Search Results (3,149)

Background/Objectives: Glioblastoma surveillance after surgery and chemoradiation remains challenging because MRI findings of tumor recurrence can overlap with pseudo-progression, treatment-related effects, and postoperative tissue changes. Methods: We developed a seven-channel postoperative MRI framework using a deep learning transfer method to support non-invasive glioblastoma treatment-effect assessment. The model used T1 contrast-enhanced, FLAIR, T1, RSI-Cell, ADC, T2, and cerebral blood flow volumes from 124 postoperative glioblastoma patients (164 MRI timepoints) as input. Images were processed using a 3D ResNet18 encoder pretrained on 588 postoperative glioma samples and fine-tuned using task-specific classification heads. The cohort comprised 124 patients contributing 164 postoperative MRI timepoints, all acquired at 3T on scanners from a single vendor. Performance was evaluated with nested five-fold cross-validation stratified and assigned at the patient level, so that all timepoints from a given patient fell in one-fold and the training epoch was selected on an inner split rather than on the fold being reported. Because some of the clinical labels were incomplete, the number of evaluable timepoints differed by task (recurrence versus pseudo-progression, 164; MGMT, 99; short-term survival, 139). The whole procedure was repeated under three independent random seeds and results are reported as the mean and standard deviation across seeds. Results: The strongest clinical endpoint was recurrence versus pseudo-progression, where nested cross-validation across three random seeds gave a pooled out-of-fold AUC of 0.935 (SD = 0.014), area under the precision-recall curve of 0.973, balanced accuracy of 0.880, sensitivity of 0.917, and specificity of 0.843. No other endpoint reached reliable discrimination. Radiation decision reached an AUC of 0.658 (SD = 0.039), while MGMT promoter methylation (AUC = 0.532, SD = 0.082) and short-term survival (AUC = 0.514, SD = 0.027) were indistinguishable from chance. Conclusions: In this single-center proof-of-concept study, postoperative MRI successfully distinguished tumor recurrence from pseudo-progression. However, the models did not reliably predict the other three outcomes related to molecular status, treatment planning, and prognosis. Overall, the model learned imaging features specifically associated with recurrence, rather than a more general representation of the tumor that can predict many different clinical outcomes. These findings support technical feasibility for a single endpoint rather than clinical readiness, for which external multi-center validation is required. Full article

Fault diagnosis in grid-connected grid-forming hybrid energy storage station (GFM-HESS) systems is challenging because fault transients are jointly affected by converter control dynamics, multi-source electrical couplings, operating-condition variations, and measurement noise. To address these characteristics, this paper proposes a virtual synchronous generator (VSG)-constrained physics-informed heterogeneous temporal graph neural network (VSG-PI-HTGNN) for multi-class fault diagnosis. Based on VSG control characteristics and electrical relationships, 18-dimensional node-level features and eight-dimensional global physical features are constructed from ten monitored signals. These signals are further represented as a heterogeneous graph with five node types and eight predefined relation types, while node-type-specific transformations, heterogeneous graph convolution, learnable relation-scaling factors, and a primary–auxiliary dual-output framework are integrated for feature learning. A MATLAB/Simulink electromagnetic transient model is established to generate 3200 samples covering normal operation and nine fault conditions. At a signal-to-noise ratio (SNR) of 18 dB, the proposed model achieves 97.25% test accuracy and a macro-F1 score of 0.9727. Ablation results show that removing all physical information reduces the accuracy to 89.83%, while, under the unified experimental setting, the proposed model obtains higher values of the reported diagnostic metrics than the seven considered benchmark methods. Further evaluations show that the accuracy remains between 95.50% and 98.75% across SNR levels of 10–30 dB and reaches 95.38% with only 20% of the training data. Validation using an independently acquired hardware-in-the-loop (HIL) dataset further yields 92.75% accuracy and a macro-F1 score of 0.9282. Overall, these results indicate that the proposed method provides favorable diagnostic accuracy, noise robustness, data efficiency under limited-sample conditions, and simulation-to-HIL transferability under the evaluated conditions. Full article

Infrared hyperspectral remote sensing is effective for chemical gas detection because gas molecules exhibit characteristic absorption features in the mid- and long-wave infrared region. This study considers coexisting gas components along a common line of sight rather than geometrically distinct spatial plumes. We propose a compact spectral one-dimensional convolutional neural network (1D-CNN) for pixelwise gas species identification and the extraction of binary detection regions directly from 121-band spectra. Across five independently simulated NETD conditions and cross-noise train–test evaluations, the model maintained consistently strong classification performance over the examined noise range. On the measured scenes, the proposed model was compared with four traditional detectors and an adapted spectral Transformer baseline, showing a favorable and comparatively consistent Precision–Recall balance across the four target gases. These results support a fixed-platform field proof of concept, but do not constitute airborne-platform, edge deployment, quantitative concentration retrieval, or broad cross-site validation. Full article

