Development and Analysis of Multifeature Approaches in SPR Sensor Development

Abstract

This paper presents the development and signal analysis of surface plasmon resonance (SPR)-based sensors in D-shaped polymer optical fibers (POFs). A gold-palladium (Au-Pd) coating was applied to the D-shaped region to obtain the SPR signal in the transmitted spectrum of the POFs, where different samples were fabricated using the same methods and parameters. In this case, the transmitted spectra of three sets of samples were compared, which indicated variations in the SPR signature that can influence the sensors’ application and reproducibility. Then, the intensity of and wavelength shift in the SPR signals were analyzed as a function of the refractive index variation, where it was possible to observe differences in the sensors’ sensitivities and the linearity of the different samples. In this regard, additional features, namely the area below the curve and the peak amplitude of the fast Fourier transform (FFT) applied to the transmitted spectra, were used to enhance the sensors’ accuracy and precision. To verify the use of such additional features in the sensor analysis, an unsupervised approach based on k-means clustering was used considering a single dataset with the results of all the sensors. The results showed clustering with the number of different refractive indices tested, which motivated the use of these features (intensity, wavelength, area and FFT amplitude) in the refractive index assessment. In this context, random forest was the supervised algorithm with the smallest root mean squared error (RMSE) among the algorithms tested, where an RMSE of 0.0057 was obtained considering all the datasets. For the analysis of each sensor (considering the three sets of sensor samples), the mean RMSE using random forest applied to the multifeature approach returned relative errors below 9%, considering the entire tested range of refractive index variation.