CNN-based multiplexed optical fiber sensors for multi-load mapping on 2D structures

Abstract

This paper reports the application of a Convolutional Neural Network specifically designed for one-dimensional optical signal processing to determine the magnitude and position of loads acting on a structure instrumented with multiplexed macrobend optical fiber sensors, offering a cost-effective and low complexity alternative for force monitoring solutions. The system effectively localized and quantified one, two, or three simultaneous loads within 16 distinct sensing areas, utilizing only five in-series optical fiber sensors, which were spectrally interrogated in transmission mode. The monitored loads ranged from 1000 to 2000 gf. The use of the Huber loss function allows the model to adaptively predict values associated with regions with or without loads. Experimental results showed an average mean absolute error of 224±65 gf during testing. By applying a straightforward post-processing method for load presence detection in each region, the system achieved average F1 scores ranging from 0.84 to 0.93 across the monitored regions, and an average Hamming score of 0.93. These findings demonstrate the system’s effectiveness in monitoring multiple loads, underscoring the potential of optical fiber sensors and CNNs in sensing applications.