Seamless steel pipes require reliable classification of outer-surface anomalies, yet class imbalance, scale variation, and visual overlap between defects and production interference complicate automated inspection. We propose WF-MobileNet for classifying two-dimensional RGB pseudo-colour surface morphology images derived from line-structured-light profiling. The model retains MobileNetV3-Small as its primary descriptor path, applies Selective WTConv to selected middle- and late-stage depthwise operations, and aggregates features at three spatial resolutions through Lite-BiFPN. A gated residual connection adds the multiscale descriptor to the terminal backbone descriptor. Evaluation used 15,463 images covering 12 defect classes and 4 production interference classes, with five training seeds on one fixed partition. WF-MobileNet achieved the highest observed mean macro-F1 and interference F1 among the six evaluated architectures, reaching (82.17 ± 0.36)% and (80.50 ± 0.96)%, respectively (mean ± sample standard deviation). The corresponding gains over MobileNetV3-Small were 4.20 and 6.84 percentage points. With 1.811 million parameters and 65.8 million multiply–accumulate operations (MACs), WF-MobileNet ranked second lowest on both complexity measures. Controlled ablation revealed larger mean macro-F1 gains from the joint configuration than from either component alone. Within the evaluated archive, WF-MobileNet improved class-balanced recognition and defect–interference discrimination while retaining low parameter and MAC counts. Full article

Bio-based alternatives to conventional rigid foams have proven to be good substitutes owing to their enhanced sustainability and competitive performance. However, because their manufacturing processes are complex and destructive testing is often impractical, this study investigates whether microstructural features can be correlated with mechanical properties in lignin-containing rigid polyurethane (PU) foams using machine learning approaches. Various types and percentages of lignin-based polyols were investigated as partial replacements for polyol, including LHO, DCA, DCA-D, LHO-O, Kraft lignin (KL), and LHO-MD, at polyol replacement levels ranging from 12.5% to 50%, together with a control formulation. Scanning electron microscopy (SEM) images and corresponding mechanical compression data were used to train a custom state-of-the-art dual-head convolutional neural network (CNN) targeting the specific prediction of density, specific compression modulus, specific yield stress, and specific compression strength. The CNN was optimized with a weighted multi-output loss function, achieving strong predictive performance with R² values ranging from 0.850 to 0.91 and correlation coefficients above 0.92, while maintaining mean absolute error percentages below ≈9%. This proves the trained network’s capability to predict and capture morphological features governing load-bearing responses. On the other hand, Grad-CAM visualization revealed that the network focused its predictions on physically meaningful microstructural regions such as cell walls and strut junctions, which confirms that the proposed network can be classified as an interpretable, non-destructive, and data-driven framework for predicting and understanding bio-based PU foams’ mechanical behavior, hence reducing the inconvenience caused by time-consuming manufacturing and destructive testing. Full article

Bearing fault diagnosis increasingly relies on multiple sensors, since compound or multi-location faults can produce signatures that are only partially captured by a single sensor. The decision-level fusion (DLF) of independently trained models offers a practical way to combine such complementary information, yet systematic comparisons across DLF techniques remain scarce, particularly for measurements from different sensor locations. This paper benchmarks six DLF strategies, i.e, Max, Average, Majority Voting, Weighted Sum, Dempster–Shafer, and Stacking, on a dual-branch one-dimensional convolutional neural network (1D-CNN) with each branch trained end-to-end on vibration signals from a distinct bearing location. On a two-sensor test bench covering seven health conditions, all methods exceed 99.7% accuracy on clean signals, while Dempster–Shafer fusion proves markedly more robust under noise, retaining up to 84% accuracy at a 5 dB signal-to-noise ratio (SNR). A conflict-coefficient analysis further provides an interpretable account of when fusion succeeds, linking performance to the confidence complementarity between branches. Full article

