Multi-Tiered CNN Model for Motor Imagery Analysis: Enhancing UAV Control in Smart City Infrastructure for Industry 5.0

Abstract

The concept of brain-controlled UAVs, pioneered by researchers at the University of Minnesota, initiated a series of investigations. These early efforts laid the foundation for more advanced prototypes of brain-controlled UAVs. However, BCI signals are inherently complex due to their nonstationary and high-dimensionality nature. Therefore, it is crucial to carefully consider both feature extraction and the classification process. This study introduces a novel approach, combining a pretrained CNN with a classical neural network classifier and STFT spectrum, into a Multi-Tiered CNN model (MTCNN). The MTCNN model is applied to decode two-class Motor Imagery (MI) signals, enabling the control of UAV up/down movement. The experimental phase of this study involved four key experiments. The first assessed the MTCNN model's performance using a substantial dataset, resulting in an impressive classification accuracy of 99.1%. The second and third experiments evaluated the model on two different datasets for the same subjects, successfully addressing challenges associated with inter-subject and intra-subject variability. The MTCNN model achieved a remarkable classification accuracy of 99.7% on both datasets. In a fourth experiment, the model was validated on an additional dataset, achieving classification accuracies of 100% and 99.6%. Remarkably, the MTCNN model surpassed the accuracy of existing literature on two BCI competition datasets. In conclusion, the MTCNN model demonstrates its potential to decode MI signals associated with left- and right-hand movements, offering promising applications in the field of brain-controlled UAVs, particularly in controlling up/down movements. Furthermore, the MTCNN model holds the potential to contribute significantly to the BCI-MI community by facilitating the integration of this model into MI-based UAV control systems.