Search Results (7,359)

Hybrid CNN–Transformer (HCT) synthesis for automated neurodevelopmental diagnostics is an effective approach to constructing intelligent detection systems that are not merely oriented toward feature classification but primarily toward solving spatiotemporal pattern recognition problems in motor disorder assessment. In neurodevelopmental diagnostics, existing automated methods rely on fixed, single-model architectures that process spatial or temporal motion features independently, failing to adapt to the heterogeneous motor irregularities characteristic of developmental coordination disorder and degrading detection sensitivity and generalization across diverse patient populations. There is therefore a pressing need for models capable of simultaneously capturing intra-frame spatial coordination patterns and inter-frame temporal movement dependencies against interrelated diagnostic criteria including accuracy, sensitivity, and motor irregularity specificity. To address this challenge, this paper proposes HCT, a novel framework that integrates ResNet-based spatial feature extraction from optical flow maps and pose estimation skeletons with multi-head self-attention Transformer encoding for modeling long-range temporal dependencies across multi-frame motion sequences. Unlike conventional single-stream approaches, where spatial and temporal processing remain confined to independent architectures, HCT decouples spatiotemporal feature learning through a cross-modal fusion pipeline, constructing a unified discriminative architecture that captures motor coordination dependencies between motion imaging inputs and multiple diagnostic criteria simultaneously. The convolutional encoder generates diverse joint displacement features, which are consolidated through cross-modal attention fusion into a robust, unified embedding with enhanced generalization and resilience to inter-individual motor variability. Integration within neurodevelopmental assessment frameworks facilitates reliable developmental coordination disorder classification, motor irregularity prediction, and interpretable diagnostic decision support, advancing the accuracy, flexibility, and clinical validity of intelligent motor disorder diagnostic systems. Full article

Accurate and non-destructive assessment of drought stress is important for improving lettuce production and supporting timely crop management. This study presents a detection-guided deep learning framework for plant-level drought-stress assessment in hydroponically grown lettuce using bird’s-eye-view RGB images. The study further investigates whether canopy segmentation can improve classification performance by reducing irrelevant background information. The framework was evaluated using 2190 images collected across three independent cultivation cycles in which drought stress was induced by isolating the plant root zones from the nutrient solution. In the first stage, YOLO-based object detection was used to localize individual plants, with YOLO26m achieving the highest detection performance of 99.4% mAP@0.5. The detected regions were subsequently used as spatial prompts for zero-shot canopy segmentation using the Segment Anything Model (SAM), with SAM ViT-B achieving a mean IoU of 0.9864. Six convolutional, transformer-based, and hybrid classification architectures were then evaluated independently using YOLO-cropped and SAM-segmented plant images. Segmented inputs consistently improved classification performance, with MaxViT-S achieving the highest binary test accuracy of 96.3%. The framework further distinguished time-defined pre-stress, early-stress, and late-stress periods with an accuracy of 92.4%. Plant-level generalization was further assessed using six-fold leave-one-plant-out cross-validation, resulting in a mean test accuracy of 90.25 ± 1.78% on unseen plants. These findings demonstrate that RGB-based plant-level analysis can support non-destructive drought-stress assessment and that canopy segmentation improves classification by reducing background influence. Full article

Interactions among molecular recognition systems, neuronal activity, and glial regulation transform early neuronal connectivity into precise functional circuits during brain development. The cerebellum is a powerful model for studying these mechanisms due to its stereotyped and accessible circuitry. Two major excitatory afferent pathways—climbing fibers (CFs), which convey error-related signals to Purkinje cells (PCs), and mossy fibers (MFs), which transmit sensorimotor information via granule cells (GCs) and parallel fibers (PFs)—undergo strengthening, competition, and refinement during postnatal development. Synaptic specificity is established by general and pathway-specific organizers. The neurexin–neuroligin system broadly regulates synapse formation, while the neurexin–CBLN1–GluD2 complex specifies PF–PC synapses and C1qL1–BAI3 signaling stabilizes the dominant CF input during competitive refinement. CF–PC synapse elimination serves as a classic model of activity-dependent competition, where weaker inputs are removed through calcium-dependent mechanisms. In parallel, glial cells regulate synaptic maturation: microglia shape inhibitory environments, and Bergmann glia support glutamate homeostasis, dendritic organization, and synapse stability. PCs integrate CF and PF inputs and provide inhibitory output to the cerebellar nuclei, where convergent excitatory collaterals from CFs and MFs are combined with PC inhibition to generate cerebellar output. Together, these coordinated molecular, cellular, and circuit-level mechanisms establish the synaptic architecture underlying cerebellar computation, motor coordination, and adaptive learning, which are the central focus of this review. Full article

