Backpropagation neural network assisted concentration prediction of biconical microfiber sensors

The response of the optical microfiber sensor has a big difference due to the slight change in fiber structure, which greatly reduces the reliability of microfiber sensors and limits its practical applications. To avoid the nonlinear influences of microfiber deformation and individual differences on sensing performance, a backpropagation neural network (BPNN) is proposed for concentration prediction based on biconical microfiber (BMF) sensors. Microfiber diameter, cone angle, and relative intensity are the key input parameters for detecting the concentration of chlorophyll-a (from ∼0.03 mg/g to ∼0.10 mg/g). Hundreds of relative intensity-concentration data pairs acquired from 32 BMF sensors are used for the network training. The prediction ability of the model is evaluated by the root-mean-square error (RMSE) and the fitness value (F). The prediction performance of BPNN is compared with the traditional linear-fitting line method. After training, BPNN could adapt to the BMF sensors with different structural parameters and predict the nonlinear response caused by the small structural changes of microfiber. The concentration prediction given by BPNN is much closer to the actual measured value than the one obtained by the linear fitting curve (RMSE 1.84×10^-3 mg/g vs. 4.6×10^-3 mg/g). The numbers of training data and hidden layers of the BPNN are discussed respectively. The prediction results indicate that the one-hidden-layer network trained by more training data provides the best performance (RMSE and fitness values are 1.63×10^-3 mg/g and 97.91%, respectively) in our experiments. With the help of BPNN, the performance of the BMF sensor is acceptable to the geometric deformation and fabrication error of microfiber, which provides an opportunity for the practical application of sensors based on micro/nanofibers.