Robust one-dimensional phase unwrapping algorithm based on LSTM network with reduced parameter number

Abstract

For interferometric optical fiber sensor applications, one-dimensional (1D) phase unwrapping is one of the most intractable steps. It is generally more vulnerable to noise impact compared with two-dimensional (2D) phase unwrapping, since the data correlation for the 1D phase unwrapping is very limited. In a recent work, we proposed the 1D phase unwrapping algorithm by combining quasi-Gramian matrix and deep convolutional neural network (DCNN). The obtained result is very robust and can work very stably under low signal-to-noise ratio (SNR) level. But the required parameter amount is large being about 10⁷ making itself inadequate for resource limited computation platforms. In this work, a lightweight phase unwrapping algorithm based on long short-term memory (LSTM) network is proposed, which utilizes only 10⁶ parameter amounts, being 1 order smaller than the DCNN. Simulation results demonstrate that in the SNR range of 0 to 16 dB, the LSTM-based method shows comparable or better performance compared with the DCNN method, and for the very low SNR of -2 dB, only small performance decrease is observed. The generalization ability of the proposed method is also verified by using the experimental data collected from an actual phase-sensitive optical time domain reflectometry (ϕ-OTDR) system.