Short-term GPU utilization forecasts are useful for scheduling, resource allocation, and capacity planning, but production traces are rarely smooth. They contain spikes, regime changes, idle periods, and incomplete observations. We study this problem on the MIT Supercloud Dataset using a direct, horizon-specific forecasting setup. After resampling the telemetry to 1 min intervals, the neural models are trained on min–max-normalized data and evaluated on the original 0–100% utilization scale after inverse transformation. Persistence and rolling mean predictors are added as non-trainable baselines and are evaluated on the same eligible targets as the neural models. Five sequence models—1D-CNN, GRU, FC-LSTM, Liquid Time-Constant Network (LTC), and Transformer—are compared at 1 min, 10 min, and 1 h horizons. To avoid a gross capacity imbalance, the primary model widths are chosen in a comparable range of approximately 55,000 trainable parameters. This controls the trainable model size only; the architectures still differ in computation, memory access, and optimization behavior. Besides the overall error, the experiments examine high-load periods, abrupt changes, and a 20% random zero-masking condition. Among the five neural models, FC-LSTM gives the lowest MAE and RMSE at the 1 min horizon. The persistence baseline reaches an MAE of 3.92 at this horizon, compared with 3.55 for FC-LSTM, corresponding to a 9.4% lower MAE for FC-LSTM. At 1 h, among the neural models, LTC has the lowest mean RMSE, while FC-LSTM retains the lowest MAE and WAPE. Paired GPU device-level comparisons indicate that the larger improvements over simple baselines are more robust than the small numerical gaps among the strongest neural models. The zero-masking robustness protocol is expanded to five masking seeds and all five primary neural models. Taken together, the results show that the preferred model changes with both the forecast horizon and error criterion. Full article

Accurate identification of multiphase flow patterns in vertical pipelines is critical for operational safety and efficiency in the oil and gas industry. Yet, conventional methods struggle with subjectivity and transitional regimes. This study develops and integrates three neural network architectures: a convolutional neural network (CNN) for spatial features, a transformer neural network (TNN) for long-range dependencies, and a physics-encoding network (PEN) for embedding physical constraints. These are combined into a hybrid framework trained on an experimental dataset of 2131 images from a wire mesh sensor, annotated using a semi-automated pipeline. Results show the hybrid model achieved 95.91% test accuracy with a macro F1-score of 0.96, the highest of the four models evaluated, with its main advantage in transitional regimes. A multi-seed ablation shows that the convolutional branch provides the dominant discriminative signal, while the transformer and physics-inspired branches added complementary improvements that are consistent across runs. This hybridisation mitigates individual model weaknesses, with the physics-inspired branch acting as a spatial regulariser that improves interpretability, providing a robust and objective tool for reliable pipeline monitoring. 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

Maritime ship detection remains challenging because of large scale variations, high inter-class visual similarity, weak target boundaries, and complex maritime backgrounds. This study proposes MCSwin-YOLOv8, an enhanced YOLOv8-based detector that combines three complementary architectural designs. First, a re-parameterizable multi-scale convolutional backbone, named RepMCSwin, is introduced to extract scale-aware semantic information and fine-grained boundary cues. Unlike the standard Swin Transformer, the MCSwin block does not use window-based self-attention but adopts cascaded multi-scale convolutions and residual feature transformation. Second, a Multi-Feature Parallel Convolutional Block Attention Module (MFPCBAM) is developed to compute channel and spatial attention in parallel, thereby preserving weak ship features while suppressing irrelevant background responses. Third, a Modified Generalized Feature Pyramid Network (MGFPN) is constructed to improve cross-level feature interaction and retain high-resolution spatial information through an additional 160 × 160 prediction branch. Experiments were conducted on the public SeaShips dataset and a private infrared maritime ship dataset. MCSwin-YOLOv8 achieved an F1-score of 94.4%, mAP@0.5 of 97.4%, and mAP@0.5:0.95 of 75.1% on SeaShips. On the infrared dataset, the corresponding results were 91.2%, 94.1%, and 66.9%, respectively. Compared with the YOLOv8 baseline, mAP@0.5 increased by 1.4 and 3.0 percentage points on the two datasets. These accuracy gains were accompanied by an increase in model complexity from 11.12 M to 19.29 M parameters and from 28.5 G to 56.7 G FLOPs, indicating an accuracy–complexity trade-off that requires further runtime evaluation. Full article

