DCNN Demodulation Method for Salinity Sensor Based on Multimode Large Misalignment MZI

• Published in: IEEE transactions on instrumentation and measurement 2024, Vol.73, p.1-7
• Main Authors: Lv, Ri-Qing, Du, Chen-Chen, Wang, Wei, Liu, Yong-Nan, Liu, Rui-Jie, Wang, Ying-Long, Lin, Zi-Ting, Zhao, Yong
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural networks; Claddings; Convolution; Demodulation; Demodulation algorithm; Feature extraction; Fiber optics; Interference; Interrogation; Mach-Zehnder interferometers; Mach–Zehnder interferometer (MZI); Misalignment; neural network; Neural networks; optical fiber sensor; Root-mean-square errors; Salinity; Salinity (geophysical); Sensors; Spectra
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2024.3353287

The large misalignment Mach-Zehnder interferometer (MZI) based on single mode fiber (SMF) is getting more attention in marine parameter measurements. Due to the existence of multiple fiber optic transmission modes in this sensing structure, traditional optical path difference (OPD) demodulation algorithms face difficulties in demodulation. Therefore, a new method to demodulate the spectra of SMF-SMF-SMF (SSS) multimode large misalignment MZI sensor using a deep convolutional neural network (DCNN) is proposed in this article. The DCNN with four convolutional layers and four max-pooling layers is established. Convolutional layers are employed to extract deep feature information from the MZI spectrum, and max pooling layers are used for feature selection and filtering. The model was trained and tested by 640 samples in total at different salinities ranging from 0‰ to 40.004‰, and the raw spectrum could be directly used without denoising. The maximum demodulation error of the model does not exceed 0.8‰, and the root mean square error (RMSE) is 0.2946‰. Meanwhile, this neural network can realize a nonlinear mapping from raw spectra to salinity and shows high potential to reduce the cost of the interrogation hardware.