Abstract

For existing concrete bridges,maintenance work plays a significant role in keeping the structures safe.However,the maintenance work mainly depends on periodic inspection[1-3],which usually consumes time and human resources.Consequently,there is a growing need for adequate maintenance management to reduce time and human resources.An artificial neural network is extremely efficient with large data sets and would provide a more accurate prediction compared to traditional statistical methods[4].Therefore,in this paper,we try to apply an artificial neural network to examine its accuracy for deterioration prediction.

An artificial neural network has been applied in the field of civil engineering[5],which is mainly divided into two targets:for laboratory experiments[6,7]and for practical engineering[8].In laboratory conditions,many researchers have studied the prediction of performance indicators of concrete,such as compressive strength[6,7]and chloride dif fusivity[9].Some researchers have also applied an artificial neural network for the mixture design of high-performance concrete[10,11].For such applications,many influential factors were investigated,and the ef fectiveness of an artificial neural network was discussed.Even though the use of an artificial neural network for laboratory experiments has profound significance,it is challenging to consider all influential factors for the application because the accelerated experimental conditions usually used are different from actual conditions on sites;in other words,the experiment is tough to simulate the real situation.Typically,the utilization of neural network in experimental conditions is deviated from the actual terms and only considered a few or even only one influential factor.

Some researchers also have applied an artificial neural network or other methods to inspection[8,9,12],evaluation[13,14],retrofit[15,16],and scheduling of repair activities[17].However,most of them neither have long-term practical inspection database nor use traditional expert-score system.

Predictions of deterioration progress have been applied to concrete bridges.However,it is still unclear how the actual conditions influence the overall deterioration of bridges.In this paper,we used an inspection database and an artificial neural network to develop a model to predict deterioration progress of concrete bridges.The database used has been set up by engineers during the longterm maintenance work.To build the model,we considered eight potential environmental factors,such as temperature,traffic volume,and years in service.Moreover,two bridge geometric parameters-length and width are also discussed.The output of the model is the overall deterioration grade of each bridge,which is classified into four grades as per the Inspection Guidelines[1].Their definitions are Grade a:healthy,Grade b:preventive action required,Grade c:early action required,and Grade d:emergency action required.

Following the flowchart shown in Figure 1,a deterioration prediction model for concrete bridges was developed.At first,we eliminated incomplete data and unreasonable data from the database.Then about 70%of bridge inspection data from all the dataset was used to train the neural network model,and 50% of the remaining dataset was applied to validate the performance of the model.Based on the training and validation sets,the overall performance of the model was adjusted.After the best performance model was got,the remaining 50% data was used to test the model.

Then,the iterated calculation was carried out to finally create the model having the best performance.In the process of this calculation,suitable sets of model parameters were obtained.Finally,by using these model parameters and the best performance model,it can be realized to make prediction of deterioration progress.

Figure 1　Flow chart of setting up the model

By applying the model to existing bridges,it was found that the prediction results show the accuracy is about 74%,as indicated in Table 1.In the table,the shaded areas represent the prediction numbers in different conditions,and the darker areas indicate the successful prediction numbers.The accuracy was calculated by all the successful predictions(20+22+30)divided by the total number(96).Even though there are still some works needed to improve the model,the applicability of the preliminary prediction model was verified for practical maintenance work of bridges.This model could be a supplementary method to help us to find appropriate timing of further inspection as well as interventions.

Table 1　The performance of the model