Guanghua Xu

Objective . Accurate motion intention recognition is challenging due to the limited use of distributed muscle signals. In this study, we proposed a framework that integrates musculoskeletal modelling, advanced feature processing, and feature selection to enhance recognition performance. Approach . A muscle screening method was developed to identify the most informative muscles for joint motion representation. Nonlinear features were extracted from surface electromyography signals using variational mode decomposition and fused with linear features derived from joint angles and moments. To reduce feature redundancy, a Random Forest with Bagging (RF‐Bagging) algorithm was applied for feature selection. Finally, a long short‐term memory regression model was trained for continuous prediction of joint angles and moments. Main results . Experiments on 20 subjects showed that using optimally screened muscles significantly improved prediction accuracy. For knee joint predictions (angle/moment), we achieved MAE of 2.63 ± 0.68 deg/0.020 ± 0.006Nm kg ⁻¹ , RMSE of 4.03 ± 1.13 deg/0.029 ± 0.010Nm kg ⁻¹ , R ² of 0.93 ± 0.05/0.95 ± 0.036, MSE of 17.49 ± 10.78 deg/0.0009 ± 0.0006Nm kg ⁻¹ , and RPD of 1.011 ± 0.049/1.019 ± 0.034, while ankle joint predictions showed MAE of 1.32 ± 0.31 deg/0.035 ± 0.012Nm kg ⁻¹ , RMSE of 1.72 ± 0.43 deg/0.050 ± 0.020Nm kg ⁻¹ , R ² of 0.93 ± 0.03/0.98 ± 0.021, MSE of 3.13 ± 1.47 deg/0.002 ± 0.002Nm kg ⁻¹ , and RPD of 1.004 ± 0.051/1.020 ± 0.040. Significance . Our findings demonstrate that that integrating optimal muscle selection, nonlinear feature enhancement, and intelligent feature selection significantly improves motion intention recognition.

Wrong content There was a mistake in Table 2 as published. Due to a formatting error, several cells that should have been vertically merged were displayed as separate cells. Specifically, the cell containing "Classification Accuracies (%)" and the cell immediately below it were not merged into a single cell. In addition, the cell containing "ωCD" and the three blank cells directly beneath it were not merged into a single cell. The corrected Table 2 The original version of this article has been updated.Adding/removing text In the published article, there was an error in the Introduction section, Paragraph 1, Sentence 3. This sentence previously stated: "Motor imagery (MI) is imaging certain movement of the body part but without any actual motion (Li et al., 2019)." The word "imaging" should be corrected to "imagining".A correction has been made to the Introduction section, Paragraph 1, Sentence 3: "Motor imagery (MI) is imagining certain movement of the body part but without any actual motion (Li et al., 2019)."

Brain-computer interfaces (BCIs) system leveraging high-frequency steady-state visual evoked potentials (SSVEP) enhance user comfort, but their signals suffer from significant amplitude attenuation, which challenges precise detection. To address this, we proposed a novel weak-feature enhancement framework, FBUSSR-CCA, which integrates a monostable (single-well) underdamped second-order stochastic resonance (USSR) model into the filter-bank processing pipeline. This model strategically converts noise energy to nonlinearly amplify weak, high-frequency target components. Validation was performed on two public high-frequency SSVEP datasets (stimulus frequencies of 40-43Hz and 40-41.5Hz, respectively). The proposed FBUSSR-CCA was compared against six benchmark methods, including the best-performing FBCCA. Results indicated that the proposed framework attained superior performance across all three experimental conditions: the mean accuracy and ITR reached 83.61 ± 13.16 % / 17.92 ± 7.81 bits min⁻¹, 98.89 ± 2.17 % / 28.77 ± 2.30 bits min⁻¹, and 70.27 ± 19.02 % / 11.88 ± 8.62 bits min⁻¹, respectively. Further analysis confirmed the USSR module is a general-purpose tool (enhancing all six methods) and is parametrically robust to perturbations. These findings show that the monostable USSR framework is a reliable solution for enhancing high-frequency SSVEP, advancing the potential for more comfortable and efficient BCI systems.

