Recognizing emotions from a single eye has been a hard work over time, mainly because muscle movements are hard to perceive and event data is susceptible to noise. Most existing methods concentrated on semantic fusion, neglecting the potential benefits of local interaction between different modalities. Considering this problem, we design a new framework called HEC-Net which hierarchically fuses different modalities’ information. Our HEC-Net begins with a dual-stream Spatio-Temporal Feature Extraction (STFE) module to encode texture and motion while suppressing noise via Top-k selection, which follows a “perceive and select” design philosophy. Then we form a three-stage local-to-global fusion process to fuse the information in a more layered approach. The whole approach includes a multi-window fusion, a pyramid structure and a global fusion. The multi-window fusion constrains interactions locally, enabling the model to firstly concentrate on details. We employ a pyramid structure to capture the changes in blinking actions over time. Finally, the global fusion process aggregates dispersed cues into a coherent global representation. HEC-Net achieves a state-of-the-art UAR of 92.30% on the SEE dataset, which remains stable and efficient under various lighting conditions.

This paper develops a local discontinuous Galerkin (LDG) method based on generalized numerical fluxes for solving traveling wave solutions of the modified Buckley-Leverett equation. To achieve efficient computation, the original equation is reformulated into a first-order system by introducing auxiliary variables, followed by spatial discretization using the DG method, while the explicit third-order Runge-Kutta method is adopted for temporal discretization. Based on the antisymmetry of the discrete spatial operator, the stability of the scheme under the energy norm is rigorously established. Numerical experiments demonstrate the robustness of the proposed method in handling convection-dominated problems with Riemann initial data, confirming its capability to accurately capture the shock structures.

To improve the accuracy of imputing missing maintenance engineering data, an ensemble learning method integrating three algorithms—XGBoost, Support Vector Machine (SVM), and Multilayer Perceptron (MLP)—is proposed. This method constructs three base learners (XGBoost, SVM, and MLP) to establish the nonlinear mapping relationship between maintenance measures and multi-dimensional influencing factors through supervised learning. A soft voting ensemble strategy based on probability weighting is adopted to optimize the comprehensive decision-making effect of the model output, and the imputation performance of each model is systematically evaluated. The research results show that the proposed ensemble learning method achieves an accuracy of 99% in missing data imputation, which is significantly superior to the single models MLP (54%), SVM (72%), and XGBoost (77%). This verifies the effectiveness and superiority of the method in imputing missing maintenance data.

The popularization speed of generative artificial intelligence technology is accelerating. The Retrieval Augmented Generation (RAG) system, with its feature of enhancing output accuracy by integrating external knowledge bases, has been widely applied in fields such as intelligent question answering and professional document generation. This study first conducted a correlation analysis and then carried out comparative experiments using multiple machine learning algorithms. The results showed that the KELM-LSTM-Transformer algorithm proposed in this paper demonstrated significant advantages in the comparison: Its accuracy rate, recall rate, precision rate and F1 value all exceed 99.5%, with some indicators approaching 100%, and the AUC is also close to 100%. This algorithm not only significantly outperforms relatively weak algorithms such as AdaBoost, decision tree, and KNN, but also outperforms medium and high-performance algorithms like GBDT, ExtraTrees, SVM, and XGBoost. Even compared with the outstanding random forest and multi-head attention LSTM, its core classification index still maintains a higher level, and the AUC index has a particularly prominent advantage, fully demonstrating its superiority in complex feature capture and sequence information modeling. This achievement provides an efficient algorithmic solution for the performance optimization of the RAG system and also offers a reference for technical research in related fields. It is of great significance for promoting the precise application of generative AI in professional scenarios.

To address low accuracy and poor robustness in tomato maturity detection under complex agricultural environments (light variations, occlusion, multi-scale targets), this study proposes a lightweight multi-scale differential fusion network Tomato Net-MSF. Its core innovation is scale-customized feature enhancement modules: (1) Small-scale KSFA module (multi-core selection + spatial-spectral attention) suppresses background and improves small-target recognition; (2) Medium-scale C3k2l module (2*2 simplified convolution + dual-bottleneck stacking) enhances fusion stability; (3) Large-scale LSConv module (global perception + dynamic aggregation) refines target contours. Experiments on Laboro Tomato dataset show: mAP50 0.75 (8% higher than baseline YOLOv11), F1 0.72 (6% higher), and 50FPS inference speed (20ms/frame) on RK3588NPU (5-fold acceleration). The model optimizes multi-scale detection balance, integrating high precision, robustness and edge deployment capability, providing an efficient solution for real-time tomato maturity detection in complex scenarios.

With the acceleration of the global energy transition process and the wide popularization of smart homes, household electricity consumption data is showing high-frequency and multi-dimensional characteristics. Hourly electricity consumption prediction has gradually become a key technical link to support the load dispatching of smart grids and help users achieve energy-saving management. Machine learning algorithms, with their powerful nonlinear fitting capabilities and efficient feature learning abilities, provide effective solutions for temporal power consumption prediction. This paper proposes the HLOA-CNN-BIGRU regression algorithm. Firstly, correlation analysis and violin plot analysis are carried out. Then, through comparative experiments of multiple machine learning algorithms, it is found that the KELM-LSTM-Transformer algorithm shows significant advantages in all indicators: Its MSE is 0.002, which is much lower than the minimum value of 0.006 in other algorithms. The RMSE is 0.047, significantly lower than the minimum value of 0.075 in other algorithms. The MAE is 0.038, which is significantly better than the minimum value of 0.056 in other algorithms. The MAPE is 2.28%, which is lower than the minimum value of 3.49% in other algorithms. R² is 99.4%, which is higher than the maximum value of 98.7% in other algorithms. Among traditional algorithms, ExtraTrees has a relatively superior overall performance, with R² reaching 98.7%, RMSE 0.075, and MAPE 3.58%, outperforming algorithms such as AdBoost and Random Forest. The overall performance of GBDT is relatively weak, with R² being only 96.8%, and MSE, RMSE, MAE, and MAPE are all the highest among all traditional algorithms. The R² of CatBoost is 97.8%, and its various error indicators are also higher than those of algorithms such as ExtraTrees and AdBoost. Overall, the KELM-LSTM-Transformer algorithm comprehensively outperforms other compared machine learning algorithms in terms of prediction accuracy and fitting effect, with stronger performance, providing important technical support for the efficient operation of smart grids and the refined management of household energy.

