Search Results (19,475)

Retrieval-augmented generation improves the factual consistency, knowledge timeliness, and scenario adaptability of large model inference services by incorporating external knowledge. However, it also introduces structural privacy risks, including private-knowledge leakage, prompt injection, and progressive information extraction in multi-turn interactions. To address these issues, this paper proposes Private-RAG, a privacy-preserving retrieval-augmented generation method for large model inference. The method constructs a composite threat model and a quantitative evaluation framework for the RAG pipeline, and further develops a layered collaborative defense mechanism consisting of controlled retrieval, sensitivity-aware context minimization, structured prompt isolation, and multi-criterion output gating. In addition, a risk feedback-driven budget accounting method is introduced to enable dynamic risk control in multi-turn interaction scenarios. Experimental results show that Private-RAG effectively reduces private-knowledge leakage, improves robustness against prompt injection, and suppresses cumulative privacy exposure while maintaining question-answering utility and a controllable deployment latency (e.g., 1165 ms), demonstrating superior privacy protection and inference robustness. Full article

Accurate remaining useful life (RUL) prediction of rolling bearings was essential for condition-based maintenance because bearing service degradation was primarily governed by progressive rolling-contact fatigue at the rollingelement–raceway interface, whereas vibration signals provided measurable responses to this degradation rather than being its physical cause. However, reliable RUL prediction remained challenging because vibration measurements were noisy, nonlinear, stage-dependent, and sensitive to operating-condition shifts. In this study, a health-indicator-guided temporal-attention framework was developed for bearing RUL prediction using public run-to-failure vibration datasets. The novelty of this work lay in integrating degradation-consistent health indicator construction, sliding-window life-cycle representation, and HI-guided temporal attention into a unified and interpretable prediction framework. First, degradation-sensitive vibration features were extracted and fused into a compact health indicator (HI) to represent the progressive deterioration trend. Then, sliding-window sequences were generated and processed by a Transformer-based temporal-attention network, through which long-range temporal dependencies were captured and higher weights were assigned to informative degradation segments near stage transitions and late-life acceleration. Experiments on the XJTU-SY and IMS datasets showed that the proposed method improved prediction stability, reduced late-life error amplification, and achieved better performance than baseline variants without HI or temporal attention. Ablation analysis confirmed that HI construction mitigated cross-stage drift, whereas temporal attention enhanced transition sensitivity during accelerated degradation. Robustness and cross-domain tests further indicated that the method maintained acceptable degradation-following behavior under noise perturbations and operating-condition changes, although explicit domain-adaptation mechanisms were still required for strongly shifted target domains. Full article

Climate change and rapid urbanization increasingly threaten water security in large river basins, yet existing assessments often fail to capture the multi-scale interactions between hydroclimatic extremes and human activities. To address this gap, we developed an integrated framework combining risk assessment, multi-method driver diagnosis (Geodetector, Multi-Scale Geographically Weighted Regression (MGWR), and Structural Equation Modeling (SEM)), and Zoned Management. Using a landscape-derived Ecological Risk Index (ERI) as a proxy indicator of runoff and non-point source potential, based on established empirical linkages between landscape metrics and hydrological processes, we applied the framework to three major urban agglomerations in the Yangtze River Basin from 2000 to 2020. Our results reveal three distinct risk mechanisms: in the Chengdu–Chongqing area (CYUA), a 165.8% increase in impervious surfaces drives altered runoff; in the Middle Reaches (MRC), the q-value of the Standardized Precipitation Index (SPI) rose from 0.017 in 2000 to 0.146 in 2020, corresponding to a 759% relative increase. Although the absolute q-value of SPI remains moderate at around 0.15, its rapid rise suggests increasing hydrological sensitivity of the MRC’s river–lake system to precipitation extremes; in the Yangtze River Delta (YRD), socioeconomic activities exert overriding pressure. Based on these diagnostics, we propose tailored strategies for water environment management, adaptive planning, and disaster mitigation. This framework offers a scientific basis for differentiated water governance in large river basins facing coupled anthropogenic and hydroclimatic pressures. Full article

