Curriculum Domain Adaptation via Wavelet Mixup for Cyber–Physical–Social Intelligence

Cyber–physical–social intelligence (CPSI) requires robust perception models that generalize across heterogeneous human-centered sensing environments. However, sensor data in cyber–physical–social systems (CPSSs) often exhibit strong nonstationarity, low signal-to-noise ratios, and substantial cross-subject distribution shifts, which can severely degrade model performance. Existing adversarial unsupervised domain adaptation (UDA) methods mainly focus on feature-level alignment, but they often underuse target-domain structural information when training the task classifier, especially under large source–target discrepancies. To address this issue, we propose wavelet-enhanced curriculum domain adaptation (Wave-CDA), a curriculum-based UDA framework with wavelet-domain mixup for CPSS-relevant sensor time series. Wave-CDA consists of three stages: 1) adversarial UDA pretraining for learning transferable representations; 2) multilevel wavelet mixup (WMix) for generating intermediate-domain samples that preserve important time–frequency structures while incorporating unlabeled target information; and 3) transferability-guided curriculum training that progressively introduces augmented samples from easier-to-transfer samples to harder samples. We further design three transferability measures, namely label confidence, label agreement, and domain confusion, to assess the reliability and usefulness of augmented samples. Experiments on cross-subject emotion recognition, human activity recognition, and human gesture recognition show that Wave-CDA consistently outperforms competitive baselines. Additional analyses of ablation, transferability weighting, and training behavior further support the effectiveness of wavelet-domain augmentation and curriculum scheduling.