Implementation issues of novelty detection technique for cable-supported bridges instrumented with a long-term monitoring system

Implementation of a sophisticated long-term monitoring system on the Tsing Ma suspension bridge in Hong Kong highlights the necessity of developing practical damage detection methodologies for large-scale civil structures. The novelty detection technique has been demonstrated to be greatly promising for damage occurrence detection of long-span bridge structures in operation with noisy measurement data. Some key issues concerning the implementation of this technique to the instrumented Tsing Ma Bridge are explored in this paper: (i) Selection of modal frequencies in constructing novelty neural network input for structural damage alarming. A novelty detection neural network using natural frequencies of only vertical modes as input vector and a novelty detection neural network using natural frequencies of mixed vertical and lateral/torsional modes as input vector are respectively constructed, and their identification sensitivities are compared; (ii) Definition of distance functions to measure difference between the input and output vectors due to anomaly. Both the Euclidean distance and the Mahalanobis distance are used to define the anomaly metric and their sensitivities to damage are identified; (iii) Construction of a novelty index capable of discerning between the novelty shift due to structural damage and the novelty shift due to inconsistent noise level. The focus of the present study is on developing an improved novelty index which uniquely indicates structural damage rather than anomaly due to change of noise level. The existing novelty index cannot distinguish tree damage from anomaly due to inconsistent noise level in training and testing stages. As a result, the sequence shift signaled by such a novelty index is definitely indicative of structural damage only when the noise level is exactly same for the training data and for the testing data in statistical sense. On the contrary, the improved novelty index developed in this study always indicates damage information no matter the training noise level is greater or less than the testing noise level.