A Weight-aware-based Multi-source Unsupervised Domain Adaptation Method for Human Motion Intention Recognition

TL;DR

This work extends the margin disparity discrepancy (MDD) framework to multi-source unsupervised domain adaptation for human motion intention recognition (HMI). It introduces WMDD, a weight-aware MUDA algorithm that adaptively assigns source-domain weights via MDD, uses a lightweight feature extractor with an ensemble of classifiers, and employs adversarial learning to enhance generalization on a target subject. The authors derive a comprehensive generalization bound for MUDA, connect theory to a practical optimization protocol, and demonstrate state-of-the-art target accuracy on ENABL3S and DSADS with real-time inference. The approach explicitly addresses inter-source variability, achieving robust cross-subject HMI recognition suitable for real-time exoskeleton control. Overall, WMDD provides a theoretically grounded, efficiency-driven solution with strong empirical performance and potential applicability beyond HMI tasks.

Abstract

Accurate recognition of human motion intention (HMI) is beneficial for exoskeleton robots to improve the wearing comfort level and achieve natural human-robot interaction. A classifier trained on labeled source subjects (domains) performs poorly on unlabeled target subject since the difference in individual motor characteristics. The unsupervised domain adaptation (UDA) method has become an effective way to this problem. However, the labeled data are collected from multiple source subjects that might be different not only from the target subject but also from each other. The current UDA methods for HMI recognition ignore the difference between each source subject, which reduces the classification accuracy. Therefore, this paper considers the differences between source subjects and develops a novel theory and algorithm for UDA to recognize HMI, where the margin disparity discrepancy (MDD) is extended to multi-source UDA theory and a novel weight-aware-based multi-source UDA algorithm (WMDD) is proposed. The source domain weight, which can be adjusted adaptively by the MDD between each source subject and target subject, is incorporated into UDA to measure the differences between source subjects. The developed multi-source UDA theory is theoretical and the generalization error on target subject is guaranteed. The theory can be transformed into an optimization problem for UDA, successfully bridging the gap between theory and algorithm. Moreover, a lightweight network is employed to guarantee the real-time of classification and the adversarial learning between feature generator and ensemble classifiers is utilized to further improve the generalization ability. The extensive experiments verify theoretical analysis and show that WMDD outperforms previous UDA methods on HMI recognition tasks.

Figures (6)

• Figure 1: The overview of the weighted multi-source unsupervised domain adaptation (WMDD) method for human motion intention recognition. The training data are from $N$ source subjects with labels and one target subject without labels. The features of the sensor signal are extracted by feature generator. The final goal is to use feature generator and ensemble classifiers to classify the target subject intention accurately. Firstly, the auxiliary classifiers are optimized by maximizing the discrepancy, then the source domain weight can be achieved according to the estimated MDD. Secondly, the feature generator and classifiers are trained to minimize total loss. Finally, motivated by adversarial learning, the classifiers and generator are optimized by minimizing and maximizing classifiers discrepancy.
• Figure 2: The comparison between the single-source unsupervised domain adaptation (SUDA) and multi-source unsupervised domain adaptation (MUDA), where different shapes represent different data categories and different colors denotes different domains. The source and target distributions of SUDA are not matched well. For MUDA, the target distribution hardly matches all subject distributions, and the discrepancy between the target distribution and each subject distribution might be different.
• Figure 3: The training steps of the proposed algorithm WMDD. $\boldsymbol{f}'$, $\phi$ and $\boldsymbol{f}$ represent auxiliary classifiers, feature generator and classifier, respectively. Steps 1 to 4 are repeated in sequence until the training is completed.
• Figure 4: The visualization of t-SNE projection of non-adapted input features and the hidden features adapted by the feature generator. All features are extracted from the training set for ENABL3S (a) and DSADS (b). The different color points represent different classes. The dark-color and light-color points denote the target and source features, respectively
• Figure 5: The comparison results of training curve under different margin $\mu$ on ENABL3S (a) and DSADS (b) datasets. The comparison results of the average margin disparity discrepancy between source and target distribution under different $\mu$ on ENABL3S (c) and DSADS (d) datasets. For the convenience of comparison, the ordinate values of sub-figure (c) and (d) are $\log{\text{MDD}}$.