Table 7 Common deep learning algorithms and their applications in HMIP

CNN

Extract local spatial features, pooling layers for dimensionality reduction

Efficiently capture spatial correlations in multi-channel signals (e.g., sEMG)

sEMG gesture classification, IMU-based gait feature extraction^247,253

RNN

Process sequential data via recurrent structures

Simple temporal analysis, low latency

Basic motion recognition (e.g., arm swinging)^110,147

LSTM

Resolve long-term dependencies through gated mechanisms

Dynamically parse continuous motion trajectories

Exoskeleton joint angle prediction^148,244,245

BiLSTM

Fuse forward/backward temporal information via bidirectional architecture

Capture complete temporal context

Gait phase segmentation, stair ascent/descent intent^150,168,241

TCN

Dilated causal convolutions, parallel temporal processing

Efficient long-sequence handling, low inference latency

Real-time gesture recognition, continuous trajectory prediction^242,243

Transformer

Global modeling via self-attention

Adaptive multimodal fusion, strong generalization

Cross-user intent transfer learning, multi-sensor fusion^152,254

GAN

Synthesize data via adversarial generation

Alleviate data scarcity, enable domain adaptation

sEMG data augmentation, cross-device adaptation^267,268,269

GNN

Model sensor topological relationships via graph structures

Explicit spatial correlation inference

Whole-body motion intent network analysis^270,271,272