Training a MobileNet CNN for simultaneous sensing of temperature and bending using experimental fiber specklegrams

Multimode optical fibers (MMFs) can be used as sensors due to their high sensitivity to chemical and physical perturbations, such as temperature and bending. It has been shown that, by training convolutional neural networks (CNNs) wiht the speckle patterns captured at the end of the fiber, high prediction accuracy for individual perturbations can be achieved. In this work, a two task CNN trained on experimental data is proposed to predict temperature and bending simultaneously. Data acquisition is automated via a MATLAB script controlling a contact heater for temperature and a linear displacement motor for bending. Temperature is varied from 27°C to 200 °C twice per bending value across 5 fixed values, producing 20000 specklegram images acquired at the fiber output. We adapted for this work the MobileNet CNN, because of its lightweight architecture. The last dense layer has been modified to produce two outputs from a single specklegram: a continuos regression for temperature, and a classification with five classes for bending. For evaluating the performance of the model, we calculate R^2 score for the linear temperature regression, and classification accuracy for bending. The model achieves an R^2 of 0.9888 for temperature and an accuracy of 100% across the five bending classes. These results show that the trained model is capable of simultaneously predicting temperature and classifying the bending of a fiber sensor from a single specklegram with high accuracy.