With protected agriculture increasingly developing toward large-scale, multi-zone, and intelligent production, reducing water and fertilizer consumption, drainage losses, and energy waste while maintaining crop yield, product quality, and long-term economic profitability has become an important challenge for sustainable greenhouse production. To address sensor abnormalities, spatial heterogeneity among irrigation zones, competition for shared hydraulic resources, and the limited ability of conventional control strategies to jointly optimize resource efficiency and economic performance, a sensor-driven spatiotemporal graph multi-agent reinforcement learning framework, termed SensorGraph-FertiMARL, is proposed for state sensing, response prediction, and sustainable coordinated fertigation optimization in multi-zone greenhouses. Reliable zone-level states are first obtained through sensor-reliability gating, missing-state reconstruction, and multi-relational graph fusion incorporating spatial, hydraulic, and microclimatic relationships. An action-conditioned fertigation–yield–quality–profit digital twin is then constructed to jointly predict root-zone responses, resource consumption, marketable yield, product quality, and net profit associated with candidate control actions. Finally, graph-based multi-agent proximal policy optimization (Graph-MAPPO) with capacity, safety, and fairness constraints is employed to coordinate fertigation allocation among multiple zones under shared pump, main-pipeline, and nutrient-solution resource constraints. Evaluated through rigorous offline digital-twin simulation and replay testing calibrated against complete multi-stage historical dataset records from an eight-zone commercial greenhouse, experimental results show that the proposed method reduced water and fertilizer use to $75.8\%$ and $78.3\%$ (achieving water and fertilizer saving rates of $24.2\%$ and $21.7\%$), respectively, relative to the uniform timer-based control baseline, while decreasing the drainage rate to $7.2\%$. The Grade-A product rate increased to $84.6\%$, net profit reached $138.9\%$ of the baseline level (a $+38.9\%$ net-profit improvement), return on investment (ROI) reached $1.94$, and the stress-violation rate was limited to $2.9\%$. For state sensing and digital-twin prediction, the sensor-state reconstruction mean absolute error (MAE) and coefficient of determination ($R^2$) were $0.036$ and $0.918$, respectively; the water-content prediction MAE was $0.041$, the electrical conductivity ($EC$) prediction root mean squared error (RMSE) was $0.119$, and the drainage prediction mean absolute percentage error (MAPE) was $8.7\%$. In addition, the $R^2$ for marketable-yield prediction reached $0.912$, the macro-averaged F1 (Macro-F1) score for quality-grade prediction reached $0.851$, the $R^2$ for net-profit prediction reached $0.904$, the mean prediction interval width (MPIW) was $0.142$, and the prediction interval coverage probability (PICP) reached $0.931$. These results demonstrate that SensorGraph-FertiMARL can simultaneously achieve water and fertilizer conservation, drainage reduction, operational-risk control, and economic improvement while maintaining marketable production and quality, thereby providing an effective intelligent optimization approach for resource-efficient, environmentally sustainable, and economically viable greenhouse production.

Muscle onset detection is a fundamental problem in electromyography signal analysis, human–machine interaction, and rehabilitation assessment. In medical and biomedical applications, slow muscle activation onset processes are widely encountered in scenarios such as rehabilitation training, postural regulation, and fine motor control. Such processes are typically characterized by slowly varying amplitudes, long temporal durations, and high susceptibility to noise interference, which poses significant challenges for accurate identification of onset timing. To address these issues, a lightweight temporal attention method for slow muscle activation onset detection is proposed and systematically validated under multimodal experimental settings. The proposed method takes surface electromyography signals as the primary input, while synchronously acquired optical motion image data are incorporated into the experimental design and result analysis, thereby aligning with the common joint use of optical imaging and physiological signals in medical and biomedical research. From a methodological perspective, the proposed framework is composed of lightweight temporal feature encoding, a slow activation-aware temporal attention mechanism, and noise suppression with stable decision strategies. Under the constraint of low computational complexity, the ability to model progressive activation signals is effectively enhanced. Experiments are conducted on a dataset containing multiple types of slow activation movements, and model performance is evaluated using five-fold cross-validation. The results demonstrate that under regular signal-to-noise ratio conditions, the proposed method significantly outperforms traditional threshold-based approaches, classical machine learning models, and several deep learning baselines in terms of onset detection accuracy, recall, and precision. Specifically, onset detection accuracy reaches approximately $92\%$, recall is around $90\%$, and precision is approximately $93\%$. Meanwhile, the average onset detection error and detection delay are reduced to about $41\mathrm{ms}$ and $28\mathrm{ms}$, respectively, with the false positive rate controlled at approximately $2.2\%$. Stable performance is further maintained under different noise levels and cross-subject settings, indicating strong robustness and generalization capability.

With the development of smart classrooms and digital learning platforms, multimodal learning behavior data provide a new sensing basis for understanding students’ cognitive states and psychological risk warnings. However, existing educational data mining methods mainly focus on performance prediction, dropout warning, or surface-level emotion recognition, while continuous and interpretable modeling of deeper cognitive biases and related psychological risks remains insufficient. To address this issue, we propose MLBS-Net, a multimodal learning behavior sensing network for teaching feedback that jointly models students’ textual expressions, behavioral sequences, classroom interactions, and psychological auxiliary signals. MLBS-Net integrates theory-guided textual cognitive bias encoding, temporal behavioral state modeling, and reliability-aware cross-modal fusion to capture psychologically interpretable cognitive patterns, characterize dynamic learning-state changes, and adaptively integrate multimodal information according to data quality and task contribution while providing interpretable feedback for teachers. Experimental results show that MLBS-Net achieves a Macro-F1 of 0.855 for cognitive bias recognition and an AUC of 0.891 for psychological risk warning, outperforming traditional machine learning, unimodal deep learning, and standard multimodal methods. Ablation results further support the effectiveness of theory-guided semantic encoding, temporal behavioral modeling, reliability estimation, and multitask learning. These findings demonstrate that MLBS-Net can jointly characterize cognitive biases and potential psychological risks from multisource learning behaviors, providing a feasible approach for learning-state sensing, risk warning, and interpretable teaching support in smart education.