This study looks at how convolutional neural networks are used in classification of motor imagery. The study was performed using three main parts- a multi-band segmentation module, an attentional spatial convolution module, and a multi-scale temporal convolution module. The multi-band module used a filter bank with overlapping frequency bands to boost the features of the frequency domain. The attention spatial convolution module adjusted convolutional kernel parameters in accordance with the input using an attention mechanism to obtain the features of the various datasets. The results were grouped together to enable multi-scale temporal convolution. The multi-scale temporal convolution results used bilinear pooling layer to extricate temporal features and execute noise elimination, and then the extracted features were classified. A new convolutional neural network (CNN) and multi-scale attentional convolutional network (MSAttNet) were introduced to improve motor imagery classification accuracy and efficiency. The method of parting bands at various Hz values proved effective, as the method was able to reach SOTA performance and high accuragies across four benchmark datasets. The MSAttNet used differing frequency bands and altered their segmentation methods to observe the impact on network performance. The study found that using overlapping frequency bands improved the network recognition, and the study shows the efficacy of different frequency division filter banks.