Neural networks-based convenient fiber optic sensing system for vehicle classification

Abstract

Accurate vehicle classification data are essential for effective road maintenance. However, obtaining such data poses challenges due to disruptions from sensor installations, high costs, and the difficulty in distinguishing between passenger cars and two-axle trucks. To address these issues, this paper presents a novel and convenient fiber optic sensing system for vehicle classification based on the principle of microbending loss. The system can be easily installed using a surface adhesion method, without disrupting traffic. The classification process comprises system deployment, data collection, data processing, and neural network-based classification. Advanced data processing techniques, including moving average smoothing, an enhanced valley detection algorithm, speed calculation, and axle spacing calculation, are employed to extract vehicle features. A three-layer neural network model is developed to enhance classification accuracy, particularly improving the differentiation between two-axle cars and two-axle trucks through the incorporation of axle load factors. In evaluations conducted on a single-lane road, the system demonstrated a high degree of accuracy, achieving an overall recognition rate of 99.4 % for eight types of vehicles. This innovative system provides a reliable source of traffic and axle load data, supporting road maintenance management, road design, overload monitoring, traffic control, and non-stop toll collection.