ANN-based Structural Element Performance Model for Reliable Bridge Asset Management

Bridge Management Systems (BMSs) have been developed to assist in the management of a large bridge network. Historical condition ratings obtained from bridge inspections are major resources for predicting future deteriorations via BMSs. Available historical condition ratings in most bridge agencies, however, are very limited, and thus posing a major barrier for predicting reliable future structural performance. To alleviate this problem, A Backward Prediction Model (BPM) technique has been developed to help generate missing historical condition ratings which is crucial for bridge deterioration models to be able to predict more accurate solutions. Nevertheless, there are still considerable limitations in the existing bridge deterioration models. In view of this, feasibility study of Time Delay Neural Network (TDNN) using BPMgenerated historical condition ratings is conducted as an alternative to existing bridge deterioration models. It is anticipated that the TDNN using BPM-generated data can lead to further improvement of the current BMS outcome. are: (1) it assumes condition state independency for simplicity, which means that future bridge condition depends only on the current condition but not on the historical condition, which is impractical; (2) only the overall condition ratings are presented for longterm prediction of bridges. To overcome the abovementioned limitations in existing bridge deterioration models, feasibility study of Time Delay Neural Network (TDNN) using BPM-generated historical condition ratings has been conducted. It can make use of the benefits of the BPM-generated historical condition records which in turn would lead to effective bridge asset management. 2 TIME-SERIES PREDICTION TECHNIQUES According to Bowerman and O’connell (1993), when prediction is used in decision-making, information related to events that have occurred in the past should be considered. Usually, historical data patterns are identified and are subsequently extended to future prediction. In most forecasting techniques, the prediction results therefore rely heavily on the assumption that the identified patterns from historical data will be extended to the future. In particular, time series analysis has capabilities for prediction which is an important activity due to requirement of accurate forecasts in many decisionmaking processes. 2.1 Auto-Regressive Integrated Moving Average (ARIMA) The Auto-Regressive Integrated Moving Average (ARIMA) model has been developed by Box and Jenkins (1970). This model is one of the most widely used techniques for reliable short-term predictions (Khashei et al. 2009). However, it has two significant limitations: (1) future value is assumed to be a linear function: and (2) large amount of historical dataset is required to obtain reliable predictions. A recent research in relation to Artificial Neural Networks (ANNs) indicates that neural networkbased models provide more precise predictions than the traditional ARIMA-based models. This is because ARIMA models cannot fully capture the nonlinearity and chaotic behaviour of data due to functional linearity of the ARIMA (Kohzadi et al. 1996). 2.2 Time-Delay Neural Networks (TDNNs) The neural-network based techniques have been applied to forecasting of nonlinear processes (Tan and Cauwenberghe 1999). One of such techniques, namely the TDNNs, originally designed for speech recognition (Waibel et al. 1989), has shown to produce reliable outcomes. The TDNN model, developed by Waibel et al. (1989), has a multilayer feedforward network. The hidden and output neurons are replicated across time (Wang et al. 2007). It has a similar structure as the Multi-Layer Perceptron (MLP). Each layer in TDNN has one or more neurons which are connected for information processing. Figure 1 shows the structure of a typical TDNN. It describes that some neurons receive delayed input from other neurons within the same layer. For instance, the network receives a single input from the outside and the residual nodes in the input layer receive the input from the delayed neuron on the left-side by the one unit time interval (Zhong et al 2005). Figure 1. Structure of typical TDNN (Zhong et al 2005) In this research, a sigmoid transfer function is selected because it produces a continuous value ranging from 0 to 1. A neuron in a hidden layer (n(j), n(j+1)... n(j+n)) is linked to neurons (n(i), n(i+1)...n(i+n) ) in the input layer. The connection from ni to nj has the weight wij that initially allocates a random value between 0 and 1. The appropriate weights (wij) are determined in the training stage of neural network. Input to ni is calculated using the following equation: