Search Results (607)

Underground mine object detection is challenged by low illumination, blur, dust scattering, and repetitive tunnel clutter, which jointly corrupt backbone features, entangle DETR queries, and weaken localization for small objects. Existing enhancement-based and detector-internal methods do not explicitly propagate degradation reliability across features, decoder queries, and box refinement. We propose Spectral-DETR, a detector-internal reliability framework built on RF-DETR. Its central design is a cross-stage reliability pathway that connects Degradation-Aware Frequency Decomposition (DAFD), Degradation-Adaptive Query Contrastive Denoising (DQCD), and Salience-Calibrated Uncertainty with Learned Uncertainty Estimation (SCU+LUE). On Mine-Objects (14 classes, 3081 images), Spectral-DETR achieves an average precision of 0.917 at an intersection-over-union threshold of 0.5 and 0.493 when averaged over thresholds from 0.5 to 0.95, exceeding YOLOv9m by 1.6 and 0.8 percentage points, respectively, under the dataset-specific evaluation protocol. In controlled RF-DETR validation, the three reliability stages improve these two measures from 0.883 to 0.913 and from 0.472 to 0.486, respectively. Spectral-DETR obtains corresponding values of 0.848 and 0.571 on ExDark and 0.973 and 0.495 on ScienceDB. DQCD and SCU remain training-only losses with no inference cost. Full article

Public generative artificial intelligence (GenAI)–student datasets invite contested links between AI intensity, usage style, and grades, yet many analyses treat predictive accuracy or significant coefficients as sufficient evidence while skipping prior achievement, co-outcome leakage checks, and absolute effect-size thresholds. This paper presents a construct-audit protocol that treats associational claim survival as a reproducible labeling task: a feature-role taxonomy, forbidden-feature gates, nested out-of-fold change-in-$R^2$ materiality thresholds, and operational labels (stable, vanished, artifact-born—the last defined but not positively observed here), with predictive models used as instruments rather than as the scientific product. On the public ai_student_impact_dataset, treated as a construct-audit sandbox (possibly synthetic or engineered; no campus-population or causal claims), a five-seed Ridge-primary run is used to validate those rules rather than to estimate GenAI effects: all eight primary intensity and style claims are non-material under locked absolute gates (AI joint change-in-$R^2\approx 0.0064$ versus prior grade-point-average lift $\approx 0.859$), while a kitchen-sink OLS significance foil stars $12/19$ coefficients that the inventory does not promote. A report-only Random Forest check shows that style and joint-block clearance can depend on the modeling instrument; inventory labels remain Ridge-primary under the locked metric. The protocol can therefore withhold GenAI–GPA claims when absolute gates fail, and a six-step laptop workflow is specified so educational researchers can apply the same checks without reproducing the full validation schedule. Full article

Conventional diagnosis of chronic kidney disease (CKD) relies predominantly on invasive blood-based examinations, limiting the scalability of kidney health screening in resource-constrained environments. This study presents embedded engineering framework for non-invasive, breath-based CKD staging framework supported by machine learning and implemented on a low-cost embedded platform. To account for physiological sex differences in baseline creatinine production, estimated glomerular filtration rate (eGFR) values and breath ammonia concentrations were derived from two independent clinical cohorts using sex-specific MDRD equations (incorporating the standard male formula and the $0.742$ female correction factor, respectively) and creatinine–BUN conversion models, with male- and female-parameterized algorithms developed in parallel. The resulting feature space was analyzed using four unsupervised clustering approaches to stratify subjects into five clinically meaningful kidney function stages. Stage-specific ammonia thresholds were implemented within an Arduino Nano-based prototype equipped with an MQ-137 gas sensor and OLED display, enabling real-time point-of-care classification. Dataset-level classification accuracy reached 82% for the male algorithm and 92% for the female algorithm. Hospital-based validation on 29 patients (22 male, 7 female) yielded a real-world testing accuracy of 90.5% ($20/22$) for male patients and 71.4% ($5/7$) for female patients, a discrepancy largely attributable to the small female sample size. Because the current evaluation lacks healthy control subjects and is constrained by sample size, these empirical results serve primarily to demonstrate hardware-software functional integration and real-world deployment feasibility rather than definitive clinical efficacy. Despite these preliminary, sample-limited clinical datasets, results suggest this approach holds promise as an accessible, non-invasive screening complement to conventional diagnostic pathways, particularly in low-resource healthcare settings. Full article

