With the emerging of wearable robots, the safety and effectiveness of human-robot physical interaction have attracted extensive attention. Recent studies suggest that online measurement of the interaction force between the robot and the human body is essential to the aspects above in wearable exoskeletons. However, a large proportion of existing wearable exoskeletons monitor and sense the delivered force and torque through an indirect-measure method, in which the torque is estimated by the motor current. Direct force/torque measuring through low-cost and compact wearable sensors remains an open problem. This paper presents a compact soft sensor system for wearable gait assistance exoskeletons. The contact force is converted into a voltage signal by measuring the air pressure within a soft pneumatic chamber. The developed soft force sensor system was implemented on a robotic hip exoskeleton, and the real-time interaction force between the human thigh and the exoskeleton was measured through two differential soft chambers. The delivered torque of the hip exoskeleton was calculated based on a characterization model. Experimental results suggested that the sensor system achieved direct force measurement with an error of 10.3 ± 6.58%, and torque monitoring for a hip exoskeleton which provided an understanding for the importance of direct force/torque measurement for assistive performance. Compared with traditional rigid force sensors, the proposed system has several merits, as it is compact, low-cost, and has good adaptability to the human body due to the soft structure.

The emergence of radio frequency machine learning has significantly propelled the application of deep learning (DL) methods in automatic modulation classification (AMC). However, under non-cooperative scenarios, the performance of DL-based AMC suffers severe performance degradation due to scarce labeled samples and noise interference. To enhance noise robustness in few-shot AMC, this paper proposes a complex-domain autoencoder-based method where a complex-valued noise reduction network (CNRN) is embedded into the AMC framework, jointly extracting complex-valued and temporal features from noisy signals to achieve signal–noise separation. Our framework executes four sequential operations: high-signal-to-noise-ratio (high-SNR) samples are first isolated from limited raw data via unsupervised classification; rotation and cyclic time-shifting operations then augment the sample space; the CNRN is subsequently trained on augmented data; and final AMC classification is implemented through DL-based classifiers. Experimental validation on RML 2016.10a dataset demonstrates: (1) for −20 dB signals, denoising achieves 20.18 dB SNR improvement with 87.74% mean squared error reduction; (2) across the −20 dB to 18 dB range, denoised signals exhibit accuracy improvements of 21.57% under DL-based classifiers. Physical validation further confirms that the proposed method exhibits enhanced noise robustness, demonstrating its practical utility in real-world scenarios.

Automatic modulation classification (AMC) serves as a fundamental task in non-cooperative communication systems, playing a crucial role in signal detection and demodulation. However, acquiring a sufficient number of high-quality labeled samples in real-world scenarios remains challenging, often resulting in insufficient training of end-to-end AMC networks and limited classification performance. To enhance the discrimination of easily confused modulation types and improve overall accuracy under few-shot conditions, this paper introduces a complex-valued circular convolutional meta multi-task network (CCC-MMTN). The proposed framework integrates meta-learning to extract discriminative features from limited data and adopts a multi-task strategy with an alternating local-global training mechanism to facilitate knowledge transfer and strengthen classification robustness. At its core, a complex-valued circular convolutional network (CCCN) is designed as the feature extractor to effectively capture temporal and complex-domain characteristics of communication signals. Extensive experiments on the RML2016.10a dataset demonstrate that, using only five samples per modulation type at a single signal-to-noise ratio (SNR) across eight modulation formats, our method achieves average accuracy rate of 77.35% (-20 dB to 18 dB). The method also achieves an overall accuracy of 84.09% through physical validation with 7 modulation types (50 samples per type) under random SNR conditions, confirming its real-world classification performance.