Northwest Helan Mountain's East Foothill is considered as one of best regions for grapes growing and fermenting. Helan Mountain's East Foothill has been divided to 5 wine origin included Shizuishan, Yinchuan, Yongning, Qingtongxia and Hongsibu. These still not establishes a scientific approach to classify and manage the wine origin because of different terroir. This study analyzed correlation and origin’s variance wine color and taste indicators producing by Helan Mountain's East Foothill, and identified the color and taste indicators relating to wine origin based on the leaves node weights of random forest. Then, the Fourier transform near infrared spectra (FT-NIR) and chemometrics methods were used to construct quantitative analysis models for each indicator. Finally, using the database composed by the predictive values of the color and taste indicators as the input layer, an artificial neural networks (ANN) model with ReLU as the optimal activation function was trained to classify the specific origin of wine from Helan Mountain's East Foothill. The research results indicated that tartaric acid esters and pH values had significant difference between the wine from different origins, and many indicators had strong correlation with other indicator. Based on the analysis of removing low weights parameters one by one, 14 wine color and taste indicators (lightness $L^{\ast }$, redness $a^{\ast }$, yellowness $b^{\ast }$, chroma $C_{ab}^{\ast }$, total anthocyanins, monomer anthocyanin, polymeric anthocyanins, ionization index, flavonol, total tannin, ethanol index, tartaric acid ester, pH value, titrable acid) were considered as relevant to the origin. The FT-NIR quantitative analysis models of all 14 wine color and taste indicators had the bigger determination coefficient (r²_c) than 0.95, the bigger relative percent deviation (RPD_c) than 5, smaller root mean squared error (RMSEC) than 5% of the mean of each indicator for calibration set; and had the bigger determination coefficient (r²_v) than 0.9, the bigger relative percent deviation (RPD_v) than 2.5, smaller root mean squared error (RMSEV) than 15.7% of the mean of each indicator for validation set. The FT-NIR models had good quantitative prediction ability for 14 wine color and taste indicators. The 3 kinds of ANN models were established with different activation function include sigmoid, tanh and ReLU. The accuracies for determining of the origin of dry red wines were 84.06%, 91.30% and 94.20% separately. The model with ReLU as the activation function was proved to be the best one. Further analysis shows that the best model has 100% sensitivity and 100% accuracy in classifying the Shizishan wine samples, 100% sensitivity and 90% accuracy in classifying the Yanchuan wine samples, 87.5% sensitivity and 93.33% accuracy in classifying the Yongning wine samples, 94.74% sensitivity and 100% accuracy in classifying the Qingtongxia wine samples, 92.31% sensitivity and 92.31% accuracy in classifying the Hongsibu wine samples. The results indicated that the model proposed in this study can accurately discriminate the specific origin of wine from the Helan Mountain's East Foothill. This study would provide technical support for the product zoning management of dry red wines in the Helan Mountain's East Foothill, and help the establishment and development of Chinese wine origin protection system.