Accurate bearing fault diagnosis under varying operating conditions remains crucial for industrial reliability; however, the scarcity of labeled target data continues to significantly hinder the generalization of deep learning models. Existing unsupervised domain adaptation techniques often exhibit negative transfer by failing to distinguish between domain-invariant and domain-specific knowledge, leading to suboptimal adaptation. Therefore, this study proposes the M ultisource D isentangled E xpert A daptation Net work (MDEA-Net). MDEA-Net employs distinct shared expert networks to capture common fault characteristics across multiple source domains and private expert networks to model source-specific knowledge. Importantly, it explicitly disentangles these representations using an expert disentanglement loss, which raises both diversity and orthogonality. For efficient target adaptation, MDEA-Net utilizes frozen shared experts and introduces lightweight, target-specific private experts using low-rank adaptation. Domain alignment is accomplished by minimizing the maximum mean discrepancy specifically between shared feature spaces. Extensive experiments on benchmark datasets (Case Western Reserve University, Jiangnan University, Xi’an Jiaotong University and Changxing Sumyoung Technology Co., Ltd) demonstrate that MDEA-Net significantly outperforms state-of-the-art methods in cross-domain bearing fault diagnosis.

Understanding complex chemical reaction cascades remains a major challenge in chemistry. Scalable investigation of their thermodynamic and kinetic properties requires the use of automated reaction prediction tools, which is a rapidly growing area in the study of chemical reactivity. However, the systematic exploration of intricate reaction networks involving highly reactive intermediates continues to pose significant difficulties. Here, we introduce RxnNet, a novel artificial intelligence-assisted platform for the automated prediction of chemical reaction mechanisms. RxnNet integrates heuristic rules with domain-specific chemical knowledge including stereochemistry, regiochemistry, conformational preferences, and isotope labeling, to construct mechanistically informed reaction networks. These networks are represented as graphs and are coupled with on-the-fly quantum chemical evaluations to identify all feasible intermediates and transition states. In this work, we apply RxnNet to carbocation chemistry, a notoriously complex and computationally demanding type of reaction. We demonstrate the method's capabilities by analyzing three multistep reactions with known mechanisms, each of which poses significant challenges even for expert computational and synthetic chemists. RxnNet provides a robust approach for uncovering reaction mechanisms, which can accelerate the understanding and design of transformations in complex chemical systems.

The emergence of the Internet of Things (IoT) and Artificial Intelligence (AI) technologies has transformed the agricultural industry by offering promising solutions to issues relating to crop recommendation and soil nutrient prediction systems. Due to the urgent need for sustainable agricultural practices, IoT and AI-based soil nutrient prediction and crop recommendation systems have drawn significant attention recently. Recent advancements have introduced innovative models capable of monitoring soil health, predicting nutrient deficiencies, and recommending suitable crop varieties tailored to a specific environment with varying climatic conditions. This paper presents recent developments in IoT and AI-driven technologies that improve smart farming by providing data-driven support and real-time monitoring of soil health. AI models are capable of analysing data from sensors, satellite imagery, and past history to accurately predict soil nutrients and recommend crops, ensuring efficient and sustainable agricultural practices. Despite the advancement in technology, data quality, model interpretability, cost, and accessibility pose challenges that hinder the widespread adoption, particularly among smallholder farmers. The future of smart farming systems lies in overcoming existing barriers and advancing technology to offer scalable, affordable, and user-friendly solutions. Various methodologies and approaches, such as hybrid and ensemble models that combine data-driven AI methods with domain-specific agronomic knowledge, have demonstrated improved reliability and accuracy. With emphasis on some of the important soil parameters such as nutrients, moisture, power of Hydrogen, temperature, relative humidity, and electrical conductivity, this paper discusses the roles of IoT and AI in enhancing the efficiency and precision of Smart Farming. Furthermore, this paper provides insight into current trends and techniques in Smart Farming by synthesising findings from a range of studies on advanced technologies.