Manufacturing sectors and ecosystems can improve their innovation performance through digital platforms, connected intelligence, and organisational settings that enable collaboration among experts and ecosystem members. The convergence of skills and capabilities distributed across humans, organisations, communities, and AI agents creates intelligent environments that can support ecosystemic and transformative innovation. To examine this hypothesis, we follow a three-stage methodology. First, we develop a modelling framework based on a vector autoregressive model, in which a weighted matrix representing directed binary couplings among human, collective, and machine intelligence drives the transition of a manufacturing ecosystem from a baseline innovation state to a more advanced one. Second, we present the SmartGreenEcos experiment, which develops an intelligent environment adapted to a specific manufacturing ecosystem. The experiment demonstrates the feasibility of the model’s abstract architecture by implementing digital platforms, e-services, and AI agents that facilitate inter-company collaboration, experimentation, and innovation. Third, we use simulations and analyse the eigenvalues and eigenvectors of the weighted matrix to examine the internal dynamics of intelligent environments and identify key thresholds and drivers of change. The results of this three-stage methodology provide insights into the design of intelligent environments and the interaction parameters through which connected intelligence can improve innovation performance. Full article

Antibiotic wastewater pollution is a serious environmental problem. This study prepared three types of P-doped biochar (P-CB, P-WB, and P-BB) from corn straw (CB), wheat straw (WB), and bamboo (BB) using phosphoric acid activation at 800 °C. Characterization by SEM, XRD, FTIR, and N₂ adsorption confirmed that high-temperature H₃PO₄ activation drastically reduced the specific surface area but transformed the original microporous framework into a mesopore-dominated structure (average pore diameter 12–15 nm). Adsorption experiments showed that all P-doped biochars enhanced norfloxacin (NOR) removal, with P-CB performing best: 87.14% removal within 30 min at C₀ = 50 mg·L⁻¹ and a maximum adsorption capacity of 305.5 mg·g⁻¹ at higher concentration. Kinetics followed a pseudo-second-order model (R² > 0.9982), and isotherms fitted the Freundlich model (R² = 0.9764–0.9855). Thermodynamics indicated a spontaneous and exothermic process, suggesting that the macroscopic driving force is dominated by physisorption. Mechanism analysis revealed that the synergistic effect of mesopore-dominated diffusion, graphitic π-π sites (from enhanced aromatization at 800 °C), and P-anchored chemisorption overrides the loss of specific surface area, enabling rapid and high-capacity adsorption. Optimal adsorption occurred at pH 5–9, while coexisting anions had a mild inhibitory effect. These findings demonstrate that high-temperature P-doping offers a strategy to rebalance pore architecture and surface functionality rather than simply maximizing specific surface area for efficient antibiotic removal from wastewater. Full article

Regenerative aluminum melting serves as a core process in recycled aluminum production. In the regenerative aluminum melting process, the furnace temperature is a key variable which affects product performance and energy costs. The extreme in-furnace temperature necessitates sensors equipped with protective jackets, which increases measurement costs and severely compromises real-time monitoring capability. Accordingly, accurate furnace temperature prediction is highly valuable for regenerative aluminum melting. In regenerative aluminum melting furnaces, periodic burner nozzle commutation and frequent material charging and discharging lead to complex and time-varying operating conditions, posing considerable challenges to high-precision furnace temperature prediction. To address these issues, a condition-adaptive multi-scale convolution-enhanced Transformer (CA-MC-Transformer) model is proposed for furnace temperature prediction. Firstly, an agglomerative hierarchical clustering algorithm based on the weighted dynamic time warping (WDTW) distance is designed to perform unsupervised clustering on historical process data, thereby extracting physically interpretable prior labels for macroscopic operating conditions. Secondly, multi-scale dilated causal convolutions are utilized to capture local dynamic features at diverse temporal resolutions. A soft attention mechanism is further introduced to dynamically assign fusion weights to condition embeddings and local features, enabling condition-adaptive feature reconstruction. Finally, the fused adaptive features are fed into an encoder-only Transformer network to capture the global long-range temporal dependencies and achieve accurate furnace temperature prediction. Comparative experiments conducted on real operational datasets from an aluminum plant verify that the proposed method effectively eliminates the inherent tracking lag of conventional deep learning models, and substantially improves prediction accuracy and anti-noise robustness under complex and variable operating conditions. Full article

