In semi-supervised lesion segmentation， the performance of the teacher network is poor， making it difficult for it to guide the student network to perform effective segmentation. To address this issue， an efficient semi-supervised medical image lesion segmentation method was proposed， employing the medical segment anything model （MedSAM）， which exhibited superior feature extraction capabilities， as the teacher network. A lightweight student network based on Mamba was constructed， and its segmentation performance was enhanced through knowledge distillation. To address the semantic mismatch caused by feature alignment across heterogeneous networks， a perturbation-consistent cross-architecture knowledge distillation method was introduced. This approach mapped teacher features to the student feature space and aligned perturbation responses， thereby improving the student network’s feature representation ability and improving segmentation performance. Additionally， to tackle the challenges of diverse lesion morphologies and low foreground-background contrast， leading to poor segmentation consistency， a distribution-based self-supervised loss was proposed for optimization. Experiments on multiple types of medical image lesion segmentation datasets demonstrate that the proposed method in this paper outperforms existing methods in segmentation performance. Meanwhile， the student network has only 1.34 M parameters， which significantly improves the model efficiency.

In medical image processing， the reliance on independent encoding components makes it impossible to achieve joint optimization of data compression and machine vision tasks. To address this issue， an end-to-end machine vision task-driven medical image compression network （MVMICNet） was proposed， achieving harmonious unification of data compression and medical image analysis in an end-to-end manner. To maintain the performance of machine vision tasks before and after medical image compression， a task-aware improved code rate-accuracy loss function was designed. By introducing task-related loss terms， it dynamically balanced the relationship among code rate， reconstructed image distortion， and machine vision task accuracy during the optimization process. Furthermore， the MVMICNet model adopted a stage-wise training approach， specifically optimizing for the different characteristics of machine vision tasks to ensure that the model can accurately capture the feature information crucial for diagnosis. This has achieved a simultaneous improvement in compression efficiency and task performance， thus demonstrating superior robustness in complex medical application scenarios. Finally， the effectiveness of the framework was verified in semantic segmentation and object detection tasks.

Medical image segmentation is an important technical basis for clinical diagnosis and treatment. Accurate medical image segmentation contributes to improving the accuracy and efficiency of disease diagnosis. Deep learning methods have made significant progress in this field. However， these methods are heavily dependent on manually labeled data， and the high cost of obtaining high-quality segmentation labels limits their practical application. Semi-supervised learning effectively alleviates the label scarcity problem by combining a small amount of labeled data with a large amount of unlabeled data. Mean teacher （MT） is a mainstream semi-supervised learning method. It leverages an exponential moving average to extract information from unlabeled data， enhancing model accuracy and generalization performance. Currently， MT has been widely adopted in medical image segmentation. In this paper， MT was comprehensively reviewed， focusing on its application and improvement in the medical image segmentation field from aspects such as consistency regularization， uncertainty， attention mechanism， multi-task learning， auxiliary correction， and model variants. Furthermore， trends in the application and enhancement of MT were briefly analyzed， and commonly used comparative experimental methods， datasets， backbone networks of MT， and evaluation metrics in medical image segmentation were summarized. Finally， existing challenges and potential future research directions in applying MT to medical image segmentation were discussed.

Existing graph representation learning methods for electronic health records （EHR） primarily rely on local information of a single patient， overlooking potential associations among patients in disease progression and treatment pathways. This limits the models’ generalizability and robustness. To address this issue， a hybrid multi-level graph neural network （H-MGNN） model was proposed， and it was applied to mortality prediction for intensive care unit （ICU） patients. The model constructed a patient-patient graph （P-P） at the macroscopic level and a taxonomy-note-word hypergraph （T-N-W） at the microscopic level， while incorporating temporal dependencies within the hypergraph to achieve multi-scale fusion of patient features. Meanwhile， a hybrid embedding （Hybrid-E） algorithm was designed to extract and integrate latent patient features and improve the prediction accuracy. Experimental results demonstrate that H-MGNN on the medical information mart for intensive care Ⅲ （MIMIC-Ⅲ） dataset significantly outperforms existing methods in terms of in-hospital mortality prediction and other tasks， validating its effectiveness and superiority in complex EHR data mining.