Introduction: Motor imagery (MI) is one of the most widely used paradigms in electroencephalogram (EEG)-based brain–computer interfaces (BCIs). In recent years, deep learning and transfer learning techniques have been increasingly adopted to further improve MI-EEG decoding performance, thereby facilitating the practical deployment of BCIs. In transfer learning, the similarity between the source and target domains is a critical factor influencing its effectiveness. Given the analogous cortical activation patterns observed in MI and motor execution (ME) tasks, cross-task transfer learning from ME to MI presents a promising yet underexplored direction. Methods: To tackle the underexplored problem of cross-task transfer learning from ME to MI, we propose a domain-aware domain–class adaptation network (DDCA Net), which consists of a domain-shared feature extractor, two classifiers, and two domain-specific feature re-weighting blocks. Domain-level alignment is achieved by minimizing the maximum mean discrepancy between source and target feature distributions, while domain-specific feature re-weighting preserves discriminative characteristics unique to each task. In addition, a bi-classifier adversarial learning framework is employed to encourage consistency of decision boundaries across domains, thereby enabling implicit class-level alignment. Results: Extensive experiments were conducted on a public dataset with over 100 subjects under varying proportions of target-domain training samples. When 80% of target-domain samples are used for training, the proposed DDCA Net significantly outperforms the within-task baseline, achieving a 7.71% improvement in classification accuracy and converting approximately 80% of previously BCI-illiterate subjects into BCI-literate users. Discussion: To the best of our knowledge, this is the first work to verify the feasibility of applying domain adaptation for cross-task transfer learning in MI-EEG classification. The findings of this study provide new insights for integrating ME and MI in advanced BCIs.

In industrial applications, incipient misalignment and rub-impact faults in three phase induction motors is critical, as they induce abnormal vibrations resulting in accelerated wear and tears, elevated energy consumption and may even result in a catastrophic failure. This paper presents the practical implementation of a novel non-invasive technique known as Holo-current spectrum method, that detects induction motor faults by using frequency spectrum of stator current. This method eliminates the use of complex external sensor systems. The core novelty of this work lies in the key-phasor averaging technique that establishes a baseline 0° reference thus enabling precise extraction of shaft position and phase from α–β phase vectors of stator current. This process results in a 2D elliptical orbital representation for quantitative eccentricity analysis. The validation of method was carried out on a hybrid experimental setup by synchronously acquiring current and vibration data. The results identify that the misalignment fault is prominent in 1^st, 2^nd, and 3^rd order elliptical components while rub-impact faults are more prominent in higher order elliptical components i.e. 4^th and 5^th order. Moreover, a correlation exceeding 0.92 was observed between orbital eccentricities and the longitudinal major axis inclination from both current and vibration signals at low angular speeds. This correlation confirms that stator current analysis alone can diagnose these faults with a precision comparable to vibration analysis. The Holo-current spectrum technique provides a direct, non-invasive assessment of fault severity level, is robust against rotor asymmetries, and serves as a significant progression in early fault identification through self-monitoring intelligence in induction motors.

Nowadays, legged robots often use speaker recognition or Ultra Wide Band(UWB) positioning to identify and track certain people. However, these identification methods need the active cooperation of the identified person, which limits their application. To solve this problem, we propose a “Symmetry-encoding and pseudo-Centroid loss optimized Gait recognition method” (SCGait). The experimental results show that our gait recognition method achieves a mean test accuracy of 82.2% on three subsets of the CASIA-B dataset, and the ablation studies show that our gait encoding method and loss function have good generalization ability. Next, we combine SCGait with Yolo to develop a person identification-tracking system for legged robots. The experimental results show that our system performs quite well in both home companion and industrial patrol legged robots. It achieves an identification accuracy of 91.8% and a FPS of 36 in our test video in a multi-person scene. The code is available at https://github.com/qplqplqpl/SCGait.

Supernumerary robotic limbs (SRLs) are extra robotic appendages that require sensory-motor integration for intuitive control, yet most lack proprioceptive feedback. Existing approaches using vibrotactile or electrotactile cues often feel unnatural and offer limited resolution. We present a real-time spatial audio system that provides users with the 3D position of an SRL's end-effector without compromising their native motor or tactile functions. Implemented in ROS2, the system was evaluated for spatial localization error and resolution. Since SRLs are controlled through continuous motion, we prioritized dynamic performance, where the mean localization error was 14.06° in azimuth, 20.49° in elevation, and 0.078 m in distance (with a corresponding dynamic resolution of 4.42°, 2.13°, and 0.044 m). For context, the system's static spatial acuity was also high, achieving a 3AFC success rate of 93.5% in azimuth, 68.0% in elevation, and 87.0% in distance, and a static resolution of 8.43°, 2.59°, and 0.031 m. In the real robotic arm end-effector aiming test, our distance perception error was 0.137 ± 0.046 m. These results demonstrate that the proposed system provides accurate and reliable artificial proprioception for SRLs, offering a promising pathway toward more seamless control.

