Gaussian Process-Based Particle-Filtering Approach for Real-Time Damage Prediction with Application

A prognostic model capable of predicting temporal damage evolution is essential to prevent catastrophic failure of structures. Data driven techniques, such as neural networks and support vector machines, are widely used for prediction of damage in a variety of aerospace and civil applications. Most of the available techniques cannot be applied for real-time prediction because they assume the measured value to be the true value, which is often not true. They also require training data from a similar set of experiments based on which predictions are made, which may not always be available. In this paper, the authors propose a novel integrated approach that intelligently combines particle filter updating with a fully probabilistic Gaussian process model to predict complex physical phenomena (e.g., temporal local pier scour) considering both measurement and prediction uncertainties. In this example, the measurement model is obtained using radio frequency identification (RFID) sensors and the state space model is the Gaussian process-based prognosis model. The performance of the algorithm was tested using corrupt training data. Different scenarios are presented with application to predicting local scour near bridge piers, which is a highly stochastic phenomenon with training data mostly unavailable. The algorithm is used to make predictions using corrupt training data. The results indicate that the proposed approach predicts the scour depth accurately under varying field conditions.