Abstract: Tribolium castaneum (Herbst) has been one of the most destructive insects for cereals and products in the world. A rapid and accurate detection can be greatly contributed to the stored grain in recent years. In this study, an electronic nose detection device was developed to detect the pests in stored wheat grains using the feature optimization of the gas sensor array. The volatile odor information of Herbst in wheat was also collected with different population density gradients. Firstly, the response signal of the gas sensor array was analyzed and preprocessed after collection. Then, the feature extraction and selection were performed on the response curve of the sensor array, in order to obtain the feature information of the sample. A regression model was established to predict the population density of Herbst in wheat. The Relative Change value (RC), Relative Integral value (RI), and Mean Differential value (MD) of each sensor were extracted as the original feature matrix (10×3), according to the response curves collected by 10 gas sensors. A Genetic Algorithm (GA) was adopted as the feature selection to optimize the original multi-feature matrix, where the recognition accuracy was used as the evaluation index. Finally, Partial Least Squares Regression (PLSR), Principal Component Regression (PCR), and Support Vector Regression (SVR) were used for the regression prediction. The results show that the PLSR model achieved the best prediction. The Principal Component Analysis (PCA) and Cluster Analysis (CA) were used to analyze the discrimination and aggregation between different types of samples. It was found that the electronic nose device effectively distinguished the wheat samples from those infected by pests. The redundant information was removed from the optimized feature matrix, according to the recognition accuracy and PCA data. The recognition accuracy of the feature matrix increased from 82.85% to 97.14% after GA optimization. The number of feature variables was reduced by 60% from 30 to 12, and the number of sensors was reduced to 8. The PLSR, PCR, and SVR were used to establish the prediction model before and after the optimization of the matrix. The prediction set correlation coefficient (r) and root mean square error (RMSE) of the PLSR model were 0.828 and 11.293, respectively. By contrast, the prediction set r and RMSE of the PCR model were 0.764 and 13.859, respectively. The prediction set r and RMSE of the SVR model were 0.804 and 12.976, respectively. The results showed that the PLSR was the best model for the prediction of the population density of Herbst. Consequently, the feasibility and effectiveness of the sensor array multi-feature optimization can be verified for the electronic nose detection of stored grain pests. The finding can also provide a strong reference for the rapid detection of grain pests.