Teacher Self-Efficacy, a key psychological construct grounded in Bandura's social cognitive theory, is critical in shaping instructional effectiveness, motivation, and professional development. Understanding the cognitive and knowledge-based antecedents of self-efficacy is particularly important during early teacher training. This study examines the extent to which pre-service English as a Foreign Language (EFL) teachers' Technological Pedagogical Content Knowledge (TPACK) predicts their teaching self-efficacy. An explanatory sequential mixed-methods design was employed, integrating quantitative data from standardized TPACK and Teacher Self-Efficacy scales (N=271) with qualitative data from focus group discussions. Structural Equation Modelling (SEM) via SmartPLS was used to analyze predictive relationships, while thematic analysis explored participants' perceptions. The findings indicate a significant positive association between overall TPACK and teacher self-efficacy. Among the TPACK dimensions, Technological Pedagogical Content Knowledge (TPCK) was the strongest predictor, followed by Pedagogical Content Knowledge (PCK) and Technological Content Knowledge (TCK), while Technological Knowledge (TK) and Content Knowledge (CK) were non-significant. Qualitative findings revealed structural and contextual barriers to technology integration, including limited integrated training, insufficient hands-on experience, technological usability challenges, and unclear policies regarding technology use. These findings advance the psychological understanding of teacher development by identifying the specific cognitive knowledge domains that underlie self-efficacy beliefs and by highlighting systemic obstacles that hinder their enactment. The study advocates for a more integrated and practice-oriented approach to EFL teacher education to enhance both instructional competence and psychological readiness for technology- integrated classrooms.

This study examined changes in practical intelligence and teacher efficacy among pre-service early childhood teachers following the implementation of gamification-based instruction. The course was applied in an undergraduate early childhood education program at a college in Gyeonggi Province. Pre- and post-survey data were collected and analyzed using paired-samples t-tests. The results indicated statistically significant increases in practical intelligence and teacher efficacy across all subdomains. Notably, domain-specific knowledge and application showed the largest change in practical intelligence, while personal teacher efficacy increased more than general teacher efficacy. These findings suggest that gamification-based instruction is associated with meaningful changes in pre-service teachers’ practical intelligence and teacher efficacy, indicating its potential as an instructional strategy for supporting professional competence in teacher education.

ABSTRACT Road crack detection is a critical task in infrastructure health monitoring. Deep learning models deployed on edge devices, however, often suffer from information loss during feature extraction and high computational costs, limiting their practicality in real-world engineering applications. To address these challenges, we propose KD-DETR, a lightweight crack detection framework based on the DETR object detector. Its structural optimizations include an Inverted Residual Module with Cross-Global Attention (IRMB_CGA) to efficiently capture long-range dependencies for cracks in complex road backgrounds, a FreqFusion-BiFPN module that enhances crack edge responses via adaptive frequency-domain filtering and improves multi-scale feature fusion through dynamic recalibration, and a model lightweighting strategy that uses channel attention-guided feature distillation to achieve both model compression and effective domain-specific knowledge transfer.Compared to the original detector RT-DETR, our KD-DETR reduces the number of parameters by 6.36% and computational FLOPs by 18.5%, effectively balancing detection precision and efficiency. Experiments on road defect datasets demonstrate that KD-DETR achieves superior mIoU, lower parameter count and FLOPs, mAP50 of 78.3% and 75.8%, and higher recall for small-scale cracks. Overall, it significantly reduces computational costs while boosting accuracy and recall, enabling efficient, high-precision road defect detection on resource-constrained edge devices.

Accurate prediction of the neurotoxicity of peptides and proteins is critically important for the safety assessment of protein therapeutics and the development of protein-based drugs. Although experimental methods can reliably identify neurotoxic peptides and neurotoxins, they are labor-intensive, costly, and unsuitable for large-scale screening. Existing computational approaches are often limited by shallow feature engineering and suboptimal multimodal fusion strategies, which restrict their predictive accuracy and generalizability in real-world applications. In this study, we propose BiToxNet, a deep learning framework that integrates evolutionary embeddings derived from a protein large language model with ten handcrafted biochemical descriptors through a bilinear attention network (BAN). This design enables effective modeling of cross-modal interactions and residue-level dependencies critical for neurotoxicity prediction. BiToxNet was evaluated on three datasets of different sequence lengths, namely Protein, Peptide, and Combined datasets, achieving accuracies of 92.3%, 96.0%, and 92.7%, respectively, and consistently outperforming existing state-of-the-art methods. Ablation studies confirmed the importance of both evolutionary embeddings and handcrafted features, as well as the critical role of BAN in feature fusion. Visualization analyses using t-SNE and hierarchical clustering further demonstrated that BiToxNet learns highly discriminative representations without reliance on domain-specific prior knowledge. Additional evaluation on an external imbalanced dataset validated the robustness and strong generalization capability of the proposed framework. Overall, BiToxNet provides a powerful and generalizable computational framework for the accurate identification of neurotoxic peptides and proteins. By effectively integrating evolutionary and biochemical information through bilinear attention, BiToxNet offers a valuable tool for neurotoxin screening and protein drug safety assessment, and presents a distinctive modeling strategy applicable to a wide range of biological sequence analysis tasks.