Background: Portal vein tumor thrombus (PVTT) marks advanced hepatocellular carcinoma (HCC) and carries a dismal prognosis. Survival varies widely even within this stage, yet simple tools for individualized risk stratification remain scarce. Methods: We retrospectively enrolled 134 HCC patients with PVTT and randomly divided them into a training set (n = 94) and a validation set (n = 40). Clinical predictors were selected by variance inflation factor screening and backward elimination Cox regression. A radiomics score (Rad-score) was constructed from portal-venous phase computed tomography (CT) images using Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression with 10-fold cross-validation. Three Cox models were built: a clinical model, an imaging model based solely on the Rad-score, and a combined model integrating both. Discrimination was assessed by C-index and time-dependent area under the curve (AUC). Calibration was examined with bootstrap-based calibration curves. Decision curve analysis evaluated net benefit. A nomogram was developed from the combined model. Results: Four clinical variables (alpha-fetoprotein (AFP), body mass index (BMI), high-density lipoprotein cholesterol (HDL-C), and alkaline phosphatase (ALP)) and two CT texture features (GLRLM_SRHGE and GLZLM_SZHGE) were retained as independent predictors. The combined model gave the highest C-index in both the training set (0.843) and the internal validation set (0.815). Its 1-year AUC reached 0.953 and 0.947 in the two sets. Calibration slopes ranged from 1.044 to 1.291 across time points, indicating a tendency toward mild overdispersion; nevertheless, decision curve analysis confirmed net benefit across clinically relevant thresholds. The combined model offered greater net benefit than either single-domain model across a 0–50% threshold range. A nomogram incorporating all five predictors was generated for individualized 12- and 24-month survival prediction. Conclusions: A combined model integrating routine laboratory variables and a CT-based radiomics score improved survival prediction over clinical or imaging models alone. The corresponding nomogram uses inputs from a basic blood panel and a single portal-venous phase CT, suggesting its potential as a low-cost prognostic stratification tool for HCC patients with PVTT, although external validation in prospective multicenter cohorts is required before clinical implementation. Full article

Reliable battery prognostics for electric vertical take-off and landing (eVTOL) aircraft require models that preserve phase-dependent electrothermal information while generalizing to cells absent from training. This study reconstructs the public CMU eVTOL battery dataset and establishes a leakage-safe benchmark for state of charge (SOC), five-mission-ahead state of health (SOH), and threshold-based remaining useful life (RUL). A protocol-based screen identified 441 valid C/5 reference-performance-test anchors across 22 cells; three cells with non-monotone diagnostic-capacity trajectories were excluded from the primary health benchmark, leaving 19 cells. Predictors were restricted to telemetry-derived phase and mission features, with cell identity and target- or future-derived quantities excluded. All health models were evaluated using outer leave-one-cell-out validation with matched 20-mission histories. Random Forest achieved the lowest SOH MAE of 0.620 percentage points, compared with 1.252 for the mission-phase Transformer and 1.471 for Attention-LSTM-MoE. The best RUL MAEs were 24.067, 46.527, and 122.198 missions at the 90%, 85%, and 80% SOH thresholds. Cell-bootstrap uncertainty showed that model differences were threshold-dependent. Row-random splitting produced substantially more optimistic errors than cell-held-out evaluation. These results show that rigorous target construction and leakage-safe validation are critical and that increased sequence-model complexity does not guarantee superior unseen-cell generalization. Full article

