Wheat is one of the main strategic stored grain varieties in China. But wheat is susceptible to fungal infection, which affects its safety as food. Early detection of harmful fungal infection in wheat is the precondition to control its hazard. However, current methods for mold detection, such as plate counting and fluorescence staining, are usually laboratory-intensive and time-consuming, and can not fulfill the need for on-site testing. Concerning this issue, this work intends to apply array fiber spectrometer series and chemometrics to establish an on-line method for detection of wheat mildew, and to provide a reference for further development of on-line sensing instruments for grain quality and safety. Sterile wheat kernels were inoculated with 5 different spore suspensions of fungal strains respectively, which were F. moniliforme 83227, F. proliferatum 195647, F. nivale 3.503, A. parasiticus 3.3950 and A. ochraceus 3.3486. Wheat samples were then stored at 28 ℃ and 85% relative humidity after inoculation to accelerate mildew process. At storage stage of 0, 1, 3, 5 and 7 d, Vis/NIR spectra of samples were collected by an on-line sensing system which was mainly conposed of an array fiber spectrometer and a diffuse reflectance probe. Total colony count of samples was also determined according to plate count method. Spectra of samples were measured at moving speed of 0.15 m·s-1 with integral time of 20 ms. Each sample was collected three times and the acquisition band ranged from 600 to 1 600 nm. Then original spectra of samples were firstly pre-processed by smoothing, multivariate scatter correction and derivative transformation to eliminate spectral noise. Subsequently, principal component analysis (PCA) was used to discriminate wheat samples with different mildew degrees (storage stage). Finally, liner discriminant analysis (LDA) and partial least squares regression (PLSR) were employed to develop qualitative and quantitative analysis models for fungal infection in wheat. Wheat samples undergone three stages during the storage period according to colony counts, which were not mildew, mildew and serious mildew. Analysis of original and second derivative spectra indicated that fungal infection resulted in significant changes in wheat spectra. PCA results showed that there was a certain trend of separation between wheat with different mildew degrees. The overall recognition rate obtained by LDA for the classification of samples with different mildew degrees was more than 90.0%. Colony counts in wheat samples was predicted by PLSR and coefficient of determination for the prediction set (R2p), root mean square error of prediction (RMSEP) and residual predictive deviation (RPD) value obtained were 0.859 2, 0.401 Log CFU·g-1 and 2.65, respectively. The combination of array fiber spectrometer series and chemometrics is feasible for on-line detection of wheat mildew. In further studies, natural infected wheat samples and samples contaminated with more representative fungal strains should be incorporated to enhance the robustness and applicability of the calibration model.