Active Probabilistic Fast Kernel Extreme-Learning Machine for Data-Efficient Bridge Condition Prediction

Bridge condition assessment in many countries, including the US, relies on routine inspections of all bridges at predetermined intervals, costing millions of dollars each year. The significant cost of this approach, along with the availability of historical data on bridge deterioration and condition assessment, has motivated researchers to develop novel methods for customized inspection planning based on the likelihood of deterioration using bridge condition forecasting models. Substituting the routine inspection policy with prediction-based planning methods will generate a significantly lower amount of data because not all bridges will be inspected routinely. Therefore, a practical forecasting model must have an efficient training mechanism that can maintain its predictive power using a considerably limited number of observations. In addition, the forecasting model must be equipped with a built-in mechanism to selectively choose the most informative data points for retraining and updating the model in order to preserve its predictive power and prevent biased prediction. None of the existing bridge condition forecasting models offer such capabilities. Therefore, our study introduces a data-efficient bridge condition forecasting approach by creating a method based on an active probabilistic fast kernel extreme-learning machine model. We implemented and empirically evaluated the performance of the proposed method using the National Bridge Inventory (NBI) data for highway bridges in the state of New York, which include a total of 13,274 bridges. The results indicated that the proposed method achieved very high predictive power with a ranked probability score (RPS) of 0.01 using a very small training data set consisting of information from only 70 bridges.