The oral odor of human beings is directly related to the disease of the human body. The content of acetone in the exhaled gas can be used as an important basis for judging diabetes. Based on the electronic nose (e-nose) technology, this paper optimizes the metal oxide semiconductor gas sensor array in the gas collection system to design a non-invasive early oral odor detection system for diabetes. The original data is reduced in dimension by principal component analysis algorithm and artificial neural network algorithm. The experimental results show that the oral odor detection system has high identification and accuracy for the content of acetone in the exhaled gas. The accuracy of sample identification on fasting is 85%, and the accuracy rate is up to 98% one hour after meal, and it is 92% two hours after meal. This study provides theoretical guidance for early non-invasive diagnosis of diabetes.