Abstract In the digital information age, artificial intelligence is increasingly being applied to national governance and judicial decision-making assistance. Existing studies lack case studies and empirical analyses of the effectiveness of large models in aiding judicial decisions. To address this research gap, this study designs a comprehensive evaluation framework encompassing five core task dimensions: Task-oriented Information Extraction, Legal Article Citation, Event Extraction, Judicial Decision Generation, and Legal Opinion Generation. By using carefully crafted prompts to activate the legal reasoning capabilities of the models, we conducted extensive testing on 13 mainstream large language models (LLMs). The experimental results demonstrate that large models perform excellently in processing legal texts and providing preliminary legal opinions, but still exhibit shortcomings in complex legal reasoning and precise decision-making. On this basis, we applied a weakly supervised learning strategy to fine-tune the LLMs for targeted improvements. The results indicate that introducing a small amount of task-specific learning can significantly enhance the performance of LLMs in judicial tasks. This further underscores the critical role of data and the acquisition of domain-specific knowledge in applying AI technology to judicial tasks. Additionally, this study briefly discusses the issue of the boundaries of AI’s involvement in judicial activities, aiming to provide theoretical foundations and practical guidance for the deep integration of AI technology with legal practice.

The spread of negative, engagement-driven content online causes significant societal harm, requiring advanced automated moderation tools. However, current classification systems often treat harmful content subtypes as independent, "flat" categories, which hinders their ability to thematically overlap content. This study designed and validated a novel integrated framework to accurately and transparently classify such complex cases. We proposed KG-DToT-HTC, a hybrid framework that synergistically combines three methodologies: a predefined Hierarchical Text Classification (HTC) taxonomy to structure the decision-making process; a domain-specific Knowledge Graph (KG) to provide factual, real-world context; and Decision Tree-of-Thought (DToT) prompting to guide a Large Language Model through an explicit, step-by-step reasoning process. On a real-world dataset of harmful Indonesian news, the proposed framework achieved a state-of-the-art Macro-F1 score of 0.934, representing a nearly 15-percentage point improvement over a zero-shot baseline. Ablation studies confirmed that each component—hierarchy, knowledge, and reasoning—provided a distinct and critical contribution to the final performance. The major conclusion of this study is that a synergistic architecture is essential for the accurate classification of complex harmful content. This work demonstrates a viable path toward "glass-box," interpretable AI moderation systems whose decisions are not only highly accurate but also fully auditable.

Amid the accelerating wave of digital transformation sweeping the global energy sector, laboratories in refining and chemical enterprises are undergoing a paradigm shift—from conventional manual operations toward intelligent, automated management systems. This study systematically investigates the construction pathways and technical architectures underlying intelligent laboratory systems within the petrochemical industry, with particular emphasis on how frontier technologies—including large-scale AI models, robotics, and the Internet of Things—are reshaping operational modalities. By integrating real-time data acquisition, mobile applications, intelligent analytics, and automated instrumentation, the proposed intelligent laboratory framework yields substantial operational improvements: a 40% increase in detection efficiency, 99.8% data accuracy, and an 80% acceleration in decision response time. Through in-depth analysis of benchmark implementations at Tianjin Petrochemical, Guangdong Petrochemical, and Fushun Petrochemical, this paper identifies key technological breakthroughs—namely the cloud-edge-end collaborative architecture, multimodal data fusion, and domain-specific knowledge graphs for quality control. It further addresses organizational challenges encountered during implementation and outlines prospective development trajectories. The findings offer a systematic and replicable reference framework for the intelligent transformation of laboratories in the refining and chemical sector.

Physics-aware machine learning integrates domain-specific physical knowledge into machine learning models, leading to the development of physics-informed neural networks (PINNs). PINNs embed physical laws directly into the learning process, enabling interpretable and physically consistent solutions to complex problems. However, the practical use of PINNs presents challenges and their applications are complex. Therefore, in this paper, we demonstrate the implementation of PINNs for systems of ordinary differential equations (ODEs), an area that is often overlooked by the physics community, which typically focuses on partial differential equations. We discuss two key challenges: the inverse problem, which involves estimating unknown parameters of ODEs, and the forward problem, which provides an approximate solution to ODEs. To provide practical insights into PINNs, we present two case studies based on a Python implementation using DeepXDE. Drawing on these studies, we discuss key challenges and identify promising directions for future research in PINN-based implementation frameworks.

