Abstract:Total volatile basic nitrogen（TVB-N）is an important indicator for evaluating the freshness of stinky mandarin fish.The existing methods for detecting TVB-N，however，are generally time-consuming and in a destructive manner.Gas-phase Fourier transform infrared spectroscopy was employed to obtain the spectral information of volatiles from stinky mandarin fish under different storage conditions，so as to achieve the rapid and non-destructive testing of TVB-N.The spectral data were then pre-processed by Gaussian filtering，robust locally weighted regression（RLWR），wavelet threshold denoising，and simulated annealing-partial least squares（SA-PLS）.The correlation model between preprocessed spectra and TVB-N was built by partial least squares regression and support vector regression.Finally，the optimal prediction model was established by RLWR combined with SA-PLS，and its decision coefficient（R2p）and relative prediction deviation were 0.942 8 and 4.005，respectively.The results showed that the established method with high accuracy and robustness provided important theoretical support for industrial online detection.