Machine-learning-assisted omnidirectional bending sensor based on a cascaded asymmetric dual-core PCF sensor

An omnidirectional bending sensor comprising the cascaded asymmetric dual Core PCF (ADCPCF) is designed and demonstrated experimentally. By cascading and splicing two ADCPCFs at a lateral rotation angle, the transmission spectra of the sensor are highly dependent on the bending direction. Machine learning (ML) is employed to predict the curvature and bending orientation of the bending sensor for the first time. The experimental results demonstrate that the ADCPCF sensor combined with machine learning can predict the curvature and omnidirectional bending orientation within 360° without requiring any post-processing fabrication steps. The prediction accuracy is 99.85% with a mean absolute error (MAE) of 2.7° for the bending direction measurement and 98.08% with an MAE of 0.03 m-1 for the curvature measurement. This promising strategy utilizes the global features (full spectra) in combination with machine learning to overcome the dependence of the sensor on the high-quality transmission spectra, wavelength range, and special wavelength dip in the conventional dip tracking method. This excellent omnidirectional bending sensor has large potential in structural health monitoring, robotic arms, medical instruments, and wearable devices.