Electromagnetic scattering from plunging breaking waves generates strong sea-surface radar returns that degrade radar-based maritime surveillance and target detection performance. This study develops a deep learning framework for automatic classification of simulated plunging-breaker scattering conditions using high-range-resolution profile (HRRP) data. The electromagnetic scattering data are generated using a physics-based Capillary Wave Modification Facet Scattering Model (CWMFSM) combined with ray-tracing techniques. Eight simulated plunging-breaker scattering conditions are constructed by combining two wind speeds, 7 m/s and 10 m/s, with four temporal conditions, Δt1, Δt10, Δt14, and Δt16. A total of 8000 HRRP samples are generated, with 100 normalized range-cell features extracted from each sample. Two deep learning classifiers, an artificial neural network (ANN) and a one-dimensional residual convolutional neural network (1D ResNet CNN), are comparatively evaluated. The ANN achieves an overall classification accuracy of 96%, compared with 91% for the 1D ResNet CNN under the simulated dataset and adopted model configurations. Robustness analysis under controlled additive white Gaussian noise (AWGN) conditions further shows that classification performance decreases as the signal-to-noise ratio is reduced, while noise-augmented training improves the robustness of both classifiers. Overall, the results demonstrate the feasibility of HRRP-based deep learning for distinguishing simulated plunging-breaker scattering conditions from sea-surface radar returns, providing a basis for further investigation of sea-clutter characterization and maritime radar applications. Full article

Winter wheat is a cornerstone of China’s grain production, contributing substantially to national food security and overall cereal output. This study modeled the nonlinear associations between multitemporal remote sensing variables and winter wheat yield. To produce high-spatiotemporal-resolution inputs, we used the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) to integrate Sentinel-2 normalized difference vegetation index (NDVI) data with MODIS NDVI data, generating NDVI composites at 8-day intervals with a 10-m spatial resolution. The NDVI, actual evapotranspiration (ET), land surface temperature (LST), precipitation (PRE), and soil moisture (SM) were selected as predictors for yield estimation because they are closely associated with winter wheat growth and yield formation during primary growth stages. By integrating the local temporal feature-learning capacity of a one-dimensional convolutional neural network (1-D CNN) with the strength of a bidirectional long short-term memory (BiLSTM) model in capturing temporal dependencies within time series, a BiLSTM–CNN model was constructed for wheat yield estimation and prediction. The BiLSTM–CNN model showed higher estimation accuracy than individual BiLSTM and 1-D CNN models, with an R² of 0.69 and root mean square error (RMSE) of 478.68 kg/hm². The use of all the parameters produced the best estimation performance among all the parameter combinations. Approximately two months before harvest, the model still provided satisfactory yield prediction accuracy. This study provides an important theoretical basis for high-accuracy regional winter wheat yield estimation and pre-harvest forecasting. Full article

Hyperspectral image (HSI) classification benefits from rich spectral information; however, high dimensionality of HSI data increases computational cost, noise sensitivity, and the risk of overfitting when labeled samples are limited. Most pretrained computer vision networks are designed for three-channel inputs, making direct application to hyperspectral cubes difficult. Conventional principal component analysis (PCA) ranks components by total variance without distinguishing useful signal variance from noise-related variance, which can reduce the reliability of the resulting representation when only a few components are retained. This paper proposes a data-augmented Noise-Adjusted Principal Component Analysis (DA-NAPCA) framework for deep learning-based HSI classification. By accounting for estimated noise covariance, NAPCA orders the transformed components by signal-to-noise ratio rather than total variance, while data augmentation mitigates the overfitting risk when labeled samples are limited. Unlike typical NAPCA/MNF applications, which select the number of retained components empirically, DA-NAPCA deliberately retains three noise-adjusted components to form a compact three-channel representation, enabling pretrained models designed for three-channel inputs to be fine-tuned without modifying their input layers. The framework is evaluated using a 3D convolutional neural network (3D-CNN) for spatial–spectral feature learning and a pretrained EfficientNet-B0 model for lightweight transfer learning. Although this paper uses 3D-CNN and EfficientNet-B0 as illustrative examples, the proposed DA-NAPCA framework is a representation-level preprocessing approach and does not require architecture-specific modification. Experiments conducted on the Indian Pines, University of Pavia, and Salinas datasets compare DA-NAPCA with RGB, band selection, PCA-based dimensionality reduction, and ablation variants. Across the three datasets, DA-NAPCA achieved mean overall accuracies of 93.11–94.71% with 3D-CNN and 95.93–97.44% with EfficientNet-B0. Compared with the second-best baseline method, DA-NAPCA improved overall accuracy by 2.75–7.58 percentage points with 3D-CNN and 1.28–2.12 percentage points with EfficientNet-B0. These results demonstrate that combining a compact noise-adjusted representation with spatial augmentation provides an effective input representation for deep learning-based HSI classification. Full article

