Attention-Based Multiscale tCNN for SSVEP Classification and Its Application to Bionic Intelligent Soft Gripper Control

To address the classification problem of short time-window steady-state visual evoked potentials (SSVEP), a novel deep-convolutional neural network (CNN) fused with residual squeeze and excitation blocks (RSEs) and multiscale convolutions is proposed. Given the difficulty in distinguishing frequency domain features of short time-window signals, AttentCNN-Multiscale begins with a filter bank (FB)-based time-domain feature extraction module. The FB comprises several sixth-order Butterworth filters with varying bandpass ranges. Then the feature tensors extracted by these filters are aggregated using a CNN with RSEs. For further feature learning, four 2-D CNNs and a multiscale convolution module are employed, with the final output generated through an adaptive fully connected layer. To demonstrate the effectiveness and superiority of AttentCNN-Multiscale, extensive experiments and comparisons are conducted on two large public datasets and our dataset. Additionally, a novel bionic intelligent soft gripper is designed and integrated with the proposed AttentCNN-Multiscale network to form a closed-loop system, enabling different grasping functionalities for various objects and demonstrating the application potential of the network in medical rehabilitation. To ensure reproducibility, the source code for AttentCNN-Multiscale is available on Github: https://github.com/raow923/AttentCNN-Multiscale.