In recent years, with the continuous development of artificial intelligence technology, Brain-Computer Interface (BCI) has attracted more and more scholars' attention. Real-time monitoring of mental load level of high-intensity mental workers and dynamic adjustment of their tasks is an important means to protect national property and the safety of operators. A study shows that the characteristic power spectrum density extracted by Electroencephalogram (EEG) is sensitive to the changes of mental load, but its dimension is too high, leading to data disaster. Traditional Principal Component Analysis (PCA) algorithm will lose some nonlinear features. Locally Linear Embedding (LLE) is a common nonlinear dimension reduction algorithm, but the algorithm is highly sensitive to noise and the result of dimension reduction is greatly affected by parameters. The Robust Locally Linear Embedding algorithm (RLLE) adds regular term optimization when LLE optimizes the weight matrix, which not only enhances the anti-noise ability of the model, but also solves the problem of matrix sickness and singularity that may appear in the process of understanding the model. Parameter k in this algorithm is often used with a smaller value to better capture the local structure of the data set and greatly reduce the calculation time of the model. However, EEG data has the characteristics of high dimension and complexity. Selecting a small k value will not only cause the model to be abnormally sensitive to noise, but also make the model ignore important large neighborhood structure, thus affecting the accuracy of dimension reduction results. In this experiment, when using this algorithm, a more reasonable interval of k value is selected in combination with model accuracy and calculation time, so that the model has stronger anti-interference ability while maintaining efficiency, and can provide more comprehensive information to describe the data set, so that the embedded results are more accurate. The experimental results show that the classification accuracy of data after RLLE dimension reduction using SVM is generally higher than that after PCA dimension reduction, which has stronger anti-interference ability.