Post-rotation jacking is a critical construction stage for load-path reconstruction and alignment adjustment in rotation-constructed bridges, particularly for ultra-wide double-deck composite girder systems. Taking a two-span continuous steel truss–concrete composite girder bridge with spans of 2 × 85 m as the engineering background, this study investigates the mechanical behavior during post-rotation jacking through theoretical derivation, finite element simulation, and on-site monitoring. Based on the force method of structural mechanics, a linear relationship between vertical synchronous jacking force and displacement is derived, and an analytical formulation for bearing reaction redistribution under laterally asynchronous jacking is established by considering the coupling effects of vertical bending, torsion, and transverse multi-bearing support. A full-bridge spatial finite element model was developed in MIDAS Civil NX 2024 V1.1 to analyze the redistribution of bearing reactions and the stress response of the concrete crossbeam under different jacking conditions. The results show that, for the investigated bridge, the jacking force–displacement response remains highly linear during synchronous jacking. The B-axis middle bearing is more sensitive to jacking displacement than the two side bearings, with its fitted stiffness being approximately 2.19 times the average stiffness of the side bearings. Eccentric jacking causes reaction concentration at the jacked point and reaction reduction at adjacent supports, and the magnitude of reaction variation increases approximately linearly with jacking displacement. When the transverse non-uniform jacking magnitude reaches 20 mm, a tensile stress of 0.3 MPa appears at the bottom flange of the concrete crossbeam; therefore, a project-specific stroke-difference limit of 20 mm is recommended for this bridge, while the actual construction achieved a stroke control accuracy of ±0.5 mm and a transverse elevation difference within 1 mm. Field monitoring results validate the proposed analytical and numerical methods. The Pearson correlation coefficients of the measured jacking forces with the finite element and theoretical results are 0.9987 and 0.9988, respectively, and the corresponding mean relative errors are 3.84% and 4.23%. For stress responses, the measured and calculated values show a strong correlation, with a Pearson correlation coefficient of 0.9980 and a mean relative error of 12.77%; the critical mid-span monitoring point shows a relative error of only 0.65%. The final bridge alignment deviation is controlled within ±3 cm. The overall mean verification coefficient is 0.968, with a 95% empirical agreement range of [0.888, 1.048], indicating that the proposed mechanical analysis framework and combined force–displacement control strategy can provide a useful reference for refined construction control of similar ultra-wide double-deck composite girder bridges with comparable span arrangement and transverse bearing layout. Full article

This study aimed to characterize the spatial patterns of soil nutrients and labile carbon in the 0–20 cm surface layer of a red-soil bench-terrace hillslope during the first year following cut-and-fill engineering. Soil nutrient redistribution is classically conceptualized as upslope depletion and downslope enrichment, yet whether this paradigm holds after bench terracing remains poorly documente d. On a granite-derived red-soil hillslope in Yudu County, Jiangxi Province, China, we established three replicated transects across four slope positions in May 2025, one year after cut-and-fill bench terracing combined with Camellia oleifera–Pinus massoniana mixed young-forest restoration. The 0–20 cm surface layer was sampled for pH, organic matter, total nitrogen, total phosphorus, water-soluble organic carbon, particulate organic carbon (POC), and mechanical composition. The results showed that organic matter, total nitrogen, and POC all peaked on the upper slope, with enrichment factors of 8.8×, 3.8×, and 5.1× relative to the hilltop, respectively; the slope base did not function as a nutrient sink. Texture displayed a monotonic downslope differentiation but decoupled from the nutrient gradient, and pH was significantly negatively correlated with organic matter and POC. The observed short-term post-restoration non-classical pattern is best interpreted as the spatially heterogeneous footprint of subsurface exposure and localized topsoil redistribution during cut-and-fill engineering, overlain by one year of incipient biological input, rather than the product of modified erosion–deposition dynamics. POC appears to be a particularly sensitive tracer of early biological activity under these short-term post-restoration conditions when superimposed on a depleted inverted-surface baseline, and the pronounced spatial heterogeneity implies that precision management based on high-resolution spatial diagnosis is warranted to address the substrate patchiness inherited from cut-and-fill operations. However, the temporal scope of this one-year baseline survey limits the inference of long-term indicator performance, and follow-up monitoring is needed to confirm whether POC retains this sensitivity as the surface layer matures. Full article