To address the challenges of inadequate spatio-temporal feature modeling and poor robustness in complex scenarios for non-contact blood oxygen saturation （SpO₂） measurement using remote photoplethysmography （rPPG），a trend-aware spatio-temporal fusion network （TAST-Net） was proposed. The proposed network adopted an innovative dual-branch fusion architecture that synergistically fused local physiological features extracted by a 3D convolutional neural network （3D CNN） branch with global spatio-temporal dependencies captured by a video vision transformer （ViViT） branch. To enhance the model’s sensitivity to signal dynamics， a weighted composite loss function combining mean squared error （MSE） and Pearson correlation loss was designed. Experimental results on two public datasets demonstrate the superior performance of TAST-Net. On the pulse rate estimation （PURE） dataset， it achieves a root mean squared error （e_{\mathrm{R}\mathrm{M}\mathrm{S}}） of 0.53%， a mean absolute error （e_{\mathrm{M}\mathrm{A}}） of 0.37%， and a Pearson correlation coefficient （R） of 0.96. On the more challenging visual information processing and learning-heart rate （VIPL-HR） dataset， the e_{\mathrm{R}\mathrm{M}\mathrm{S}}， e_{\mathrm{M}\mathrm{A}}， and R reach 0.84%， 0.57%， and 0.82， respectively， outperforming other comparative methods. These findings indicate that TAST-Net provides an effective solution for accurate and robust SpO₂ estimation from facial videos and validates the advantage of integrating local and global features in rPPG signal processing.

Determining key injection parameters of bone cement， such as volume， pressure and angle， in percutaneous vertebroplasty still largely relies on empirical judgment， and they have a critical role in preventing leakage and related complications. An advanced simulation framework that integrated a multiphase non-Newtonian fluid model was introduced. The framework employed a Voronoi-based biomimetic bone representation to predict cement flow behavior within trabecular structures， enabling high-fidelity visualization and analysis of cement dispersion under varying operating conditions. Experimental results demonstrate the framework’s potential value in preoperative planning optimization， surgical training assistance， and leakage risk assessment. For preoperative use， it can assist clinicians in rehearsing injection procedures and refining operation strategies； in training scenarios， it can simulate cement dispersion patterns under different parameter settings， thereby helping novice surgeons understand procedural logic； furthermore， its capability to predict leakage risks holds promise for providing technical support to improve clinical surgical safety.

A robust fault diagnosis algorithm based on an observer was designed for the nonlinear dynamic system of distributed drive electric vehicles with actuator faults. Firstly， by considering the actuator fault and system disturbance， a mathematical model of 7 degrees of freedom vehicle system was established. Then， a fault diagnosis strategy based on a nonlinear observer was designed by using linear matrix inequality and Lyapunov function to realize multiple estimations of system state， disturbance information， and fault factor， and the relevant theoretical proof was given. Finally， by referring to the actual driving conditions of the vehicle， the co-simulation was carried out by MATLAB/Simulink and CarSim. The results show that the designed observer can accurately diagnose and estimate the actuator faults， rapidly track the actual operating state of the system， and improve the anti-interference ability of the system.

In order to solve the problem of voltage deviation and current distribution imbalance in direct current （DC） microgrid under false data injection attack， a distributed resilient cooperative control method was proposed by taking the islanded DC microgrid with multiple distributed generators as the research object. This method could effectively eliminate the impact of false data injection attacks and will not interfere with the operation of the system under normal circumstances. The Lyapunov stability theory was used to prove that the DC microgrid can operate stably when it is attacked by any constant false data injection， achieving the two control objectives of voltage regulation and current distribution. The simulation model was built by MATLAB/Simulink， and the validity of the control method was verified.

Existing video-text retrieval methods fail to effectively model temporal information and relevance information in a unified manner.To address this issue， a video-text retrieval method based on a cross-modal attention mechanism was proposed.Firstly， embeddings of video frames and text were extracted using a large-scale pre-trained image-text model， and knowledge transfer was leveraged to alleviate the heterogeneity between different modalities.Then， a joint text-frame cross-modal attention module was introduced to simultaneously encode temporal information among video frames and relevance information between video frames and text， enabling the capture of more competitive video representations.Finally， the cross-entropy loss function was used to constrain the model training.Comparative experiments for verification demonstrate that the proposed method can effectively capture temporal and relevance information of video frames， achieving competitive performance on the microsoft research video to text （MSR-VTT） and large-scale movie description challenge （LSMDC） datasets.