Small Language Models (SLMs), typically ranging from hundreds of millions to several billion parameters, emerging as transformative tools in educational settings. Unlike their larger counterparts, SLMs offer distinct advantages including enhanced privacy preservation, reduced computational requirements, and cost-effective deployment on consumer-grade hardware. This paper examines the current landscape of SLM applications across diverse educational domains including health and medical education, programming education, mathematics education, science education, language instruction, and financial literacy. Drawing from recent research and implementations, we analyze the technical approaches employed, key advantages realized, and challenges encountered in deploying SLMs for educational purposes. Our analysis reveals that when properly fine-tuned and augmented with domain-specific knowledge through techniques such as Retrieval-Augmented Generation (RAG), SLMs can achieve performance comparable to large language models while maintaining significantly lower resource requirements. We identify critical future directions including the need for standardized evaluation frameworks, improved reasoning capabilities, and scalable infrastructure solutions. This paper contributes to the growing discourse on democratizing AI in education by highlighting how SLMs can provide accessible, privacy-preserving, and pedagogically effective educational support on a scale.

Hazard recognition on construction sites is crucial for ensuring worker safety. Traditional methods widely rely on expert assessments, on-site inspections, and checklists, which can be time-consuming and susceptible to human error. The integration of multimodal Large Language Models (LLMs), such as GPT-based systems, offers a promising opportunity to overcome these limitations. Therefore, this study evaluates the effectiveness of GPT-4o in recognizing workplace hazards from image inputs, with a specific focus on construction sites. The results indicate that the model can serve as a valuable decision-support tool for safety professionals by providing scalable and real-time insights. However, the study also highlights key limitations, including the model’s reliance on general visual features rather than domain-specific safety knowledge, and the continued need for human supervision. Additionally, ethical concerns, including bias in AI-generated hazard assessments, data privacy, and the risk of over-reliance on AI, must be carefully managed to ensure these tools contribute responsibly and effectively to proactive risk management strategies.

Named Entity Recognition (NER) in the rare disease domain poses unique challenges due to limited labeled data, semantic ambiguity between entity types, and long-tail distributions. In this study, we evaluate the capabilities of GPT-4o for rare disease NER under low-resource settings, using a range of prompt-based strategies including zero-shot prompting, few-shot in-context learning, retrieval-augmented generation (RAG), and task-level fine-tuning. We design a structured prompting framework that encodes domain-specific knowledge and disambiguation rules for four entity types. We further introduce two semantically guided few-shot example selection methods to improve in-context performance while reducing labeling effort. Experiments on the RareDis Corpus show that GPT-4o achieves competitive or superior performance compared to BioClinicalBERT, with task-level fine-tuning yielding the strongest performance among the evaluated approaches and improving upon the previously reported BioClinicalBERT baseline. Cost-performance analysis reveals that few-shot prompting delivers high returns at low token budgets. RAG provides limited overall gains but can improve recall for challenging entity types, especially signs and symptoms. An error taxonomy highlights common failure modes such as boundary drift and type confusion, suggesting opportunities for post-processing and hybrid refinement. Our results demonstrate that prompt-optimized LLMs can serve as effective, scalable alternatives to traditional supervised models in biomedical NER, particularly in rare disease applications where annotated data is scarce.

Automated anomaly detection is vital to industrial quality control, yet conventional deep learning detectors often struggle with scalability. These models, typically following a rigid "one-model-per-task" paradigm, require separate systems for each product line, increasing operational complexity and cost in diverse manufacturing environments. To address this limitation, we propose a unified defect detection framework based on a Multimodal Large Language Model (MLLM). Our approach utilizes a two-stage fine-tuning strategy: Supervised Fine-Tuning (SFT) to impart domain-specific knowledge, followed by a novel Reinforcement Fine-Tuning (RFT) process that refines visual reasoning. This RFT stage is guided by a multi-faceted verifiable reward function designed to optimize localization accuracy, classification correctness, and output structure. On a challenging real-world glove manufacturing dataset, our RFT-enhanced MLLM achieves a mean Average Precision (mAP) of 0.63, which is comparable to a highly specialized YOLO baseline (0.62). More importantly, a single, unified MLLM trained on a mixed-product dataset maintains competitive performance (mAP 0.61), demonstrating its ability to dynamically handle different products and defect types via natural language prompts. This study validates the feasibility of using a single, flexible MLLM to replace multiple rigid models in complex industrial inspection, offering a scalable and cost-effective paradigm for future intelligent quality control systems. The open-source code will be released at https://github.com/GloamXun/Glove-MLLM.

