Truck weigh-in-motion using reverse modeling and genetic algorithms

The ability to accurately determine the loading attributes of a truck (namely the axle configuration, the spacing between the axles, and the load imposed by each axle) while it is in motion is an important function for the design and structural health monitoring of bridges, and highways. Truck weigh-in-motion (WIM) as it is termed is an inverse problem where the load is identified from the observed response of the structure over which it is travelling. The problem has been reasonably well solved using neural network techniques, but there is still significant room for improvement in terms of reducing the number of misclassifications of trucks and increasing the precision of the axle spacing and load estimates. The problem can be formulated as an optimization problem. Genetic algorithms (GAs) are proven robust and efficient search optimization techniques. The potential of the GA approach for reverse identification of axle configuration and loading from bridge girder stress envelopes has been investigated and compared to an existing neural network solution. The investigation is a pilot study that considers a simply supported steel girder bridge with a concrete deck. The bending stresses of the bridge are simulated numerically and are used as the input for reverse modeling. The identification procedure is carried out using GAs by minimizing error between the measured bridge response and reconstructed bridge response. The performance of the GA depends on the tuning of genetic operators, hence different operator settings are considered and tuned for optimality. Advance strategies such as migration and multiple species with real coded representation variables are adopted to improve the performance. The effect of measurement parameters such as sampling frequency (50-400 Hz), levels of noise (5-25%), time varying load and measuring sections on accuracy of identification are also investigated. The performance of the GA approach is found to outperform the existing neural network solution. The significance of this is that, unlike the neural network approach, the GA solution can be applied to any bridge configuration for which a reasonable stress model exists. Moreover, the computational time for the GA is found to be on average 3-4 seconds which, although is several orders of magnitude slower than the neural network solution, it is well within what could be considered an acceptable delay for generating a solution.