Background: Ibandronate is a nitrogen-containing bisphosphonate used in osteoporosis; however, its relationship with vertebral-fracture-related outcomes, pain trajectories, and broader inflammatory–skeletal signaling remains incompletely characterized. Methods: This retrospective observational study included patients with osteoporosis categorized according to ibandronate exposure. The primary outcome was new vertebral fracture occurrence, and the secondary outcome was change in pain severity assessed using the Visual Analog Scale (VAS). Multivariable regression, sensitivity analyses, and exploratory network-pharmacology, transcriptomic, and molecular docking analyses were performed. Results: Forty patients (20 ibandronate, 20 control) were included. Ibandronate use was associated with numerically lower vertebral fracture occurrence, although this did not reach statistical significance in crude or adjusted analyses. Greater pain reduction was observed in unadjusted analyses but was attenuated after multivariable adjustment, and baseline heterogeneity should be considered when interpreting between-group differences. Radiological outcomes did not differ significantly between groups. Exploratory systems-level analyses identified enrichment patterns involving inflammatory signaling, osteoclast differentiation, cytokine-associated pathways, and skeletal regulatory processes; however, these findings should be interpreted as hypothesis-generating and not as evidence of causal biological mechanisms. Conclusions: In this exploratory, hypothesis-generating study, ibandronate use was associated with trends toward lower vertebral fracture occurrence and greater unadjusted pain improvement, although these findings were attenuated after adjustment. The combined clinical, transcriptomic, and computational observations are compatible with the possibility that inflammatory and skeletal regulatory pathways may intersect within a broader systems-level framework relevant to vertebral-fracture-related outcomes in osteoporosis. However, these findings should not be interpreted as direct mechanistic evidence of ibandronate-specific molecular activity or clinical efficacy. Larger prospective studies integrating clinical, radiological, and mechanistic approaches are required to clarify the biological and clinical relevance of these observations. Full article

Early life stress (ELS) and adverse childhood experiences are critical determinants of neurodevelopmental trajectories and long-term somatic and psychiatric health outcomes. This narrative review synthesizes current evidence, identified through searches in PubMed, Scopus, and Web of Science, on the neurobiological and epigenetic mechanisms through which early environmental exposures shape developmental programming and stress responsivity across the lifespan. A central framework is the dysregulation of the hypothalamic–pituitary–adrenal (HPA) axis, which mediates adaptive and maladaptive stress responses. During sensitive developmental periods, including prenatal, perinatal, and early postnatal stages, increased neuroplasticity confers heightened vulnerability to environmental influences, resulting in persistent alterations in stress regulation systems, brain circuitry, and endocrine function. The review further examines the role of maternal stress during gestation, with emphasis on placental regulatory mechanisms and fetal programming processes that establish long-term physiological set points. In parallel, emerging evidence on paternal stress is considered, highlighting potential contributions of germline epigenetic modifications and postnatal environmental transmission pathways. At the molecular level, epigenetic mechanisms—including DNA methylation, histone modifications, and non-coding RNA regulation—are discussed as key mediators linking early environmental exposures to stable changes in gene expression without alterations in DNA sequence. Collectively, the evidence supports ELS as a fundamental biological embedding process with enduring consequences for health across the lifespan. A deeper understanding of these mechanisms, alongside the identification of reliable biomarkers, is essential for early detection and the development of targeted preventive and intervention strategies in pediatric populations. Full article