Reliable short-term position forecasting may provide anticipatory state information for collision-risk assessment, communication management, and prediction-assisted control in Internet of Things (IoT)-enabled unmanned aerial vehicles (UAVs); these downstream functions are not evaluated directly here. This study reformulates UAV position prediction as a flight-wise, multi-horizon, three-dimensional forecasting problem and tests whether position, velocity, gravity-resolved acceleration, and quaternion-orientation histories improve predictive accuracy while measuring model-level edge-inference cost rather than end-to-end system latency. The dataset contains 3100 AirSim flights with high-rate kinematic, inertial, attitude, pressure, and magnetic-field measurements under variable horizontal wind. The reported generalization is flight-disjoint within one AirSim domain; route/scenario disjointness and transfer to physical UAVs are not established. Signals are converted to a common navigation frame, gravity-resolved, low-pass filtered, resampled to 50 Hz, and partitioned by flight identifier before normalization and window construction. Each learned model receives 2 s of history and predicts the complete next 1 s trajectory, with errors evaluated at 0.1, 0.5, and 1.0 s. The sensor-enriched LSTM (LSTM-PVAQ) is compared under matched conditions with persistence, constant-velocity, constant-acceleration, extended Kalman filter, reduced-feature LSTM, GRU, temporal convolutional network (TCN), and compact Transformer baselines. LSTM-PVAQ achieved 3D RMSE values of 0.043, 0.168, and 0.371 m at 0.1, 0.5, and 1.0 s, respectively. At 1 s, its RMSE was 21.7% lower than LSTM-PV, 13.1% lower than GRU-PVAQ, 9.3% lower than TCN-PVAQ, and 16.8% lower than Transformer-PVAQ. Its one-second ADE and FDE were 0.216 and 0.339 m. On a Raspberry Pi 5 CPU using one FP32 thread and batch size one, median neural forward-pass latency was 0.88 ms, well below the 20 ms model-update interval. The results show that gravity-resolved inertial and orientation histories improve multi-horizon prediction, while TCN-PVAQ remains an attractive lower-latency alternative. Full article

Under intensified climate change, improving yield estimation accuracy is essential for food security and precision agricultural management. This research aimed to enhance summer maize yield estimation accuracy. GF-1 imagery was used to derive vegetation indices for leaf area index (LAI) modeling; vegetation index–LAI models used partial least squares regression (PLSR), random forest (RF), extreme gradient boosting (XGBoost), support vector regression (SVR), convolutional neural network (CNN), and categorical feature-enhanced gradient boosting (CatBoost). Retrieved summer maize LAI and measured soil moisture (SM) were assimilated into a calibrated World Food Studies (WOFOST) model using ensemble Kalman filtering (EnKF) and four-dimensional variational assimilation (4D-Var). The influence of ensemble size and assimilation period on the accuracy of yield simulations was analyzed. Results showed that all machine learning models effectively retrieved LAI, with CNN achieving the highest accuracy under the evaluation conditions of this study. Compared with the non-assimilation scenario, data assimilation significantly improved WOFOST-based maize yield estimation; EnKF performed best for yield estimation, whereas 4D-Var yielded slightly lower LAI simulation errors than EnKF under the experimental conditions of this study. Among the ensemble sizes evaluated, an ensemble size of 100 produced the lowest yield prediction error, while assimilation during the tasseling-to-grain-filling period resulted in lower yield prediction errors than continuous whole-growth-period assimilation under the conditions examined. Overall, data assimilation effectively improved WOFOST-based maize yield estimation and supported regional crop monitoring and precision agricultural management. Full article

Cross-condition bearing fault diagnosis suffers from severe performance degradation due to domain shift across different operating conditions, especially when only a few labeled target-domain samples are available. To address this challenge, this paper proposes a progressive attention-guided two-stage transfer learning framework for few-shot bearing fault diagnosis across different fixed operating points. First, the raw time-domain vibration signals are fused with frequency-domain representations extracted by short-time Fourier transform (STFT) to enhance fault feature representation. Then, a progressive attention-guided feature learning strategy is developed by integrating dual efficient channel attention (ECA) modules into a deep one-dimensional convolutional neural network (1D-CNN), enabling the network to adaptively emphasize fault-sensitive features while suppressing redundant information. Subsequently, a two-stage transfer learning strategy is designed, consisting of transferable feature learning from the source domain and few-shot adaptation to the target domain. During target-domain adaptation, key feature extraction layers are frozen, and a sample-balanced optimization mechanism is introduced to alleviate the dominance of source-domain samples during joint training. Experimental results on the Case Western Reserve University (CWRU) bearing dataset demonstrate that the proposed method achieves an average accuracy of 99.96% across three cross-condition transfer tasks. Furthermore, experiments conducted on a self-built shaft system dataset show that the proposed method achieves an average accuracy of 87.11% under three representative transfer scenarios. The results verify that the proposed framework effectively mitigates domain shift and enables accurate bearing fault diagnosis with limited labeled target-domain samples. Full article

