During bimanual movement, brain state is known to be different from the unimanual movement. Thus the conventional arm movement classifier for unimanual arm movement decoding method seems to be insufficient to decode bimanual movement. In this research, we suggested the convolutional neural network (CNN) for movement state classification to improve the decoding accuracy for bimanual movement estimation. We recorded the monkey's cortical signal while the bimanual task, and convert to spectrogram dataset for decoding. To evaluate the CNN, we stacked several layers for deep structure and figured out the best configuration. As a result, this method showed improved the arm movement state classification performance for bimanual tasks. This technique could be applied to arm movement brain computer interfaces (BCIs) in real world and the various neuro-prosthetics fields.