Background: Therapeutic resistance following cyclin-dependent kinase 4/6 inhibitor (CDK4/6i) plus endocrine therapy (ET) represents a key unmet need in hormone receptor-positive, human epidermal growth factor receptor 2-negative (HR+/HER2−) metastatic breast cancer (mBC). Treatment paradigms have advanced from non-targeted options, such as fulvestrant monotherapy or everolimus-based combinations, to precision medicine strategies, including inhibitors of the PI3K/AKT pathway, oral selective estrogen receptor degraders (SERDs), and novel ER-modulating agents, often guided by biomarkers and molecular surveillance. Methods: This narrative review synthesizes evidence from randomized clinical trials, real-world studies, and biomarker-driven analyses published from 2010 to 2026, with emphasis on next-generation sequencing (NGS)-guided genomic profiling, targeted pathway therapies, and circulating tumor DNA (ctDNA)-based proactive interventions in the post-CDK4/6i setting. This review was conducted and reported in accordance with the SANRA recommendations for narrative reviews. Results: Early second-line standards, including fulvestrant and alpelisib for PIK3CA-mutated tumors, established the basis for biomarker-guided treatment in hormone receptor–positive, HER2-negative metastatic breast cancer. With the widespread use of CDK4/6 inhibitors in the first-line setting, the optimal post-progression strategy has shifted toward molecularly selected combination approaches rather than single-agent endocrine therapy, as endocrine monotherapy has shown limited efficacy in acquired resistance. Multiple randomized studies have demonstrated that adding targeted agents to endocrine therapy improves progression-free survival compared with hormonal therapy alone, supporting combination regimens as the preferred strategy after CDK4/6 inhibitor progression, except in carefully selected patients with low disease burden, indolent biology, or frailty where tolerability is a major concern. Precision-based trials have further refined this approach. Elacestrant improved progression-free survival in ESR1-mutated disease in the EMERALD trial, capivasertib plus fulvestrant demonstrated significant benefit in tumors harboring AKT/PIK3CA/PTEN pathway alterations in CAPItello-291, and inavolisib plus palbociclib and fulvestrant achieved both progression-free and overall survival improvement in PIK3CA-mutated patients with early relapse in INAVO120. Real-world analyses further support the effectiveness of these biomarker-directed strategies across diverse clinical subgroups. Comprehensive genomic profiling has identified multiple resistance mechanisms, including ESR1 mutations, PI3K/AKT/mTOR pathway activation, RB1 loss, and FGFR alterations, which may co-occur and reduce sensitivity to endocrine monotherapy. While ESR1 and PI3K pathway alterations now guide approved therapies, FGFR alterations remain investigational targets, with ongoing trials evaluating selective FGFR inhibitors. Proactive switching approaches evaluated in SERENA-6 and PADA-1 demonstrate that serial circulating tumor DNA (ctDNA) monitoring can detect emergent ESR1 mutations before radiographic progression, providing a clinically actionable lead time for early therapeutic modification and extending endocrine-based disease control by approximately 5 to 7 months. Conclusions: Post-CDK4/6i management increasingly relies on NGS-guided precision approaches, integrating pathway-specific therapies and ctDNA surveillance to tailor sequencing based on resistance profiles, prior ET response, and tumor heterogeneity. Future investigations into novel ER degraders and multi-targeted combinations hold potential to further optimize algorithms, extend non-chemotherapy options, and enhance survival in HR+/HER2− mBC. Full article

Dietary emulsifiers, common in processed and ultra-processed foods, improve food texture and shelf life but may affect gut health by interacting with the microbiota and intestinal barrier. While emulsifiers have long been considered safe, growing evidence links their presence in ultra-processed foods to chronic disease risk. This review aims to evaluate the current understanding of the factors and mechanisms underlying individual differences in intestinal mucosal susceptibility to dietary emulsifiers. A search of PubMed and Embase through February 2026 identified eight relevant studies. Overall, the available evidence indicates a heterogeneous and highly individualized host response to dietary emulsifiers. These differences appear to be strongly influenced by the gut microbiota and its functional properties, while animal studies further suggest that host factors such as sex-related differences in microbial composition may also contribute to variability in response. Importantly, not all emulsifiers have the same effects, underscoring compound-specific impacts on gut physiology. The findings demonstrate that sensitivity to dietary emulsifiers varies substantially between individuals, challenging the long-standing assumption that these additives are universally safe. Given the multifactorial nature of this susceptibility, particularly the role of the gut microbiota, future research should adopt an integrative approach that combines microbial profiling with host genetics, immune responses, and early-life exposures. Such efforts will be essential to identify at-risk individuals and to inform more personalized dietary recommendations aimed at preserving intestinal health and reducing disease risk. Importantly, there is a clear need for larger, well-powered studies that can validate and expand upon these initial observations. Full article

Laser hard overlaying is an advanced, perspective technology with wide industrial applications, for example, dies. The aim is to improve surface properties like wear resistance using special layers of powder sintered or remelted by laser beam. At present, dies are manufactured by machining with following bulk heat treatment, which is an expensive process particularly due to use of expensive high-alloyed tool steels. Repairs performed using arc or plasma welding introduce a big amount of heat to the part, which can cause dimension changes and material degradation. These methods often fail also due to low weldability of the materials. An advantage of laser overlaying is minimization of these difficulties. The paper contains a comprehensive evaluation of several types of hard overlayed powder of H13 tool steel on a S355 structural steel and on H11 tool steel using a laser beam. Macro- and microstructure, hardness and fatigue resistance are evaluated, including fatigue damage mechanisms. In the case of the H13 welds on the S355 steel plate, the quality of the welds was mostly acceptable, without pores or segregate impurities and with a good interconnection between the weld track and base material. Results are completed with basic measurement of residual stresses using destructive strain-gauge methods. High tensile residual stresses of 1465 MPa were measured at the boundary of the first track of the single-layer overlay. Fatigue resistance is sensitive on surface and subsurface defects, which can significantly reduce endurance limit. Fatigue strength of specimens with the single layer overlay was considerably lower than fatigue strength of the S355 steel. The decrease was between 25% and 50%. In the case of overlay of H13 on H11 tool steel, the decrease in fatigue strength was between 25% and 30%. Full article

