This study proposes a spatio-temporal dynamic multiscale convolutional neural network (DMS-CNN) classification model based on a dimensional attention mechanism to improve classification accuracy and applicability to practical scenarios, in order to address the problems of non-stationarity, time-varying complexity, and low classification accuracy of electroencephalogram (EEG) signals, as well as the shortcomings of traditional machine learning methods in extracting complex features. First, the data are bandpass-filtered to eliminate artifacts and pre-processed using downsampling and channel selection. The processed data are then input into the constructed spatiotemporal convolution model, to further enhance the feature extraction capability of the network and multidimensional attention mechanisms of timing. This is followed by the incorporation of channel and regularization technology. To address the problem of insufficient data, a frequency-band exchange method is used to enhance the data, thereby improving the generalization performance of the model. Average classification accuracies of 90.97% and 90.21% were obtained for the HGD and self-collected laboratory datasets, respectively. Compared with other algorithms, the classification accuracy of this method was significantly improved.