Temperature demodulation for optical fiber F-P sensor based on DBNs with ensemble learning. (July 2023)

• Record Type: Journal Article
• Authors: Wang, Lixiong; Liu, Hanjie; Pan, Zhen; Xu, Ye; Fan, Dian; Zhou, Ciming; Li, Yuan
• Is Part Of: Optics & laser technology. Volume 162(2023)
• Publication Date: 2023-07
• Subjects: Deep belief network -- Ensemble learning -- Temperature demodulation -- F-P sensor; Optics -- Periodicals; Lasers -- Periodicals; Electronic journals; 621.366
• DOI: 10.1016/j.optlastec.2023.109275
• Languages: English
• ISSNs: 0030-3992

Abstract

Highlights: The demodulation performance of traditional methods is limited in temperature environment because the thermo-optic coefficient and the coefficient of thermal expansion are set to a fixed value. The learning ability of individual demodulation model is insufficient. A deep belief networks (DBN) has a strong feature extraction capability and powerful nonlinear mapping ability, which shows the great potential in signal demodulation. Ensemble learning is a learning approach that demodulation results are combined by multiple learning algorithms to improve the demodulation performance of single DBN model. Abstract: Aiming at the problem of traditional temperature demodulation due to ignoring the irregular change of the thermo-optical and thermal expansion coefficients, a demodulation method based on deep belief networks (DBNs) with ensemble learning is proposed. The DBN can establish a nonlinear mapping model between spectrum and temperature to learn the information that the thermo-optic and thermal expansion coefficient are changed with temperature, the accurate temperature demodulation can be realized. In order to improve the learning ability of the individual model, a stacking ensemble method is utilized to learn the knowledge of five DBN models with different activation functions. The proposed method achieves the demodulation performance that the demodulation precision is 0.30%F.S., and the mean absolute error is as low as 0.98 °C in the range of temperature from 30 °C to 1100 °C. It shows that the proposed method can promote the learning ability of algorithm to improve the demodulation performance of single DBN model and make F-P high temperature sensing system more accurate and reliable.