This study investigates the impact of music training on the neural processing of musical syntax, as captured by electroencephalography (EEG) signals. It addresses the complex, long‐term brain state changes and inter‐regional interactions during syntactic processing, which remain poorly understood. We compared the syntactic sensitivity of musicians and nonmusicians by exposing them to music fragments with varying degrees of syntactical violations according to music theory. Our findings reveal that music training enhances brain sensitivity to syntactic violations, with musicians exhibiting lower low‐band power and higher high‐band power in EEG signals compared to nonmusicians. Additionally, musicians showed higher dispersion entropy and fractal dimension in EEG signals, indicating greater complexity in neural processing. Brain connectivity, as measured by covariance, decreased for nonmusicians but increased for musicians when listening to the most syntactically irregular version of a familiar piece of music, suggesting differential neural engagement. Phase locking value analysis demonstrated consistently higher brain connectivity in musicians. This research provides insights into the advanced cognitive processes underlying music syntax processing and offers a novel approach for evaluating music cognition levels, contributing to the understanding of enhanced cognition in musicians.

Goal: Motor unit (MU) discharge information extracted via real-time electromyogram (EMG) decomposition shows superiority in dexterous finger motion decoding. However, some critical constraints might limit the performance of multi-degree-of-freedom (DoF) muscle force estimation, including the neglect of the temporal cumulative effects of MU discharges and the rigid MU pool allocation across fingers. Methods: A hybrid encoder-decoder framework integrating EMG decomposition with TFR (Temporal Firing Rate)-Net was proposed for multi-DoF force estimation. The initial encoder involved decomposing EMG into MU spike trains and firing rates, which are subsequently processed through TFR-Net for hierarchical encoding and decoding. This framework uses temporal firing rate to model force accumulation instead of a twitch model, and allocates MU weights dynamically and flexibly in multi-fingers and multi-force level tasks instead of rigid MU pool allocation. Results: Two baselines (firing rate-based regression method and twitch force model method) were evaluated on 15-min EMG from 10 subjects performing alternating multi-finger isometric extensions. The results showed the proposed method demonstrated superior comprehensive performance with a higher correlation (R²: 0.80 ± 0.12 vs. 0.75 ± 0.14 vs. 0.62 ± 0.19) and a lower prediction error (root-mean-square error: 6.32% ± 1.89% vs. 7.24% ± 1.82% vs. 9.96% ± 2.54 % maximum voluntary contraction) compared with the two comparison methods, and an improved computational efficiency (Flops: 1.350k vs. 408.093k vs. 0.016k) compared with the twitch force model method. Significance: Further development of the proposed method could potentially provide a robust human machine interface for dexterous finger force prediction in realistic applications.

Background: Dynamic visual acuity (DVA) is functionally distinct from static visual acuity (SVA), though SVA is often used clinically as a reference. Methods: To identify EEG biomarkers for DVA, we presented participants with a high-contrast checkerboard moving horizontally at speeds ranging from 4°/s to 30°/s, engaging motion-sensitive pathways while preserving spatial detail. Six EEG features—ERPs (N200 and P300), TRCA, Hjorth activity, mean curve length, and Tsallis entropy—were extracted from eight occipito-parietal channels and evaluated for speed sensitivity. Results: Hjorth activity and Tsallis entropy showed consistent monotonic trends with respect to speed. Hjorth activity exhibited the strongest univariate correlation (r = 0.88, p < 0.05). In a Lasso regression model using all speed-sensitive features, the predicted speed correlated with actual speed at r = 0.588, with TRCA-weighted features retained for their multivariate contribution. Notably, Hjorth activity peaked at PO7/PO8 (3.558 and 1.478 µV² at 30°/s), aligning with V5/MT+ activation. Conclusion: Given its high sensitivity, neuroanatomical plausibility, and simplicity, Hjorth activity is recommended as a primary candidate for EEG-based DVA biomarker development. This study provides a foundation for objective neurophysiological evaluation of dynamic vision.