2026-04-25 投稿版
ALL.tex L503–L507
Six classifiers were benchmarked, including three classical machine-learning methods (KNN [26], SVM-RBF [27], and Random Forest [28]) and three learning-based methods (1D-CNN [29], GAF-CNN [30], and GAF+PCA+SVM). To accommodate their different input topologies, the 20-channel time series (resampled to 200 timesteps) were represented in three forms: (1) flattened feature vectors (Xflat ∈ ℝ^(N × 4000)) for classical classifiers; (2) raw multi-channel sequences (Xseq ∈ ℝ^(N × 20 × 200)) for the 1D-CNN; and (3) Gramian Angular Field (GAF) tensor encodings [30, 13] (XGAF ∈ ℝ^(N × 20 × 64 × 64)), which map each channel time series into a 2D angular-correlation image for 2D-CNN-based models. All models used 5-fold class-stratified CV, with normalization fitted on training folds only. Since each gesture has only two repetitions per subject, same-subject trials may share within-subject folds; we therefore treat the subject-disjoint LOSO evaluation (Section IV-D) as the primary cross-user indicator, and the within-subject numbers as an in-session upper bound.
中文注释共比较六种分类器，包括三种传统机器学习方法（KNN、RBF 核 SVM 和随机森林）以及三种学习方法（1D-CNN、GAF-CNN 和 GAF+PCA+SVM）。为适应不同输入结构，将重采样为 200 个时间步的 20 通道时间序列表示为三种形式：（1）供传统分类器使用的展平特征向量 X_flat∈R^(N×4000)；（2）供 1D-CNN 使用的原始多通道序列 X_seq∈R^(N×20×200)；（3）GAF 张量编码 X_GAF∈R^(N×20×64×64)，把每个通道的时间序列映射为二维角相关图像。所有模型均采用五折类别分层交叉验证，归一化只在训练折拟合。由于每名参与者每类手势只有两次重复，同一参与者的 trial 可能分布在不同折，因此将参与者隔离的 LOSO 作为主要跨用户指标，把受试者内结果视为会话内上限。
当前工作区版本
ALL.tex L501–L506
Six classifiers were benchmarked, including three classical machine-learning methods (KNN [26], SVM-RBF [27], and Random Forest [28]) and three learning-based methods (1D-CNN [29], GAF-CNN [30], and GAF+PCA+SVM). To accommodate their different input topologies, the 20-channel time series (resampled to 200 timesteps) were represented in three forms: (1) flattened feature vectors (Xflat ∈ ℝ^(N × 4000)) for classical classifiers; (2) raw multi-channel sequences (Xseq ∈ ℝ^(N × 20 × 200)) for the 1D-CNN; and (3) Gramian Angular Field (GAF) tensor encodings [30, 13] (XGAF ∈ ℝ^(N × 20 × 64 × 64)), which map each channel time series into a 2D angular-correlation image for 2D-CNN-based models. Supplementary Tables S3 and S4 report the handcrafted-feature, dimensionality-reduction, and channel-ablation analyses. The t-SNE visualization in Fig. 12 uses globally pooled Conv1 embeddings from a 1D-CNN, with full projection parameters provided in the Supplementary Material. All models were evaluated using both a class-stratified five-fold split at the trial level, which allowed participant overlap, and strict LOSO, which kept participants disjoint between training and testing. The five-fold results are therefore treated only as an in-distribution upper bound. In LOSO, one complete participant and all corresponding trials were held out in each of ten folds, making LOSO the primary cross-participant indicator. Normalization, PCA, ANOVA, and feature selection, when applicable, were fitted using training-fold data only. The “RF + handcrafted” configuration in Table II uses a 570-dimensional feature set comprising time-domain, frequency-domain, and cross-channel structural features.
中文注释共比较六种分类器，包括三种传统机器学习方法（KNN、RBF 核 SVM 和随机森林）以及三种学习方法（1D-CNN、GAF-CNN 和 GAF+PCA+SVM）。为适应不同输入结构，将重采样为 200 个时间步的 20 通道时间序列表示为三种形式：（1）供传统分类器使用的展平特征向量 X_flat∈R^(N×4000)；（2）供 1D-CNN 使用的原始多通道序列 X_seq∈R^(N×20×200)；（3）GAF 张量编码 X_GAF∈R^(N×20×64×64)，把每个通道的时间序列映射为二维角相关图像。Supplementary Tables S3 和 S4 给出人工特征、降维和通道消融分析。图 12 的 t-SNE 可视化采用 1D-CNN 第一卷积块 Conv1 特征在时间维全局池化后的表示，完整投影参数见补充材料。所有模型同时采用允许参与者重叠的试次级类别分层五折划分和训练、测试参与者完全隔离的严格 LOSO。五折结果仅视为同分布上限。LOSO 的十个折中，每次留出一名完整参与者及其全部 trial，因此 LOSO 是主要跨参与者指标。归一化、PCA、ANOVA 和特征选择如有使用，均只在训练折拟合。表 II 中“RF + handcrafted”使用 570 维特征集，包括时域、频域和跨通道结构特征。