Single-Polarization Fiber Current Sensing With Error Correction Using an Optimized Constant-Term-Driven Neural Network

For fiber current-sensing approaches, optical path simplification and measurement error correction are two key points that enhance the measurement performance. This article presents a highly simplified single-polarization structure with error correction using an optimized constant-term-driven neural network (CTNN). To achieve optical path simplification, the adopted structure modulates light with minimal devices: a single-polarization circulator and a Faraday rotator mirror. The circulator receives and polarizes the light reflected by the mirror for optical modulation. In the CTNN, the temperature-dependent constant terms from the sensing output are fed into the neural network used for adjusting the sensing output to lower the measurement error over a wide temperature range. Also, the improved Newton-Raphson algorithm (INRA) with diverse solution distribution is used to enhance the fitting performance of the CTNN by optimizing weights and biases. Lastly, compared to existing fiber sensing approaches, the proposed approach reduces the number of optical devices from 4 to 2, achieving the simplest optical path. On this basis, the test data show that the measurement error is reduced from 1.3% to 0.56% by the INRA-assisted CTNN, compared to 0.97% achieved by other optimized neural networks.