In order to meet the specific business needs of different tenants， software as a service （SaaS） usually provides simulation customization functions. Through simulation customization， tenants can personalize SaaS according to their own business needs， so as to better meet their business needs. However， existing tenant customization services have the problems of failing to fully meet tenant requirements and having slow algorithm operation and response speeds. Therefore， an SaaS multi-tenant service simulation and customization technology based on an ant colony optimization （ACO） algorithm and variability traversal was proposed. The simulation optimization service deployment was achieved. By introducing a variability model， the adaptability and reusability of service customization and assembly were realized. The experimental results show that in the evaluation of SaaS tenant service resource utilization， the average values of the proposed technology are slightly higher than those of the comparison algorithm on instances a and b. The execution time fluctuates among different configuration schemes， with the shortest being 1 426 ms and the longest being 1 652 ms. The duty cycle of switching resources is relatively stable， with a fluctuation range between 1.12% and 1.51%. A lower duty cycle means that SaaS can more effectively utilize resources and reduce performance losses caused by resource switching at the same time. Based on the configuration schemes and running time data of different SaaS tenants， it is indicated that tenants can effectively derive service configuration schemes. The proposed technology can provide technical references for optimizing the simulation customization performance of SaaS.

In light of the rapid development of air transportation and information technology， the efficient utilization of massive and heterogeneous aviation safety data in aviation emergency management faces challenges. The problem of knowledge extraction for an aviation accident knowledge graph was studied， specifically named entity recognition and relation extraction， and the following methods were proposed： 1） An improved BiGRU-IDCNN-CRF model based on bidirectional encoder representations from Transformers （BERT） was presented， achieving a named entity recognition accuracy of 94.69%； 2） A reinforcement learning-based clustering distant supervision relation extraction model was constructed， in which data noise was reduced by integrating improved K-means clustering with distant supervision labeling， and the denoising process was optimized via reinforcement learning； a combination of piecewise convolutional neural network （PCNN） and an attention mechanism was applied to achieve a relation extraction accuracy of 84.16%. Experimental results indicate that the quality of information extraction for the aviation accident knowledge graph is effectively improved， providing accurate information support for aviation safety management.

The position of the sintering terminal point is a key parameter that affects the quality and production efficiency of sinter. To improve insufficient guidance， poor timeliness， and weak visualization effect in sintering terminal point prediction， a five-dimensional digital twin model was constructed， including physical entity， virtual environment， multi-step prediction， twin data， and virtual and real connection， which provided process parameter monitoring and optimization guidance for the sintering process. In terms of prediction， the data was first preprocessed， and then the feature variables were screened by grey relation analysis （GRA）. Finally， the deep temporal convolutional network （DeepTCN）by using population based training（PBT） was constructed for multi-step prediction of the sintering terminal point. The experimental results show that the proposed digital twin model has high prediction accuracy under different prediction steps， and it provides advanced ideas and technical methods for digital and intelligent transformation in the sintering field.

Because the sintering process has complex and high-dimensional process variables and many uncertain factors， it is difficult for a single feature selection method to effectively select the best feature set， which affects the prediction accuracy of the model. Therefore， a prediction model of attention mechanism-based bidirectional gated recurrent unit model （BiGRU-Att） sinter drum index based on hybrid feature selection was proposed. Firstly， the maximum information coefficient （MIC） was used to select candidate features from the original feature set. Then， the feature selection method based on simultaneous perturbation stochastic approximation （SPSA-FS） was used to further optimize the candidate feature set. Finally， the best feature set was used as the input of BiGRU-Att to predict the sinter drum index. The results of comparative analysis with multiple models and single feature selection methods show that the hybrid feature selection method proposed in this paper can select the best feature set， and the established model has higher prediction accuracy， providing reliable decision-making support for the sintering process.

