A lightweight multi-scale convolutional attention network for lower limb motion recognition with transfer learning

Surface electromyography (sEMG) signals can directly reflect intention of human motion, therefore are widely used in motion recognition and robot control. However, the accuracy of recognizing lower limb gait patterns tends to decrease when applying a pre-trained model to new users. Additionally, gathering sufficient data from new users to calibrate a new model in practical applications is challenging. To address this issue, this study proposes a lightweight deep learning model that utilizes multi-scale depthwise convolution to extract multi- scale features from multi-channel sEMG signals. An attention mechanism is employed to fuse these features across channels. A transfer learning strategy involves transferring the parameters of the feature extractor from the pre-trained model to the target model. The label classifier is fine-tuned using limited data from unseen subjects. The experimental results show that the proposed model can effectively extract the features from sparse multi-channel sEMG signals and accurately identify motion intentions of lower limbs. The average recognition accuracies achieved on the Gravity dataset and ENABL3S dataset are 91.49% and 93.07%, respectively. The proposed model outperforms the baseline model in terms of recognition accuracy and training time burden. Moreover, our proposed transfer learning approach has greater efficiency compared with a training model specialized for each new users from scratch, which owes to its ability to achieve satisfactory recognition accuracy within a short calibration time by utilizing only six repetitions of training data. The research provides a viable solution to address the issue of data scarcity in practical applications and enhance user-friendliness.