Curvature prediction of long-period fibre grating based on random forest regression

This study proposes a long-period fibre grating (LPFG) curvature estimation method based on random forest regression (RFR) to address the shortcomings of the existing curvature evaluation method, namely, polynomial fitting; these shortcomings cause difficulty in achieving adequate model regularity and application universality. The resonant wavelength and resonant peak amplitude of the LPFG are used as input variables in this method to develop an RFR model for curvature estimation, allowing for accurate curvature prediction of the sample. The results show that the RFR-based LPFG curvature prediction model can better characterise the input–output regression relationship than back-propagation neural networks. The average R² value of the RFR model is 0.9826, and the actual measured curvature value is highly correlated with the model predicted curvature value. Compared to that exhibited by back-propagation neural networks, the RFR model exhibits higher accuracy for curvature estimation, with average values of 0.1314 and 0.1173 for root mean square and mean absolute errors, respectively. This method can provide a more comprehensive theoretical basis for the application of robot learning in the curvature measurement of LPFG and has practical value.