Temporal action localization (TAL) requires a detector to recognize action categories and estimate temporal boundaries in untrimmed videos. Mamba supports linear-complexity long-sequence modeling, yet a uniform allocation of state-space operators does not explicitly differentiate the context requirements associated with temporal scales and prediction branches. We present Multi-Scale Decoupled Routing Mamba (MSDR-Mamba), a multi-scale branch-decoupled routing state-space detector. The method combines a phase-dilated multi-rate Mamba temporal pyramid, multi-band state-time initialization, level-wise local–global gating guided by duration priors, and a branch role-decoupled head. The final head uses CNNs for classification and center-offset estimation, with Mamba used for class-specific start/end boundary neighbor modeling. With frozen InternVideo2-6B features on THUMOS14, MSDR-Mamba achieves a five-threshold mAP of 73.09%, exceeding TriDet by 0.43 percentage points. Supplementary experiments on ActivityNet-1.3 and P²ANet further evaluate the complete configuration under longer-duration and dense short action distributions. The results support scale- and branch-aware state-space modeling as a practical design strategy for TAL. Full article

We set out to fix a “contamination” problem in reconstruction-based Temporal Convolutional Network VAEs (TCN-VAEs) for cross-dataset IoT flow anomaly detection: when attack flows share an encoder window with benign flows, the shared latent code is allegedly distorted, inflating benign reconstruction error and producing false positive rates (FPRs) of 22–65% despite an ROC-AUC above 0.93. Our proposed fix, TCN-Pred, excludes the target flow from the encoder and scores it by next-flow prediction error, reducing FPR to 0.65–13%. We subjected this causal explanation to a battery of controlled ablations, holding architecture, decoder, loss, and thresholding fixed while varying one factor at a time. Each one falsified the original hypothesis: target inclusion/masking changes FPR by at most 0.001; context shuffling/reversing/zeroing changes it by at most 0.003; a context-blind constant-output predictor matches TCN-Pred’s FPR and F1 to three decimal places on all three datasets. The actual cause, confirmed on the original trained models with no retraining, is a scoring-granularity mismatch: the TCN-VAE threshold is calibrated from per-window errors averaged over 20 flows but applied to per-flow errors at evaluation (standard deviation $\approx \sqrt{20}\times$ higher, measured ratio 4.46 against a predicted 4.47). Recalibrating the identical model at matching granularity drops FPR from 22.7/47.6/64.6% to 0.65/5.0/12.5% on BoT-IoT, IoT-23 and ToN-IoT, closing 89–97% of the reported FPR gap without changing a single model weight. We report this diagnostic chain, together with an attack-prevalence sensitivity analysis, sample-disjoint calibration, normality diagnostics, and label-free and redundancy-aware (mRMR) feature-selection benchmarks, as a methodology other work should apply before attributing fixed-threshold performance to architecture. The pipeline is supervised source-domain feature selection followed by benign-only detector training, not fully unsupervised, a distinction we quantify later in the paper. Investigating dataset representativeness, we found that all three provided files reduce to only ≈6000 genuinely distinct flows via an undocumented row-duplication procedure, causing 97.8% BoT-IoT train/test near-duplicate overlap; a leakage-free re-evaluation changes FPR by only 0.23 percentage points. We also found that the TLS-metadata columns are already transformed upstream of every available artefact, so the proportion of genuinely TLS-encrypted flows cannot be recovered, and we soften the paper’s encrypted-traffic framing accordingly. Full article

