基于SE-TCN-Transformer融合网络的下肢外骨骼机器人运动意图识别研究

Research on Motion Intention Recognition for Lower-Limb Exoskeleton Robots Based on SE-TCN-Transformer Fusion Network

摘要: 针对下肢外骨骼机器人人机交互系统中对运动意图识别的高精度与实时性要求，本文提出了一种基于SE-TCN-Transformer的新型融合网络架构。现有的卷积神经网络(CNN)和长短期记忆网络(LSTM)在处理长时序多模态传感器数据时，往往面临局部特征提取不足、全局依赖捕捉困难以及难以有效抑制传感器噪声等挑战。为解决上述问题，本研究设计了“局部特征增强–全局语义建模”的双阶段特征提取策略。首先，引入嵌入Squeeze-and-Excitation (SE)通道注意力机制的时间卷积网络(Temporal Convolutional Network, TCN)作为前端，通过因果空洞卷积提取高频局部运动特征，并利用SE模块自适应重标定通道权重，有效筛选关键传感器信息并抑制冗余噪声。其次，采用Transformer编码器对特征序列进行全局建模，利用多头自注意力机制精准捕捉步态相位间的长距离时序依赖。在公开数据集HuGaDB上的实验结果表明，该模型对8种复杂日常步态的平均识别准确率达到96.73%，F1分数达到95.43%。此外，基于Shapley Additive exPlanations (SHAP)的可解释性分析揭示，足部垂直加速度与大腿角速度是步态识别中作用较大，表明模型提取的关键特征分布与人体下肢生物力学的先验知识具有较高的一致性。特别是在坐下、起立等具有瞬态突变特征的转换动作识别中，该模型表现出显著优势，优于传统的SVM、LSTM及CNN-LSTM混合模型。该研究不仅验证了混合架构在多模态步态分析中的有效性，也为外骨骼机器人的实时精准控制提供了有力的技术支撑。

Abstract: To achieve high-precision and real-time motion intention recognition for lower-limb exoskeletons, this paper proposes a novel hybrid architecture: the SE-TCN-Transformer. Addressing the limitations of conventional CNNs and LSTMs in processing long-sequence multimodal data, we introduce a “local enhancement-global modeling” strategy. First, a Temporal Convolutional Network (TCN) integrated with Squeeze-and-Excitation (SE) attention extracts high-frequency local features while adaptively suppressing sensor noise. Second, a Transformer encoder captures long-range temporal dependencies via multi-head self-attention. Experiments on the HuGaDB dataset demonstrate an average accuracy of 96.73% and an F1-score of 95.43% across eight gaits. SHAP-based interpretability analysis identifies foot vertical acceleration and thigh angular velocity as the most critical features, demonstrating a higher consistency with biomechanical priors. The proposed model significantly outperforms traditional SVM, LSTM, and CNN-LSTM baselines, particularly in recognizing transient actions like “Sit” and “Stand”, thereby providing robust technical support for real-time exoskeleton control.