As image data proliferates rapidly and retrieval demands grow ever more complex, building efficient, accurate image retrieval techniques has become a core challenge in computer vision. A multi-task image retrieval model, integrating dynamic hash learning and a visual attention mechanism, is proposed to overcome the limitations of existing methods in terms of feature expression capability, retrieval efficiency, and multi-task collaborative optimization. Firstly, a dynamic hash center update mechanism is proposed, which learns more discriminative hash representations to build an efficient index foundation for subsequent approximate nearest neighbor search (ANNS), thereby achieving adaptive optimization of the feature space. Secondly, introducing channel attention networks enhances the representation ability of multi-scale features. Finally, parameter sharing strategy is combined to achieve collaborative learning of multi-task features. The experiment outcomes reveal that the model achieves retrieval accuracy of 95.3% and 92.6% under 64 bit hash codes on the Microsoft Common Objects in Context and the National Institutes of Health Chest X-ray Dataset 14, respectively, which is significantly better than traditional methods. In complex scenario testing, the retrieval accuracy of the model for single subject and multi-subject images remained stable at over 90%, and the training speed increased by 25.8% compared to the baseline, while the memory usage rate decreased to 22.4%. The proposed method effectively solves the problems of low efficiency in high-dimensional feature mapping and weak multi-task correlation, providing a solution that combines efficiency and robustness for large-scale image retrieval-tasks.
Digital twin technology is revolutionizing cyber-physical systems (CPS) by enabling real-time system modeling, adaptive control, and intelligent decision-making. This paper presents a novel framework that merges Adaptive Hierarchical Policy Optimization (AHPO)—a new hierarchical reinforcement learning (HRL) algorithm with CPS-Chain, a lightweight blockchain protocol, to develop adaptive digital twins that address core CPS challenges of scalability, latency, and security. AHPO models multi-level system-environment interactions by dynamically decomposing complex CPS tasks into high-level strategic and low-level tactical policies, enabling context-aware control in domains such as smart cities and industrial automation. CPS-Chain facilitates decentralized, tamper-proof data exchange through an optimized Proof-of-Authority (PoA) consensus mechanism, ensuring system reliability under high-frequency data conditions. The proposed framework was evaluated using simulation-based smart traffic management and industrial predictive maintenance scenarios. Traffic and machine telemetry datasets were synthetically generated to emulate realistic cyber-physical system conditions. Experimental results demonstrated 92% failure prediction accuracy, a 25% reduction in traffic congestion, and approximately 0.5-second blockchain transaction latency. These findings highlight the potential of integrating hierarchical reinforcement learning and blockchain technologies to improve the adaptability, security, and resilience of digital twin-enabled cyber-physical systems. Furthermore, the proposed framework provides a scalable foundation for future intelligent CPS applications requiring real-time decision support and secure data management.
This study develops a deterministic model based on ordinary differential equations to investigate the transmission dynamics of mastitis in dairy cattle while accounting for environmental contamination. An optimal control framework is further incorporated to evaluate the effects of five intervention strategies: prevention, detection of subclinical infections, treatment, culling of persistently infected cows, and improved hygiene in cattle housing environments. Stability analysis confirms the mathematical and epidemiological validity of the model, with disease persistence determined by the effective reproduction number (R0). Specifically, when R0<1, the disease-free equilibrium is stable and mastitis can be effectively controlled, whereas when R0>1, the disease persists and continues to spread within the herd. Sensitivity analysis identifies the transmission rate as the most influential parameter, highlighting the critical role of preventive measures in disease control. Using Pontryagin’s Maximum Principle, the study evaluates the effectiveness of various intervention combinations in reducing mastitis transmission. Cost-effectiveness analysis further identifies the most practical and economically viable control strategies. The results demonstrate that integrated control measures substantially reduce the prevalence and spread of mastitis, emphasizing the importance of comprehensive prevention and management programs. In addition, the adoption of advanced therapeutic approaches, such as nanoparticle-based treatments, may provide promising alternatives for mitigating the growing challenge of antimicrobial resistance. Overall, strengthening preventive measures, improving farm hygiene, and implementing effective treatment protocols can significantly enhance dairy cattle health, welfare, and productivity while promoting sustainable dairy production.
Aiming at the problem of low control accuracy and high energy consumption of LED lighting caused by the coupling of multiple factors in indoor environment, a self-adaptive energy-saving control method for indoor environment LED under Internet of Things technology is proposed. Build an IoT control platform consisting of a perception layer, a network layer, and an application layer, using light sensors and human infrared sensors to collect data on regional light intensity and personnel distribution; Using adaptive weighted fusion algorithm to calculate regional illumination values, combining Gaussian filtering and temperature variance to extract human infrared features, and using support vector machine to complete human motion recognition; Using lighting data and personnel behavior patterns as inputs, adaptive adjustment of LED brightness and color temperature is achieved through fuzzy neural networks. The experiment was conducted in a 120 square meter indoor space, equipped with a dimmable LED lighting platform, human infrared sensors, and lighting sensors. The dataset covers six types of personnel behavior patterns: walking, reading, resting, washing, makeup, and eating. The results showed that the proposed method's recognition of walking, reading, resting, and washing behaviors was consistent with actual behaviors, and the personnel positioning trajectory was clear and had no significant deviation; Under six behavioral modes, the power consumption after control decreased from 150, 200, 250, 180, 220, and 200 W · h to 100, 50, 150, 120, 160, and 140 W · h, with energy-saving rates of 33.3%, 75%, 40%, 33.3%, 27.3%, and 30%, respectively. Compared with the wireless sensing intelligent dimming control method and the power line communication high-pressure sodium lamp energy-saving regulation method, the proposed method has lower energy consumption and more significant energy-saving effect, which verifies its effectiveness and practical value in indoor LED adaptive energy-saving control.