LiDAR-based 3D object detection for autonomous driving must balance localization accuracy with real-time inference, while sparse point measurements make small-scale objects such as pedestrians and cyclists particularly challenging to represent at long range. This paper presents PES-PointPillars, an enhanced PointPillars detector with three coordinated design changes. First, pinwheel-shaped convolution (PConv) replaces selected backbone convolutions to expand horizontal and vertical receptive fields for sparse structural patterns. Second, an Improved Inter-Layer Feature Correlation (I-EFC) module uses soft gating and adaptive thresholding to fuse multi-level features through continuous, input-dependent weights. Third, a Smooth L1-NWD (SNWD) loss applies normalized Wasserstein distance to planar position and scale while retaining Smooth L1 regression for vertical position, height, and orientation. Using the parameter settings and configuration of the original PointPillars implementation, the locally executed PES-PointPillars experiment achieves Moderate 3D average precision values of 77.1% for cars, 46.7% for pedestrians, and 62.9% for cyclists at 68.3 FPS on the KITTI validation split. Relative to the source-reported PointPillars reference, the corresponding numerical differences are 2.1, 3.2, and 3.8 percentage points. The reported component-wise and staged ablations show category-dependent gains, with the complete model providing the strongest aggregate performance among the evaluated configurations. Full article

Undiagnosed diabetes and prediabetes are common among trauma patients and increase perioperative complication rates, yet routine glycemic screening at hospital admission is rarely performed. We compared 10 supervised machine learning (ML) algorithms for early detection of impaired glucose metabolism in 1789 trauma surgery patients (1095 normal; 694 prediabetes/diabetes) using questionnaire data and preoperative laboratory values within a nested cross-validation framework with hyperparameter optimization. On the combined feature set, XGBoost achieved the highest area under the receiver operating characteristic curve (ROC-AUC; 0.815), followed by multilayer perceptron (0.814) and Gradient Boosting (0.813); adding blood parameters significantly improved performance for nine of ten algorithms (DeLong’s test, p < 0.05). SHapley Additive exPlanations (SHAP) and XGBoost feature gains identified six overlapping predictors, yielding a simplified model with comparable mean discrimination. Inter-model differences were small, and a fully interpretable logistic regression model remained within one percentage point of XGBoost. The best-performing model demonstrated strong calibration and a higher net benefit than either universal testing or no testing across the evaluated threshold probabilities. In routine trauma practice, such a model could identify patients who would benefit from confirmatory glycated hemoglobin (HbA1c) testing, enabling earlier detection of dysglycemia, risk-stratified glucose management, and timely referral for lifestyle intervention in patients with prediabetes. Full article

Emotion recognition in children’s drawings is difficult because affect is carried by sparse strokes, symbolic objects, and overall composition rather than by the stable appearance statistics of photographs. We built a reproducible four-class benchmark (Angry, Fear, Happy, Sad) on a single public corpus of 818 children’s drawings and compared three transfer-learning regimes under identical stratified five-fold splits with nested model selection: ResNet-50, ViT-B/16, and an end-to-end fine-tuned SigLIP image encoder (SigLIP-FT). SigLIP-FT reached the highest macro-F1 (0.773 ± 0.028), ahead of ViT-B/16 (0.700 ± 0.040) and ResNet-50 (0.598 ± 0.038), and was the best calibrated (ECE 0.119). Frozen-feature linear probes preserve this ordering (0.541, 0.648, 0.731), locating the advantage in the pretrained representations rather than in fine-tuning, while zero-shot SigLIP reaches only 0.510, so task-specific supervision remains necessary. Margin-based abstention raised retained-set macro-F1 to 0.810 at 78.0% coverage and 0.844 at 61.4% coverage—post hoc operating points computed on the pooled out-of-fold predictions; deployment thresholds must be fixed on independent data—and SigLIP-FT attains the lowest area under the risk–coverage curve (AURC 0.126 versus 0.199 and 0.278). Residual errors are highly structured: 87.0% lie within the negative-emotion triad, and Fear is the hardest category for all three architectures. All findings are established on this single corpus; their transfer to other collections is an open question. Coverage–performance behavior, rather than a single full-coverage score, is the appropriate reporting standard for ambiguous visual domains of this kind. Full article