The study examines discrepancies between personally reported and declared use of generative artificial intelligence (GenAI) among university students from a public university located in the Peruvian Andes, operationalized as the AI Use Gap, an exploratory discrepancy indicator based on two self-reported measures. Drawing on a sequential explanatory mixed-methods design, the study combines survey data (N = 150), experimental vignette evaluations, and qualitative follow-up interviews to explore how students manage the visibility and disclosure of AI use in academic contexts. Findings indicate relatively high levels of AI use alongside a consistent discrepancy between personally reported and declared use, suggesting patterns of differential reporting across contexts. Quantitative analyses did not show clearly differentiated exploratory relational patterns between the AI Use Gap and the psychosocial/contextual indicators examined, including perceived stigma, concealment, normative ambiguity, and peer pressure. Given the exploratory nature and limited internal consistency of the contextual indicators, these findings were interpreted cautiously as provisional exploratory patterns rather than as evidence of stable psychosocial relationships. Qualitative findings suggest that AI disclosure practices are shaped by socially evaluative and context-dependent processes, including fear of judgment, uncertainty regarding acceptable AI use, and selective disclosure strategies. Participants frequently described AI use as widespread but not consistently disclosed across academic settings. Overall, the findings suggest that discrepancies between AI use and disclosure may be better understood as forms of visibility management shaped by institutional ambiguity and social expectations rather than by stable individual-level characteristics alone. Rather than validating stable psychosocial mechanisms, the study explores an emerging and context-sensitive phenomenon using provisional contextual indicators intended to capture heterogeneous patterns of perception and disclosure. The study contributes to ongoing discussions regarding transparency, academic integrity, and the social regulation of AI use in higher education, particularly in under-researched Global South contexts. Full article

A series of substituted pyrrolo[2,3-d]pyrimidines was designed, synthesized, and evaluated in vitro against two protozoan parasites: Plasmodium falciparum and Trypanosoma brucei brucei. Pharmacological studies revealed antiprotozoal activity with IC₅₀ values in the submicromolar to micromolar range. Additionally, the in vitro cytotoxicity of these new compounds was assessed using human HepG2 cells. Among them, the pyrrolopyrimidine derivative 1d emerged as the most potent antimalarial compound, exhibiting a selectivity index (SI) of 600.81 against the P. falciparum chloroquine-resistant W2 strain. For the chloroquine-sensitive 3D7 strain, the most notable selectivity index (SI) was observed for pyrrolo[2,3-d]pyrimidine 1c, with a value of approximately 123. Furthermore, compound 1b demonstrated the most interesting activity against Trypanosoma brucei brucei, with an SI of 39.52, marking it as a promising trypanocidal agent. FRET melting assays confirmed that these nitrogen-containing heterocyclic compounds bind to telomeric G-quadruplexes in P. falciparum and Trypanosoma. However, no clear correlation was found between G-quadruplex binding and antiparasitic activity or selectivity, suggesting that G-quadruplex targeting is unlikely to be the main mechanism underlying cytotoxicity. Full article