This study assessed the integration of Unmanned Aerial Vehicles (UAVs) and satellite images (Sentinel-1 and Sentinel-2) in advanced machine-learning techniques to monitor the ABV of quinoa crops (Jacha Grano variety) across the Bolivian Altiplano. The proposed method follows a two-step procedure. First, UAV RGB images were used in photogrammetric and deep-learning (Convolutional Neural Networks-CNN) models to estimate reference quinoa ABVs at a 10 m spatial resolution from the crop canopy 3D model and classification, respectively. Secondly, several spectral and polarization/texture indices derived from Sentinel-2 and -1 images were integrated into three decision-tree-based machine-learning models (Random Forest-RF, Gradient Boosting-GB, eXtreme Gradient Boosting-XGB), and one CNN-based machine-learning model to estimate ABV. Additionally, a Stacking Model (STM) build on top of the three decision-tree-based models was considered for comparison. Model evaluation was also performed in a two-step approach. First, a 10-fold cross-validation strategy was used to highlight ABV sensitivity to Sentinel-2 and Sentinel-1 alone and in combination. Secondly, a Leave-One-Plot-Out Cross-Validation (LOPOCV) strategy was used to avoid autocorrelation between the training and evaluation dataset and therefore provided more insight into ABV mapping potential. The results showed that the combination of Sentinel-1 and Sentinel-2 features in the CNN model achieved the best predictive performance with R² and RMSE values of 0.64 and 0.39 m³ ∙ 100 m⁻², respectively. These findings highlight the potential of integrating multi-source information in advanced artificial intelligence algorithms for quinoa ABV monitoring, offering new insights toward the identification of sustainable practices across remote regions with complex socio-economic contexts. Full article

Respiratory-sound event classification is challenged by non-stationarity, class imbalance, heterogeneous acquisition, and participant-correlated recordings. This study evaluates whether direct fusion of short-time Fourier transform (STFT), mel-frequency cepstral coefficient (MFCC), and wavelet-packet features improves a temporal one-dimensional convolutional neural network (1D-CNN), and whether temporal convolution offers an advantage over conventional classifiers. A radial basis function support vector machine (RBF-SVM) and random forest were included deliberately to separate the value of the engineered representation from classifier complexity. Experiments used 920 recordings and 6898 annotated cycles from the International Conference on Biomedical and Health Informatics (ICBHI) 2017 Respiratory Sound Database. The predefined 60:40 recording-level benchmark partition was retained; model selection used five-fold participant-grouped cross-validation, and 95% confidence intervals were estimated from 1000 participant-level bootstrap resamples. The complete hybrid 1D-CNN achieved a macro-averaged F1-score of 0.313. MFCC alone yielded 0.320, but the paired difference was not statistically resolved. The RBF-SVM and random forest achieved 0.401 and 0.378, respectively. These findings apply to direct early concatenation with the shared 1D-CNN backbone and do not imply that feature fusion is generally ineffective. The study provides leakage-aware baselines, controlled ablations, clustered uncertainty estimates, and frozen artifacts for reproducible comparison. Full article

This work proposes a comparative framework for malware classification that evaluates the synergy between traditional feature engineering and modern deep learning architectures. Our methodology follows two primary paths: first, we integrate opcode n-grams with word-embedding techniques (Word2Vec, Doc2Vec, and FastText) to capture local execution patterns in dense vector spaces. Second, we evaluate end-to-end representations using transformer-based models (BERT and ViT) and a raw opcode-based 1D Convolutional Neural Network (1D-CNN) to determine if effective features can be learned without explicit n-gram engineering. Both pathways are rigorously tested across a suite of classifiers, including Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbor (k-NN), and CNNs. Experimental results for multi-class classification demonstrate that while transformer-based models offer high automated feature extraction capabilities, the combination of opcode n-grams with word embeddings remains a highly effective and interpretable approach for detecting real-world malware. Full article