Machine learning approach in multi-channel fiber-optic SPR sensors

Abstract

Fiber-optic surface plasmon resonance (SPR) sensors have been increasingly used due to their advantages such as compact size, stable physical and chemical properties, high sensitivity, and resistant to electromagnetic interference. In this paper, a dual-channel side-polished fiber-optic SPR sensor system is established to detect the refractive index of liquids and the liquid temperature. High sensitivity of ∼2000 nm/RIU (refractive index unit) and linear sensitivity of ∼−2.0 nm/°C is achieved for two-channel SPR sensors, which are integrated for simultaneously RI temperature sensing and easily scaled to multi-channel SPR sensor array. The resonant wavelengths of the two sensors are located in the range of 600–700 nm and 700–800 nm respectively, which could avoid overlapping the resonance dips. Then, we propose an approach using five machine learning algorithms to predict the resonant wavelength of the second channel of SPR temperature sensor using the normalized transmission spectrum data with a wavelength range of 760–889 nm. After error analysis, it is found that the Categorical Boosting (CatBoost) is the best among five algorithms in terms of accuracy and resolution. The results provide clear evidence that accurate SPR resonant wavelength can be obtained with machine learning approach using a more compact spectrometer.