Neural networks and principal components analysis for strain-based vehicle classification

A large database has been acquired and compiled of vehicles crossing over a simply supported bridge deck system. Over the course of 1.5 years, deck strains caused by traffic, along with time-synchronized video images have been archived (400,000 records). Herein, this dataset is presented and used to develop a strain-based vehicle classification approach, as a machine learning application. To achieve this goal, the principal components analysis technique is applied to extract essential features from the strain time histories. Using these features as input, a two-layered back-propagation neural network is built and trained to sort vehicles into five classes. In this regard, availability of the video images provides essential information for developing the needed labeled datasets. The trained network is tested, and satisfactory results are achieved, showing viability of the classification approach for this bridge deck system.