Mountain ecosystems are sensitive response units and critical ecological barriers to global climate change. Located in the mid-latitude climate transition zone, these ecosystems feature high ecological sensitivity and complex driving mechanisms, creating an urgent need to conduct long-sequence, high-precision dynamic assessments in order to support ecological conservation and climate adaptation decision-making. However, three key research gaps remain in the field: first, traditional assessments are dominated by static observation, lacking the capacity for long-sequence dynamic analysis and future projection; second, the coupled interaction mechanism among multiple ecological factors remains unclear, with insufficient quantitative and physical mechanism characterization; third, existing ecological zoning has not been validated for robustness, rendering it incapable of addressing climate disturbances and extreme scenarios. In order to study the regional atmospheric ecosystem, this study takes Shangluo in the eastern Qinling Mountains as the study area and constructs an integrated assessment framework integrating multi-dimensional diagnosis, simulation and projection, dynamic zoning and robustness validation based on long-sequence multi-factor data covering the years 1965–2024. The study aims to reveal the long-sequence evolution patterns and four-dimensional coupling mechanism of the Qinling Mountains atmospheric ecosystem, developing a reproducible and transferable dynamic assessment model. The results show that the study area exhibits the characteristic of elevation-dependent warming, and the correlation coefficients between elevation and air temperature, and between vegetation coverage and air quality reach −0.89 and −0.76, respectively.; ecological quality presents a spatial pattern of being high in the southwest and low in the northeast, with a coefficient of variation across the whole study area lower than 0.03. The results of 1000 Monte Carlo random disturbance validation runs show that even under intensified climate stress, the zoning pattern still maintains extremely strong disturbance resistance. This study reveals the steady-state multi-factor interaction mechanism in mountainous regions, addressing the defects of traditional static assessments that ignore ecosystem evolution and lag effects. The dynamic projection model constructed in this study can be transferred to similar mid-latitude mountainous regions worldwide, providing theoretical and technical support for regional ecological governance. Full article

Roof instability in the heading area of fully mechanized excavation roadways, together with insufficient coordinated operation between excavation and support, severely restricts tunneling safety and construction efficiency. A novel track-mounted advanced support equipment structure with an articulated curved roof beam is proposed in this study. Considering actual underground working conditions, including uneven roof contact, eccentric loading and local support failure, a three-degree-of-freedom dynamic model covering vertical, pitch and roll motions is established based on Lagrange’s equations. Dynamic characteristics under varying load amplitudes, excitation frequencies, static load offsets and typical support failure modes are systematically analyzed. The results reveal that only vertical vibration emerges under the full support condition, and the resonance frequency of the system is approximately 10 Hz. The maximum steady-state vertical displacement reaches 0.6406 mm with an RMS of 0.5472 mm under an intact support state. The pitch vibration amplitude caused by the failure of the first support group is three times that of the second group, proving front supports dominate anti-overturning capacity. Side beam failure triggers remarkable roll-coupled vibration, while middle beam failure mainly enlarges vertical displacement. This paper clarifies the vertical–pitch–roll coupling vibration mechanism induced by local support failure. Parameter sensitivity analysis reveals that static load offset has the highest sensitivity, while excitation frequency (within 4–6 Hz) and damping ratio exhibit negligible influence on the steady-state response. The obtained quantitative results can provide a reliable theoretical reference for structural optimization, stability regulation and safety monitoring of track-mounted advanced support facilities. Full article

Accurate surface defect classification is a critical requirement in structural health monitoring and infrastructure inspection, where defects, including cracks, spalling, delamination and noncrack regions, often appear with low-contrast and complex background textures. Motivated by the need for a robust and discriminative framework that can enhance defect visibility and focus learning on damage-critical regions, this research proposes a novel damage-aware DenseNet-201 (DA-DenseNet-201) model for surface defect classification. As a critical novelty, a damage-aware adaptive contrast-limited adaptive histogram equalisation (DAC) filtering strategy is introduced as a preprocessing stage. The proposed DAC filter dynamically adjusts contrast enhancement parameters based on damage indicators, selectively amplifying crack edges and defect textures while preserving healthy surface regions and suppressing noise. Building on this method, enhanced images are processed using a pretrained DenseNet-201 backbone, retaining the benefits of dense feature propagation and efficient gradient flow. To strengthen the discriminative learning of DA-DenseNet-201 further, an attention refinement block is integrated into the network, combining channel attention to emphasise defect-relevant feature responses and spatial attention to localise damage regions accurately. In addition, a multiscale feature fusion mechanism aggregates feature maps from multiple dense blocks to capture fine-grained crack patterns, texture-level degradation and high-level semantic damage information. Extensive experiments conducted on surface defect datasets demonstrate its effectiveness, achieving a superior classification accuracy of 98.93%, along with notable improvements in sensitivity, specificity and the intersection over union compared with state-of-the-art models. These results confirm that the proposed DA-DenseNet-201 provides a reliable and high-performance solution for automated surface defect classification. Full article