Multimode Optical Fiber Specklegram Pressure Sensor Using Deep Learning
Mohammad Istiaque Reja, Darcy L. Smith, Linh V. Nguyen, Heike Ebendorff‐Heidepriem, Stephen C. Warren‐Smith
Abstract
Optical fiber pressure sensing is of significant interest for industrial process monitoring and acoustic sensing. However, the direct detection of pressure changes along an optical fiber with minimal crosstalk is a significant challenge due to the low stress-optic coefficient of glass and interferences from temperature, bending, and strain. Here, we demonstrate that deep learning combined with microstructured optical fiber specklegram sensing is an effective approach to overcoming environmental crosstalk for the challenging application of pressure sensing. Specklegrams created from multimode interference are inherently sensitive to both the measurand of interest and other environmental perturbations. By employing a deep neural network, namely a multilayer perceptron, we show that the environmental crosstalk of the specklegram-based pressure sensor can be mitigated. Furthermore, we demonstrate the practical implementation of the machine learning-based pressure sensor where we need to continuously update the model. We show that the network can learn and reject environmental disturbances and predict pressure values correctly through continuous training updates. This technique approaches the instrument error of our calibration pressure gauge, with an error below 3 kPa over a range of 1 MPa.