Sensors

Reducing user-specific calibration is critical for practical steady-state visual evoked potential (SSVEP)-based brain–computer interfaces (BCIs), yet few-shot cross-subject decoding remains challenged by heterogeneous source transferability and inter-subject variability. We propose SQ-HAF, a source quality-guided framework that combines target-relevant source selection, frequency neighborhood-regularized spatial filtering, covariance alignment, and multi-branch decision fusion. Candidate source subjects are ranked primarily by target–source template similarity; when more than one labeled calibration trial per stimulus is available, split-half template consistency (STC) provides a bounded confidence adjustment. The retained source data are used to construct aligned generalized and source-specific templates. For recognition, SQ-HAF fuses a five-subband harmonic reference CCA score with generalized source template and source-specific template scores. We evaluated SQ-HAF using leave-one-subject-out validation on the 35-subject Benchmark and 70-subject BETA datasets. Under the 1.0 s protocol with one labeled calibration trial per stimulus, the complete SQ-HAF configuration achieved 81.57% accuracy (ACC) and 227.83 bits/min information transfer rate (ITR) on Benchmark, and 65.56% ACC and 163.42 bits/min ITR on BETA. In a matched 0.5 s analysis with two calibration trials per stimulus, the five-subband harmonic reference design improved ACC over single-band processing on both datasets after Holm correction. These results indicate that target-relevant source screening and complementary harmonic/template evidence can support low-calibration cross-subject SSVEP decoding.

Manual sorting is still the mainstream scheme for component identification in down feather quality evaluation, which suffers from low detection efficiency and poor stability. To address these drawbacks, this paper proposes a fine-grained component detection algorithm based on YOLO12 for down feather quality classification. Five typical down feather components are selected as detection targets, including discolored feathers, down filaments, immature down, feathers and pure down. A dedicated down feather object detection dataset consisting of 1140 images is established accordingly. To improve the feature representation capacity of the model for tiny objects, faint boundary features and subtle distinctions between analogous categories, this work integrates the Convolutional Gated Linear Unit (CGLU) into the A2C2f module of YOLO12. While maintaining the original feature aggregation pathways and residual architecture, the conventional MLP feed-forward branch within ABlock is replaced with convolutional gated transformation. Experimental results demonstrate that the proposed A2C2f-CGLU model achieves precision of 96.50%, recall of 94.49%, mAP50 of 98.05% and mAP50-95 of 57.89% with the optimal weights on the validation set. Compared with the original YOLO12, the mAP50-95 metric is elevated by 3.04 percentage points, and the overall performance surpasses two comparative variants, A2C2f-DFFN and A2C2f-KAN. Visualizations of PR curves, confusion matrices and real test samples validate that the proposed method effectively enhances the recognition stability of tiny down feather targets and similar classes. This research provides a visual inspection foundation for subsequent component proportion calculation, quality grade discrimination and the development of intelligent detection systems.

As core weighing equipment in the logistics and industrial sectors, the accuracy of truck scales is significantly affected by environmental noise, sensor errors, and nonlinear factors. This paper proposes an adaptive smoothing-constrained broad learning system (PSO-MCC-SCBLS) to enhance the precision and robustness of truck-scale weighing. Particle swarm optimization (PSO) is employed to optimize the number of feature windows, feature nodes, enhancement nodes, and the smoothing coefficient of the SCBLS; the maximum correntropy criterion (MCC) replaces the minimum mean square error (MMSE) criterion for training the output-weight matrix; and a smoothing constraint derived from the physical continuity of the weighing system is introduced to improve generalization in small-sample scenarios. The method was validated on real data from an 8-channel, 40-ton truck scale under a corrected evaluation protocol that uses group-wise five-fold cross-validation, selects all hyperparameters by an inner cross-validation on the training folds only, and matches the effective regularization strength across MCC and non-MCC variants. A complete component ablation over the seven BLS-family variants (BLS, MCC-BLS, SCBLS, PSO-BLS, PSO-SCBLS, PSO-MCC-BLS and the full model) is reported alongside RBLS, CatBoost, CNN_Attention and LSTM, together with anti-interference tests under a composite Gaussian-plus-impulsive disturbance injected in three scenarios: disturbed calibration only (A), disturbed calibration and deployment (B), and disturbed deployment only (C). The results delimit the contribution of each component. Automated structural search is the one component whose benefit is large and consistent, reducing clean-data RMSE by 47.7% over a plain BLS. On clean data the full model does not lead: PSO-MCC-BLS attains the lowest RMSE ($0.0209\times {10}^3$ kg) while the full model records $0.0292\times {10}^3$ kg, indicating that under a leakage-free protocol this calibration task is already close to a low-complexity regime. The advantage of the full model is specific and is reported as such: in Scenario B at a noise ratio of 0.5 it achieves the lowest RMSE ($1.7670\times {10}^3$ kg) and the best Friedman rank among all eleven models ($p<0.05$), whereas at the weaker intensity and in Scenario C the deep-learning baselines lead. Once the effective regularization is matched, the MCC term contributes negligibly on this dataset, so the observed robustness is attributable to the tuned BLS-family model as a whole rather than to correntropy weighting in isolation.