Bending classification from interference signals of a fiber optic sensor using shallow learning and convolutional neural networks

Bending monitoring is critical in engineering applications, as it helps determine any structural deformation caused by load action or fatigue effect. While strain gauges and accelerometers were previously used to measure bending magnitude, optical fiber sensors have emerged as a reliable alternative. In this work, a machine-learning-based model is proposed to analyze the interference signal of an interferometric fiber sensor system and characterize the bending magnitude and direction. In particular, shallow learning-based and convolutional neural network-based (CNN) models have been implemented to perform this task. Furthermore, given the repeatability of the interference signals, a synthetic dataset was created to train the models, whereas real interferometric signals were used to evaluate the models’ performance. Experiments were conducted on a flexible rod in fixed–free and fixed–fixed ends configurations for bending monitoring. Although both models achieved mean accuracies above 91%, only the CNN-based model reached a mean accuracy above 98%. This confirms that monitoring bending movements through interference signal analysis by means of a CNN-based model is a viable approach.