The communication between the brain of an individual person and the external devices is facilitated by brain-computer interface (BCI) through the utilisation of electroencephalograph (EEG) data decoding. The study of motor imagery EEG (MI-EEG) enables the identification of the subject’s active intention. There is not a clear link between signals and how the brain works, so it’s hard to figure out what they mean. Specific EEG patterns produced by simulated motions should be identified properly to increase the effectiveness of the BCI devices. In this work, a novel deep learning technique based on one-dimensional separable convolution is proposed to classify the motor imagery signals into four different classes for the correct identification. The proposed deep learning architecture has been tested using the BCI Competition IV 2a, the publicly available motor imagery EEG Dataset. The work exceeds the state-of-the-art models by obtaining high classification performance and requiring only fewer trainable parameters.