Translation in the medical field is an important research direction for large language models. Medical translation texts are usually very professional and have different meanings for the same word. We have very high requirements for the trustworthiness of translation models. But the existing methods are not good at making professional term accuracy and semantic consistency unified. Medical experts can hardly trust the decisions made by the models. For this reason, this study puts forward an interpretable framework with two-stage LoRA fine-tuning based on LLaMA-3-Chinese-8B-Instruct-v3—TSX-MedTrans. The first stage is term-level fine-tuning to make the representation of professional terms better. The second stage is sentence-level fine-tuning to make the context semantic coherence better. This framework also adds an interpretability module. It is used to show the attention distribution visually, so as to check the model's decision-making mechanism. The experimental results show that TSX-MedTrans improves the translation quality a lot: after term-level fine-tuning, BLEU rises from 29.624 to 34.728, and COMET rises from 0.836 to 0.891; after sentence-level fine-tuning, BLEU rises from 34.583 to 38.232, and COMET rises from 0.846 to 0.876.

During the training process of federated learning, the model parameters uploaded by clients or the changes brought by parameter updates may still leak the relevant data information of the corresponding clients. As a mainstream privacy-preserving technology at present, differential privacy usually adopts the operation of adding uniform noise to parameters. However, this method often leads to slower convergence speed of model training, unsatisfactory model accuracy, and increased communication overhead. To address these issues, this paper proposes a dynamic differential privacy federated learning method guided by gradient-increment mask based on sensitivity. Starting from the update status of model parameters, this method establishes a sensitivity scoring mechanism according to the amplitude of parameter updates and gradient information, which is used to measure the degree of privacy leakage risk loss of different parameters in the training process. On this basis, a gradient-increment mask mechanism is added to selectively protect the parameters with high sensitivity scores, so as to reduce the unnecessary disturbance of noise distribution on low-sensitivity parameters. Furthermore, this paper constructs a dynamic privacy budget scheduling strategy guided by sensitivity, which flexibly adjusts the noise intensity according to the parameter sensitivity distribution to achieve refined and rational allocation as well as efficient utilization of privacy budget. Based on the analysis of these aspects, to resolve the contradiction between global noise injection and selective protection in differential privacy, this paper proposes a sensitivity-guided privacy budget allocation algorithm. This algorithm identifies the key parameters for model training through the mask consensus mechanism and allocates stronger noise protection to these parameters to meet the requirements of dynamic differential privacy. Theoretical analysis shows that this scheme satisfies the guarantee requirements of differential privacy. Experimental results indicate that under the same privacy budget, compared with the conventional full-coverage differential privacy federated learning, the model accuracy trained by this scheme in the same number of rounds is 4%-6% higher on several benchmark datasets, which basically proves the feasibility of the scheme proposed in this paper.

This paper investigates the local regularity of weak solutions to a class of inhomogeneous generalized Laplace equations where the right-hand side depends on the weak solutions themselves. The core innovation lies in handling the strong coupling between the inhomogeneous term of the equation and the weak solution. We first establish the local boundedness of the solution as a key a priori estimate for the analysis. On this basis, by constructing ingenious test functions and applying delicate iterative techniques, we overcome the additional nonlinear interference caused by the inhomogeneous term, and finally prove the Harnack inequality satisfied by the weak solution in the interior domain. The results of this study provide a theoretical tool for analyzing the more refined properties of related nonlinear problems.

The Elephant Trunk demonstrates unique advantages in complex environments thanks to its flexibility, high degree of freedom and versatile transport characteristics. In the industrial world, it is known as nature's universal robotic arm. However, the ability to achieve three-dimensional smooth movement and stiffness adjustment similar to that of the elephant trunk puts forward higher requirements on the structural design, drive integration and functional adaptability of the robot. Here I proposed and implemented a bionic elephant trunk robotic system that incorporates a modular architecture, soft actuators, and adjustable stiffness mechanisms, with 3D motion achieved by three 120 degrees soft bellows actuators working in concert, and thus possessing continuous deformation and spatial localization capabilities. In addition, in order to cope with the demand of rigid-flexible switching for different grasping tasks, the system introduces a variable stiffness mechanism based on skeleton-particle coupling, which is capable of state adjustment and structural toughening under external loads. In the end-operation environment, the robot integrates soft suction cups with adsorption functions. This effectively improves the stability of non-regular target attachment and grasping. The system is modeled using SOLIDWORKS software and the soft actuator is constructed using 3D printing and silicone injection moulding. Experimental results show that the robot exhibits excellent adaptability, load capacity, and locomotion in multi-scenario operations. It shows its wide potential in the direction of complex environment manipulation such as service robots.