms. Meanwhile, the pre-processing method of samples is improved. After the ENN is obtained, the ground settlement in the next advance step could be predicted. The observations are compared to the predictions, and standard deviation d is got. If the d is greater than a preset limit in processing, the primary sampling set is refined and a new ENN is achieved then. Otherwise, the prediction is continued while d is in limits. The ability to learn from relatively few training examples is important because training examples are often hard to find or expensive to collect. Thus, it is important for a learning system to be able to extract useful generalizations from a small set of examples. The ENN model also showed good ability in prediction trained by small samples. A primary use is made to the cross-over tunnel in down line zone near Nanpu Bridge Station of the Second phase of Shanghai Metro Pearl Line. The testing results showed a good predicting performance by using actual collected data from the project, with maximal average error 2.08% for training samples and -5.27% for testing samples. The hogging or sagging curves in the longitudinal direction are also drawn for analysis. The method would be especially valuable to the tunnelling by DOT (double-O-tunnelling) method under condition of large span and thin cover of earth, which may result in greater impacts to the structures or pipelines nearby. (A) Reprinted with permission from Elsevier. For the covering abstract see ITRD E124500.