In this research, an embedded metal oxide semiconductor (MOS) electronic nose (e-nose) was designed to detect Chinese pecan quality. To improve the performance of e-nose, three types of features were extracted to form initial feature matrix, including mean-differential coefficient value, stable value, and response area value. Furthermore, followed by the non-search feature selection strategy, optimized feature matrix was obtained through the procedure of mean analysis, variation coefficient analysis, cluster analysis and correlation analysis. It was observed that pecans were better classified after the optimization of initial feature matrix, shown by principal component analysis (PCA) score plot. And also the regression models of optimized feature matrix established by partial least squares regression (PLSR) (R-2 = 0.9377) and back propagation neural networks (BPNN) (R-2 = 0.9787) presented a better prediction capacity than these of initial one (PLSR: R-2 = 0.8887; BPNN: R-2 = 0.9093). In conclusion, the optimization method not only reduced data dimensionality but also improved electronic nose performance. (C) 2017 Elsevier Ltd. All rights reserved.