ABSTRACT This paper presents a knowledge graph construction method for spindle assembly in winding machines, aiming to address issues of dispersed knowledge and underutilization during the assembly process. To overcome the complexities of domain-specific knowledge and the challenges of relational triple extraction, the authors propose a comprehensive framework that includes knowledge modeling, knowledge extraction, and visualization. First, an ontology library tailored to the spindle assembly domain is developed, defining relevant entity types and relationship types. Then, the authors propose an enhanced extraction model, DGAD, which utilizes dual-gated dynamic convolution and multi-head attention mechanisms to automate the extraction of entities and relationships, effectively integrating local context and global features. The extracted triples are visualized using the Neo4j database, helping users intuitively understand the relationships between entities in the assembly process. Experimental results show that the proposed model achieves F1 score improvements of 5.83% and 7.03% in named entity recognition and relationship extraction tasks, outperforming baseline methods. The visualization of the knowledge graph provides a solid foundation for downstream applications, such as intelligent question-answering systems and fault diagnosis.

To estimate the influence of various loss functions on the performance of deep learning (DL) models for dose prediction in intensity-modulated radiotherapy (IMRT) for prostate cancer. A retrospective dataset comprising 110 prostate cancer patients was utilized. DL model was trained using various loss functions: mean absolute error (MAE), mean squared error (MSE), and combinations of MAE with predefined domain-specific knowledge, including dose-volume histogram (DVH) loss and moment loss function. The planned target volume (PTV) and dosimetric metrics for organs at risk (OARs) were used to assess the model's performance. A one-way analysis of variance (ANOVA) was applied to perform statistical comparisons between the clinical and predicted plans. In terms of dose deviations for OARs and PTV, the model trained with MAE plus moment loss performed better than models trained with MAE + DVH loss, MSE, or MAE. The MAE ± standard deviation (SD) between clinical and predicted dose distributions in the test cohort were (1.76 ± 0.5) Gy, (1.78 ± 0.5) Gy, (1.93 ± 0.6) Gy, and (2.02 ± 0.4) Gy for MAE + moment, MAE + DVH, MSE, and MAE models, respectively. Compared to the ground truth plans, the accuracy of all predicted plans was clinically acceptable. This study highlights how important loss function choice is to the optimization of DL-based prostate IMRT dose prediction models. The performance of the model is greatly improved by incorporating domain-specific knowledge into the loss function, which supports the possible practical application of such models for more precise and personalized radiation planning.

Reinforcement learning with formation energy feedback for material diffusion models.

Low-Dose Computed Tomography (LDCT) reduces radiation risk but introduces high noise levels that compromises diagnostic quality. To address this, we propose a Hybrid Generalized Efficient Layer Aggregation Network-UNet (GELAN-UNet) model, which incorporates medical priors into a progressive modular architecture. This design uses medically enhanced modules in shallower layers to capture fine details and computationally efficient blocks in deeper layers to reduce cost. Key innovations include a novel low-frequency retention path and an edge-aware attention mechanism, both crucial for preserving critical diagnostic structures. Evaluated on the public Mayo Clinic dataset, the proposed method achieves a superior peak signal-to-noise ratio (PSNR) of 45.28 dB - a 12.45% improvement over the original LDCT - while maintaining an optimal balance between denoising performance and computational efficiency. The critical importance of the low-frequency path, as revealed by ablation studies, validates the rationality of the hybrid strategy, which is further supported by comparisons with full medical and frequency-aware variants. This work delivers a high-performance denoising model alongside a practical, efficient architectural paradigm - rigorously validated through systematic exploration - for integrating domain-specific medical knowledge into deep learning frameworks.

Text-to-image person retrieval (TIPR) aims to identify the target person using textual descriptions, facing challenge in modality heterogeneity. Prior works have attempted to address it by developing cross-modal global or local alignment strategies. However, global methods typically overlook fine-grained cross-modal differences, whereas local methods require prior information to explore explicit part alignments. Additionally, current methods are English-centric, restricting their application in multilingual contexts. To alleviate these issues, we pioneer a multilingual TIPR task by developing a multilingual TIPR benchmark, for which we leverage large language models for initial translations and refine them by integrating domain-specific knowledge. Correspondingly, we propose Bi-IRRA: a Bidirectional Implicit Relation Reasoning and Aligning framework to learn alignment across languages and modalities. Within Bi-IRRA, a bidirectional implicit relation reasoning module enables bidirectional prediction of masked image and text, implicitly enhancing the modeling of local relations across languages and modalities, a multi-dimensional global alignment module is integrated to bridge the modality heterogeneity. The proposed method achieves new state-of-the-art results on all multilingual TIPR datasets.