Falls are a major cause of severe injury and mortality among older adults, requiring rapid state recognition and alerts in healthcare and caregiving environments. However, many existing fall-recognition approaches primarily focus on post-fall detection or assume access to cloud/GPU computing, leaving limited evidence for short-term pre-fall recognition on low-end CPU-based edge devices. This study proposes a lightweight CNN–TCN–Attention framework for recognizing fall and pre-fall states on resource-constrained edge devices. Using MediaPipe, three-dimensional coordinates and visibility scores of 33 human body landmarks are extracted from input videos. Raw RGB frames are not directly used by the classifier, thereby reducing the processing of visually identifiable information. A total of 2300 fall-related videos from the AI-Hub dataset were reorganized into three classes: normal, pre-fall, and fall. The pre-fall onset was defined as the point at which the downward vertical velocity of the nose landmark exceeded a dataset-specific threshold. The model uses a CNN to extract frame-level skeletal patterns, a temporal convolutional network (TCN) to learn temporal changes in posture, and an attention module to aggregate temporal features. It was deployed on a Raspberry Pi Zero 2 W using ONNX Runtime. The proposed model achieved 94.71% accuracy and a macro F1-score of 93.63%, with an average state-decision latency of 325.88 ms/decision. It improved accuracy by 2.83 percentage points over the lightweight CNN–LSTM–Attention baseline while maintaining comparable latency and achieved approximately 1.75× faster processing than the MobileNetV3–TCN–Attention model. Full article

Plant disease symptoms combine local texture changes with patterns distributed across a leaf, while practical recognition models must remain compact. We introduce DC-FEN, a MobileNetV3-based design that models spatial-token relations and channel interactions in parallel and injects them through gated residual fusion. We also examine output-distribution, direct-feature, and token-relation transfer under same-backbone and heterogeneous teachers. PlantVillage and Plant Pathology 2021 (FGVC8) are evaluated with duplicate-audited, group-aware 70/15/15 splits, an explicit unresolved-leaf sensitivity check, validation-only selection, five training seeds, class-sensitive metrics, and paired seed-wise descriptive summaries. On PlantVillage, the no-additional-attention student, DC-FEN teacher, and DC-FEN joint student obtain macro F1 scores of $96.46\pm 0.91\%$, $96.90\pm 0.40\%$, and $96.55\pm 0.25\%$. On FGVC8, the corresponding scores are $87.29\pm 0.63\%$, $87.14\pm 0.52\%$, and $87.20\pm 0.26\%$. At the prespecified FGVC8 threshold of 0.5, DCAB changed sample-wise F1 by $-0.02\pm 0.55$ percentage points relative to the unmodified backbone; validation-selected global and label-specific thresholds changed this contrast to $+0.28\pm 0.55$ and $+0.55\pm 0.29$ points, while threshold-free macro mAP remained essentially unchanged. A duplicate-audited PlantDoc pressure test reduced frozen-checkpoint accuracy to $30.34\pm 1.10\%$ and $29.57\pm 1.10\%$, showing that external generalization remains unestablished. A ResNet50 teacher gives logit-only students $97.42\pm 0.51\%$ macro F1 on PlantVillage and $89.82\pm 0.43\%$ sample-wise F1 on FGVC8. After separately weighting the direct and relation terms, the corresponding joint students obtain $97.37\pm 0.56\%$ and $89.94\pm 0.27\%$, recovering the degradation seen with unit internal weights while remaining close to logit-only transfer. Thus, the study evaluates the benefits and limits of explicit spatial–channel interaction and shows that adding intermediate transfer constraints does not guarantee a stronger student. Full article

