Novel event classification for structural health monitoring systems

This thesis reports results obtained in applying neural networks to the problem of vehicle type classification from strain measurement data such as that obtained during structural health monitoring (SHM) of a vehicle bridge. It builds upon previous work which addressed the issue of reducing vast amounts of data collected during an SHM process by storing only those events regarded as being "interesting," thus decreasing the stored data to a manageable size. This capability is extended here by providing a means to group and classify these novel events using artificial neural network (ANN) techniques. In absence of actual strain measurements from a structure in service, simulated strain responses of cars, vans, buses and large trucks passing over sensors was generated and used for training and evaluation purposes.;Three types of neural systems consisting of a combination of supervised and unsupervised learning were investigated. The first consists of 2 layers of artificial neurons using both supervised and unsupervised learning. In this system, the first layer is a feature extraction layer and the second is the event classifier. The system was able to achieve an identification success rate of 63% for a dataset containing 3001 isolated vehicle strain patterns. The second system that is investigated is an extension of the first that included an extra data preprocessing stage. In this system, input data presented to the system is first scaled to the maximum value before being presented to the first layer. The scaling factor is retained and later presented to the second layer as an extra input. This system was able to achieve a success rate of about 92% for an isolated vehicle dataset containing 3001 data patterns. It was further found that proper identification of one vehicle when two are present within a single observation period was possible, even when the strain responses are overlapping. The vehicle type selected by this system in that case corresponds to the vehicle with the highest magnitude strain signature. Modifications to the system were explored in efforts to improve recognition while removing the emphasis on magnitude alone. In doing so, a classifier was produced which selects the most consistently identified input pattern over a series of four sensors. This final system investigated is made up of sub-systems which consist of a data preprocessing stage and a two layer artificial neural network. Recognition accuracy for this system was found to be 85% for 3001 simulated vehicles. The system was found to do comparatively better than the neural classifier system with scaling for observation windows containing two vehicles.