In order to effectively characterize the threedimensional （3D） fluorescence spectrum of gas in apple storage room, a feature information representation method based on wavelet packet decomposition coefficient was proposed. On this basis, the research on apple spoilage early warning methods was carried out. Firstly, the triangular interpolation method and Savitzky-Golar (SG) convolution smoothing were used to preprocess the spectrum data to eliminate the influence of Rayleigh scattering and ambient noise on the original fluorescence spectrum. After preprocessing, the 3D fluorescence data was converted into one-dimensional (1D) data vector, and the converted method was the corresponding emission spectrum, which was smoothly connected end to end according to the order of excitation wavelength, the 1D vector was decomposed by 3-layer sym4 wavelet packet, and the low-frequency coefficient set after the wavelet packet decomposing was extracted as fluorescence characteristic information. Secondly, partial least squares (PLS) was used to analyze the characteristic information and six physic-chemical indexes, and spectral clustering analysis (SCA) was adopted to determine the spoilage benchmark based on the results of PLS. Finally, a spoilage early warning model was constructed by using Mahalanobis distance (MD). The results showed that the fluorescence characteristic information represented method of wavelet packet decomposition coefficient was effective. With the increase of storage days, the Mahalanobis distance of the sample to the spoilage benchmark was gradually decreased, which better described the change trend of apple quality during storage, and it could realize the early warning of apple spoilage during storage.