Compressor surge poses a critical safety risk for aero-engines. However, conventional physics-driven detection methods—relying on single-domain features and fixed empirical thresholds—struggle to adapt across varying compressor configurations, wide operating ranges, and complex interference environments. This paper proposes a multi-feature fusion adaptive surge detection method that integrates time-domain amplitude, frequency-weighted power and slope features within a joint threshold criteria, enabling reliable and adaptive surge detection according to the statistical characteristics of the signal itself. A wavelet-based preprocessing strategy is established with the db4 wavelet and four-level decomposition identified as the optimal setting through systematic evaluation. A novel feature FWP is introduced herein, which applies frequency-dependent weighting to the power spectral density to suppress noise components while amplifying energy changes within surge-relevant bands, achieving 1.7 to 6.1 times greater magnitude variation near the surge point compared with total spectral power. The slope feature is further discovered to distinguish surge from transient interferences such as rapid valve throttling, fuel stepping and rapid acceleration. Among 100 samples, the three-feature joint detection strategy integrated with adaptive threshold criteria improves accuracy from 61% to 98%. A Bayesian optimization framework using Gaussian process surrogate models is developed for efficient cross-engine hyperparameter tuning, converging to optimal solutions within merely 11 to 13 iterations across two distinct compressors. Lastly, the method is implemented on an NI cRIO-based real-time platform and validated on two distinct ten-stage high-pressure compressors, covering surge tests across a wide speed range of 45% to 98%. Comparative tests against an industry-standard reference device demonstrate earlier warning lead times of 41 to 99 ms. The results confirm that the proposed method herein achieves high accuracy, strong robustness against operational interferences, and good cross-platform adaptability for practical application. Full article

Class-incremental three-dimensional point cloud semantic segmentation requires models to learn newly introduced categories while preserving previously acquired knowledge without storing historical point clouds. This setting is challenged by representation drift during incremental optimization and semantic background shift caused by incomplete annotations of previously learned categories. To address these problems, this study proposes confidence-gated relational distillation, an exemplar-free teacher–student framework that combines feature-level relation preservation with semantic-level background correction. The relational component transfers normalized neighborhood-affinity distributions and weights each point according to teacher reliability, thereby reducing the influence of uncertain predictions. The background-compensation component reconstructs reliable old-class targets using class-specific thresholds and calibrates competition between previously learned and newly introduced classes. Experiments on the Stanford Large-Scale Three-Dimensional Indoor Spaces dataset and ScanNet show competitive performance across multiple incremental settings, with more consistent improvements on ScanNet. Under the ScanNet 10-1 protocol, the proposed method achieves an average mean intersection over union of 44.6% across eleven learning states and 33.3% at the final state. Under the same PointNet++ configuration, it also reduces training time and peak graphics processing unit memory. These results indicate that reliable relational transfer and adaptive background correction provide an effective balance between old-class retention and novel-class acquisition without introducing replay data or separate architectural branches during incremental optimization. Full article

This paper proposes a novel Topological Data Analysis (TDA) pipeline to extract robust structural features from functional near-infrared spectroscopy (fNIRS) signals for the classification of Alzheimer’s Disease (AD) stages. Alzheimer’s disease is increasingly understood as a disconnection syndrome, where the disruption of functional brain networks precedes gross anatomical atrophy. However, traditional graph-theoretic approaches rely on arbitrary connectivity thresholds, which can obscure critical multi-scale topological information, and are sensitive to noise. To address this, our framework leverages Persistent Homology (PH) to analyse the topological evolution of brain networks across a continuous range of scales. By modeling 48-channel hemoglobin concentration time-series as high-dimensional point clouds via Granger causality metrics, we construct filtration sequences of Vietoris–Rips complexes. The resulting topological invariants, including 0—dimensional connected components, 1—dimensional loops, and 2—dimensional voids, are first examined through Persistence Diagrams. For classification, significant H0 and H1 features are converted into Persistence Images using Gaussian kernel smoothing, while H2 features are retained for qualitative topological interpretation. This transformation enables the integration of complex topological features into standard machine learning workflows. Our experimental results were evaluated on a subject-level held-out test set consisting only of original, non-augmented recordings. Data augmentation was applied only to the training set to alleviate class imbalance. The proposed topology-driven feature extraction method achieved 86% accuracy in multi-class diagnosis (NC vs. MCI vs. AD). This study validates the efficacy of TDA as a sophisticated signal processing tool for revealing intrinsic neurodegenerative patterns in hemodynamic data, offering an exploratory methodological proof-of-concept for AD stage classification. Full article