The rapid development of information technology in today's era urges people to pay more attention to the field of human-computer interaction. It is of great significance to monitor the mental load of operators at any time and adjust the workload of operators according to this. Some studies have shown that the power spectral density of EEG signals is more suitable for the task of mental load classification, but the feature dimension of EEG is high, which is prone to dimensional disasters. At present, principal component analysis (PCA) is the most widely used algorithm in dimension reduction in machine learning. Aiming at the unadaptability of PCA in EEG signal classification and the sensitivity of SVM to the relationship between features, this paper proposes a wrapped dimension reduction method based on PCA-SVM and stepwise enumeration method. In the feature engineering stage, the concept of fixed verification set is introduced to assist wrapped dimension reduction, and the feature engineering scheme is adjusted with the accuracy of verification set as an index, So as to improve the separability of data after dimensionality reduction. Due to the introduction of the concept of supervised learning, the experimental results show that the accuracy of data classification after dimensionality reduction based on PCA-SVM and step-by-step enumeration is generally higher than that based on traditional PCA, which provides a new idea for dimensionality reduction of high-dimensional bioelectric data.