Neural Network Novelty Filtering for Anomaly Detection

A sophisticated long-term monitoring system is devised by the Hong Kong SAR government authorities to monitor the structural health and performance of the suspension Tsing Ma Bridge and the cable-stayed Kap Shui Mun Bridge and Ting Kau Bridge. The sensing system comprises about 900 permanently installed sensors including accelerometers, strain gauges, displacement transducers, anemometers, level sensors, temperature sensors and weigh-in-motion sensors. Based on this system, the Hong Kong Polytechnic University is devoted to developing a neural network based damage alarming system by use of the measurement data of modal properties, strain and fatigue properties, static displacement and deflection, and cable static force respectively. This paper describes how to establish a neural network novelty filter for anomaly detection of Tsing Ma Bridge cables from the measured multi-mode frequencies of the cables. An auto-associative neural network is developed for this purpose. Only a series of measured modal data in structural healthy state are required in training the neural network. Another series of measured modal data in testing phase are fed into the trained network to obtain a novelty index sequence that signals whether anomaly occurs. A salient feature of this approach is that it needs neither structural model nor damage model in both training and testing phases. Moreover, it is intrinsically tolerant of measurement noise and uncertainty in ambient conditions. Numerical simulation using noise-corrupted analytical data of the Tsing Ma Bridge cables is performed to demonstrate the effectiveness of the proposed novelty filter and its robustness to measurement and structural uncertainty.