The digital transformation of financial services has increased the need for intelligent information systems capable of supporting credit risk assessment in dynamic and data-intensive environments. Traditional credit scoring approaches often face limitations when evaluating customers with limited financial histories, heterogeneous data sources, and rapidly evolving behavioral patterns. In response to these challenges, this study proposes an adaptive credit risk assessment framework that integrates machine learning, an Adaptive Champion–Challenger strategy, and MLOps practices within a unified information systems architecture. The proposed framework was evaluated using real operational data obtained from a Mexican FinTech company specializing in mobile phone financing. Three machine learning algorithms—Logistic Regression, XGBoost, and TabNet—were implemented and continuously evaluated through a rolling Champion–Challenger process supported by out-of-time validation and statistically validated model promotion criteria. Experimental results indicate that different algorithms became optimal during different evaluation periods, indicating that model effectiveness varied over time as customer behavior and portfolio characteristics evolved. While XGBoost served as the initial static baseline model, TabNet and Logistic Regression achieved superior performance during several evaluation periods, illustrating the potential benefits of adaptive model selection under changing data conditions. The proposed Adaptive Champion–Challenger Framework achieved a mean AUC of 0.817, compared with 0.798 obtained by the static baseline model. Statistical validation using the DeLong test for correlated ROC curves confirmed that the observed performance improvement was significant (p = 0.0021), providing evidence that the performance gains achieved by the adaptive strategy were unlikely to be attributable to random variation. From a Digital Transformation and Information Systems perspective, the findings suggest that maintaining predictive effectiveness in dynamic FinTech environments requires not only high-performing machine learning algorithms but also governance mechanisms that support continuous model evaluation, monitoring, traceability, and adaptive model selection. The results indicate that periodic model replacement based on statistically validated out-of-time performance can help maintain predictive effectiveness under changing data conditions while supporting model governance and operational reliability. Overall, the proposed framework provides a practical and scalable approach for implementing adaptive credit risk assessment systems that support continuous model governance, data-driven decision-making, and the management of machine learning models in alternative financing environments. Full article

This study presents a sensor-fused field investigation and simulation-based analysis of four horizontal and vertical traffic-calming devices—two raised speed tables, a speed hump, and a raised crosswalk—installed along a 5250-ft two-lane residential collector in Nashville, TN, USA. A dual-sensor architecture combining a Miovision Scout video-based vehicle counter and WAAS/EGNOS-augmented GPS probe-vehicle logging (5 m 3-D RMS horizontal accuracy, 1 Hz sampling) was used to reconstruct 30 quality-controlled free-flow vehicle trajectories and 12-h per-lane volume counts. A spatial kinematic transform (a = v·dv/dx) was applied to extract device-specific approach-deceleration and post-device recovery-acceleration rates, and a three-parameter log-logistic cumulative-distribution function was fitted to the field-observed desired-speed percentiles (root-mean-square error below 0.043 for both speed-table devices). The camera- and GPS-derived observations were used to calibrate and statistically validate a PTV VISSIM microsimulation replica of the corridor, achieving a mean-speed calibration error of 0.71% or better at every device, a GEH statistic below 1.5 at all four analysis turning movements, and independent travel-time validation errors of 5.7–12.1%, within the accepted 15% threshold. The validated model was then used to reconstruct device- and spacing-specific May–Keller macroscopic speed–density–flow relationships, calibrated against simulated capacities of 650–775 vehicles per hour per lane at 350-, 700-, and 1050-ft device spacing. Results show capacity reductions of 20–33% relative to free-flow conditions and yield kinematically derived maximum recommended spacings of 265–630 ft to maintain crossing speeds at or below 15 mph, depending on device geometry. The findings demonstrate a reproducible, low-cost sensor-fusion workflow for quantifying the safety–capacity trade-off of traffic-calming corridors and for informing the design of sensor-in-the-loop adaptive-calming infrastructure. Full article

