To solve the problem of low data classificationaccuracy and poor model stability of functional near-infrared brain functional imaging signals in brain-computer interfaces, this paper proposes to study the changes of cerebral blood oxygen concentration in the prefrontal brain region during the stimulation period and carry out a study of the state binary classification identification of task and rest on the experimental data of the prefrontal brain region, and take the extracted single feature and multi-feature fusion as the inputs of the model, respectively, and validate the results of the classification through the model The conjecture that multi-feature fusion can improve the classification accuracy to some extent. Firstly, this paper designs an experimental paradigm for the prefrontal brain region: word fluency cognition experiment. The device used to collect the data is NirSmart, a Danyang Huichuang device from China, with a sampling rate of 11 Hz and a spatial resolution of 3 cm. The device uses avalanche diode and ultramicro photo detection technology, and the sensitivity is up to 0.1 pW. After that, the collected data are preprocessed using the Homer2 toolbox, the features are extracted using MATLAB, and the Random Forest model is constructed using two metrics of feature importance and error curve to evaluate the classification results. Importance and error curves were used to evaluate the performance of the model. Finally, the average of 20 runs of the random forest model was used as the final classification result. The experimental results show that multiple features can improve the final classification result compared with a single feature. The best second classification result for the case of three-feature fusion is 93.84%, 2.32%, 4.25%, and 5.33%, higher than the single-feature mean, slope, and peak-to-peak value, respectively. From the results of the experimental data, it can be seen that multi-feature fusion can improve the accuracy of near-infrared brain functional imaging classification to a certain extent. The performance of the random forest model is stable, which is expected to solve the previous problems of low classification accuracy and poor model stability and promote the development and application of brain-computer interface systems.