The cotton pests have the characteristics of concealment, migration, and sudden burst, and there are many influencing factors involved. The accurate diagnosis of cotton pests is a difficult problem in the agricultural field. Previous studies have demonstrated that cotton plants produce blends of volatile compounds in response to herbivores serve as cues for parasitic and predatory insects. Therefore, it is possible to obtain information about cotton pests by detecting volatile compounds in cotton. In this study, an electronic nose was used to detect the volatiles emitted by cotton plants damaged by cotton bollworm at the flowering period. The cotton samples were divided into four infested cotton treatments. According to the number of pests in each pot of cotton seedlings, the treatments inoculated with 0, 1, 2, and 3 bollworm larvae were marked as 0-P, 1-P, 2-P, and 3-P, respectively. The 0-P was healthy cotton as a control treatment. The cotton bollworm feeding lasted 48 h. During this period, the electronic nose detection tests were performed every 6 h, and a total of 8 repeated tests were performed. Appropriate pattern recognition techniques were applied to construct reliable algorithms for interpreting the acquired signal in cotton. Principal Component Analysis (PCA), discriminant function analysis, cluster analysis, and Radial Basis Function Neural Network (RBFNN) were applied to evaluate the data. The results of PCA and discrimination values of the healthy cotton treatment showed that the volatiles released by healthy cotton had obvious circadian rhythm. For the three infested cotton treatments, whereas the distribution patterns of cotton samples were different from that of the healthy cotton treatment. The three infested cotton treatments had regular distribution trends that cotton samples changed along the direction of the first and second principal components. Cluster analysis results showed that the four cotton treatments were all finally divided into two categories, the healthy cotton treatment, and the three infested cotton treatments. All these results suggested that there was a significant difference between healthy and damaged cotton samples. Then RBFNN was used to analyze four treatments of cotton samples at 8 different times. The results showed that the total correct rate of the test sets was 73.4%, the correct rate of the healthy cotton treatment was 100%, and the misjudgment samples appeared among the three infested cotton treatments. Moreover, two unified consecutive prediction models were established regardless of the time factor. The RBFNN model was established by using four treatments of cotton samples. The correct rate of the training sets was 66.1%, and the correct rates of the test sets were 100 %, 79.7 %, 32.8 %, and 20.3 % for the 0-P, 1-P, 2-P, and 3-P treatments, respectively. In another RBFNN model based on 0-P, 1-P, and 3-P treatments, the correct rate of the training sets was 87.8%, and the correct rates of the test sets were 100 %, 78.1%, and 82.8% for the 0-P, 1-P, and 3-P treatments, respectively. Comparing the results of the two RBFNN models, the prediction accuracy of the second model had been greatly improved. At the same time, it was also found that the prediction accuracy of all RBFNN models for healthy cotton treatment reached 100%. Therefore, the electronic nose could be used as an effective monitoring method for the occurrence of cotton bollworm in the cotton plants. It should have a potential application for crop pest monitoring in the field. [ABSTRACT FROM AUTHOR]