Accurate estimation of subsurface acoustic impedance is fundamental to quantitative reservoir characterization in seismic exploration. Nevertheless, a single network architecture cannot adequately represent both the local details and the global trends of seismic records within a unified framework, while the severe scarcity of annotated well-log data substantially constrains the generalization capability and predictive accuracy of deep-learning-based inversion approaches. To overcome these limitations, a semi-supervised acoustic impedance inversion framework based on a hybrid deep learning architecture is proposed. The framework employs a cascaded architecture consisting of a multi-scale depthwise separable convolution with channel attention (MSDSE) module and a convolution-augmented Transformer encoder. Seismic data are first processed by the MSDSE module to extract local multi-scale temporal features, and are subsequently passed to the convolution-augmented Transformer encoder, which captures global long-range sequence dependencies while retaining complementary local temporal information. The two modules progress hierarchically and jointly achieve a feature representation that spans from local details to global trends, and the initial low-frequency model is fused with the network output via channel-wise concatenation. Meanwhile, an initial-model constraint together with a physical-consistency constraint are simultaneously imposed within the loss function, thereby improving training stability while fully leveraging the physical information embedded in unlabeled traces. Experiments on both synthetic and field data confirm the effectiveness of the proposed method. The results show that, even with a small number of labels, the method produces stable impedance estimates and outperforms conventional deep learning methods in both generalization and prediction accuracy. Full article

Oil palm fresh fruit bunch (FFB) ripeness classification is essential for improving the oil extraction rate, oil quality, and mill processing efficiency. However, RGB-based classification is often limited by insufficient colour information from dark unripe fruitlets and shadowed regions of the bunch surface. This study evaluated the effect of image input representation on convolutional neural network (CNN)-based FFB ripeness classification. Four models with the same CNN architecture were compared using RGB, RGB + infrared (IR), CIEL*a*b*, and CIEL*a*b* + IR inputs. RGB and IR images were acquired using a dual-camera setup positioned 3 m from the FFB sample. The CIEL*a*b* + IR model achieved the best overall performance, with weighted-average precision, recall, and F1-score values of 0.84, 0.81, and 0.81, respectively, compared with 0.79, 0.76, and 0.76 for RGB-only. The addition of IR improved model performance by providing complementary near-infrared reflectance information, while CIEL*a*b* colour-space transformation provided a more discriminative colour representation. Class activation heat maps showed that the models focused mainly on fruitlet regions, with the CIEL*a*b* + IR model producing more distinct activation over ripeness-relevant areas. These findings demonstrate that the proposed four-channel CIEL*a*b* + IR image representation improves CNN-based FFB ripeness classification by combining lightness-separated colour information with near-infrared reflectance features, although underripe FFB remains challenging, where it produced the highest rate of false positives because of its transitional and heterogeneous characteristics. Full article

High-dimensional environmental covariates are increasingly available for digital soil mapping (DSM), but their effective use depends on both the feature-selection strategy and the predictive model architecture. However, systematic evidence remains limited regarding how different metaheuristic feature-selection methods interact with standalone and hybrid learning models across multiple soil and groundwater prediction tasks. This study systematically evaluated the interactions between 10 metaheuristic feature-selection algorithms and 13 predictive models, including random forest (RF), convolutional neural network (CNN), recurrent architectures, CNN–recurrent neural network (RNN) hybrids, squeeze-and-excitation (SE)-enhanced hybrids, and iTransformer-based hybrids, across four prediction tasks involving soil organic carbon (SOC), soil–water extract electrical conductivity (ECe), apparent electrical conductivity (ECa), and groundwater level (GWL) in Xinjiang, China. A total of 149 candidate environmental covariates were considered for ECe, SOC, and ECa, whereas 122 candidate covariates were considered for GWL. The results showed that no single feature-selection method consistently performed best across all four targets; instead, predictive performance depended on the interaction among the feature-selection strategy, predictive architecture, and target variable. CNN–RNN hybrid architectures generally achieved higher predictive performance than standalone models, although their benefits varied among prediction targets. The best-performing combinations yielded coefficient of determination (R²) values of 0.9826, 0.6981, 0.8429, and 0.8085 for GWL, SOC, ECe, and ECa, respectively. These findings indicate that target-specific compatibility, rather than aggressive dimensionality reduction or a universally superior algorithm, is a key determinant of predictive performance in high-dimensional DSM. By demonstrating that feature-selection effectiveness is jointly influenced by model architecture and target characteristics, this study provides a methodological reference for developing target-specific digital soil mapping models in arid regions. Full article