Real-time vehicle identification using two-step LSTM method for acceleration-based bridge weigh-in-motion system

Recently, accelerometers have been employed for bridge weigh-in-motion (BWIM) systems to provide more durable field measurements comparing with conventional strain-based sensors. As the basis of BWIM system, accurate vehicle identification provides fundamental support for vehicle loads monitoring and overweight traffic detection. However, research efforts on axle recognition in real time are still inadequate, especially for accelerometer-based BWIM system. In this paper, we propose a two-step solution for real-time vehicle identification designed for acceleration measurements. In this method, a sequence-to-label long-short-term memory (LSTM) network is constructed to identify axle-induced responses in a multi-lane system directly. The input sequence is wavelet coefficients after performing wavelet transform on the raw data. Based on the trustworthy axle identification results, an auto-grouping step is then proposed and applied for vehicle-type identification. Model training and method evaluation are conducted using filed measurements from a highway bridge in Tokyo. Two data sets are utilized, i.e., 191 vehicles with 456 axles and 596 vehicles with 1380 axles. Results show that 98% axles can be identified correctly using proposed LSTM method from both data sets, while accuracy of vehicle-type identification is 96% for both data sets, which can demonstrate the robustness of proposed methods. Moreover, the driving lane detection of all detected vehicles is 100% without any failed cases. Comparing with all-in-one deep network using acceleration measurements as input sources directly, the proposed two-step LSTM method requires less training data, hence it is a computationally efficient solution, which would enable its generalization capability for applying on other bridges.