The effect of hydrogen charging time on the hydrogen embrittlement （HE） of the multiphase microstructure of QP980 steel was studied by electrochemical hydrogen charging and slow strain rate tensile （SSRT） tests. The stress-strain curve shows that the tensile strength and elongation decrease significantly with the increase of hydrogen charging time， but the presence of hydrogen does not affect the work hardening rate before fracture. The fracture morphology analysis shows that in the absence of hydrogen， the fracture mode at the center of the specimen is a mixed dimple and quasi-cleavage， and the edge region exhibits a dimple fracture morphology. After hydrogen charging， the mixed fracture zone of the specimen expands， and the unit facet size of quasi-cleavage increases with prolonged hydrogen charging time. Observations of secondary cracks and microstructure revealed that at lower hydrogen concentrations， the phase interfaces between ferrite and martensite served as the primary sites for crack initiation. These cracks propagated along the ferrite-martensite interfaces but were blunted by the ferrite phase， resulting in microvoid-coalescence type cracks. This indicates that the hydrogen-enhanced localized plasticity （HELP） mechanism was the dominant hydrogen embrittlement mechanism under these conditions. When hydrogen concentration is high， the cracks transform into hairline cracks and pass through the matrix ferrite structure， indicating that the hydrogen-enhanced decohesion （HEDE） is the dominant mechanism.

A numerical simulation was conducted to investigate the solidification characteristics inside the melt pool during twin-roll thin strip casting under side dam vibration. The effects of vibration frequency， amplitude， and casting temperature on molten steel solidification in the melt pool were analyzed. The mechanical vibration experiment was designed to validate the accuracy of the numerical model. Results show that applying vibration accelerates molten metal solidification rate， and higher vibration frequency further enhances the solidification rate. Due to the high heat flux density at the side dam， the solidification rate of the molten steel increases， leading to a decreasing trend in liquid fraction from the side dam to the center of the melt pool along the horizontal direction. Elevated casting temperature reduces the melt pool’s mushy zone size and lowers the Kiss point position. The study finds that the optimal vibration frequency is 13 Hz， the optimal amplitude is 0.5 mm， and the casting temperature should not exceed 1 860 K.

To investigate the influence of structural parameters on the heat transfer characteristics between integral spiral finned tube bundles， 13 groups of such tube bundles with varying structural parameters were designed and fabricated. A 1∶1 experimental bench for heat transfer characteristics was established to conduct experimental research on the heat transfer characteristics of these tube bundles. The findings indicate that the Nusselt number （Nu） on the gas side of the tube bundles improves with an increase in flue gas velocity， fin height， fin pitch， and longitudinal pitch within the production’s adjustable range. Conversely， the transverse pitch does not significantly affect the Nu on the gas side of the tube bundles. Based on the experimental data， a correlation equation describing the Nu on the gas side of the integral spiral finned tube bundles is proposed， providing a theoretical foundation for subsequent research on the optimization of numerical calculations of the thermal parameters of these tube bundles.

In this paper， a mathematical model for predicting the flow and solidification of the mold under the control of the composite electromagnetic field of electromagnetic stirring and electromagnetic braking was established to address issues related to solidification homogeneity and surface defect control in continuous casting slabs. The wall-adapting local eddy-viscosity （WALE） large eddy simulation model， in conjunction with a simplified inclusion capture criterion， was employed to analyze the mechanism by which the composite electromagnetic field controlled surface defects in the cast slabs. The findings reveal that the capture locations of inclusions are mainly concentrated in the outer surface layer of the slabs； the difference between the thickness of the solidified shell in the wide and narrow surface under the composite magnetic field is reduced by 4.59% compared with that without electromagnetic field， and the total amount of inclusion capture under the composite magnetic field is reduced by 47.21% compared with that without electromagnetic field. It is found that electromagnetic braking inhibits the inclusion capture in the solidified shell， and the core role of electromagnetic stirring is to promote the flow and solidification homogeneity.

Existing studies overlook the well-rock coupling process and deep wellbore heat transfer effects， making it difficult to accurately evaluate the heat extraction performance of enhanced geothermal system （EGS） in hot dry rock. To address this issue， a three-dimensional thermal flow coupling model for well-rock combination based on COMSOL was developed， and the impact of various factors on the heat extraction performance of the EGS was analyzed. The simulation results show that the heat extraction process of EGS primarily relies on heat transfer through fractures and wellbores， with heat exchange through fractures contributing 73% and that through wellbore accounting for 27%. In a single injection and production mode， the best heat extraction occurs at a 500 m spacing between wells， with an outlet temperature reaching 145.5 ℃. Increasing the injection flow rate raises the outlet temperature， but shortens the heat reservoir’s lifespan. Optimizing well layout and flow heat transfer paths reveals that the single injection and production layout has a longer lifespan and higher outlet temperature， while the single injection and four production mode achieves the highest heat extraction， 3.1 times that of the single injection and production.