Search Results (34)

Body area networks (BANs) require secure intra-body communication, yet sensor nodes are too resource-constrained for conventional public-key cryptography, and pre-shared key schemes conflict with plug-and-play clinical workflows. This paper introduces PhysioKey, a TinyML-based key agreement framework that derives symmetric session keys from physiological signals without pre-shared secrets or trusted third parties. A lightweight 1D-CNN (6320 parameters, INT8-quantized, 31.2 KB flash) extracts embeddings from ECG and PPG windows on ARM Cortex-M4 class devices, which are reconciled through fuzzy commitment with BCH error-correcting codes. Patient-level 5-fold cross-validation on PTB-XL (500 patients, dual-ECG) achieves EER of $7.8\%\pm 0.8\%$ with ROC AUC $0.978\pm 0.004$; on BIDMC (53 patients, ECG + PPG), a dual-encoder architecture reduces cross-modal EER to $30.6\%\pm 1.2\%$. Since standalone PhysioKey yields only 7–24 effective key bits, the recommended deployment mode is a hybrid PhysioKey + ECDH protocol providing 128-bit security while PhysioKey adds physical on-body authentication; standalone operation suits energy-constrained scenarios with its $27\times$ advantage over ECDH. HKDF-SHA-256 post-processing yields session keys passing all six NIST SP 800-22 tests (≥96% at the 1024-bit level). Full article

Knee joint moment is an important biomechanical parameter for sports assessment, rehabilitation monitoring, and human–machine interaction. However, direct measurement is often restricted to laboratory-based settings. Surface electromyography (sEMG) offers a non-invasive alternative for indirect joint moment estimation, but many existing deep learning models remain too computationally demanding for potential wearable edge deployment. To address this gap, this study proposes Topo2DCNN-LSTM, a lightweight two-dimensional (2D) convolutional neural network model, designed for sagittal-plane knee flexion moment estimation. The model used a feature-based sequential representation, transforming raw sEMG signals into compact Root Mean Square (RMS) feature frames. The input was processed by a lightweight 2D convolutional neural network (CNN) encoder and paired with long short-term memory (LSTM) units. The model was trained on a public walking dataset of healthy subjects with synchronized sEMG and joint kinetics at two treadmill speeds. When compared with selected deep learning baselines, the quantized model achieved a mean RMS Error of 0.088 ± 0.020 Nm/kg at 1.2 m/s and 0.114 ± 0.034 Nm/kg at 1.8 m/s. On a SparkFun Thing Plus–SAMD51, it achieved an average inference latency of 28 ms using 71,316 bytes of random-access memory (RAM) and 257,172 bytes of flash. These results support its use as a proof of concept for personalized unilateral knee moment estimation with isolated on-device inference feasibility under resource-constrained and limited walking conditions. Full article

A method for simultaneous classification of fabric material and color based on hyperspectral imaging and visual detection is proposed. Fabric material classification is performed using hyperspectral imaging (HSI) combined with a one-dimensional convolutional neural network (1D-CNN), while fabric color recognition is achieved using an red-green-blue (RGB) camera and a color classification model. Material and color features from the same fabric sample are matched to realize synchronous classification. Experiments were conducted on three fabric materials (cotton, polyester, and cotton–polyester blend) and eight colors. At a conveyor speed of 1 m/s, the sorting success rates reach 95.0% for cotton, 97.5% for polyester, and 85.0% for cotton–polyester blended fabrics. The proposed method demonstrates reliable performance for single-material fabrics and good industrial applicability for automated fabric sorting. Full article

Leaf mass per area (LMA) plays an important role in vegetation productivity, carbon cycling, and remote sensing-based ecosystem monitoring. However, remotely predicting LMA from hyperspectral reflectance remains challenging due to the weak and strongly overlapping spectral response of LMA and spectral variability across species. To address these limitations, this study proposed an integrated framework that combines a fractional-order spectral derivative (FOD) with a one-dimensional convolutional neural network (1D-CNN) to enhance LMA prediction accuracy and cross-species generalization. Leaf hyperspectral reflectance was processed using FOD with 0–2 orders, and the relationship between FOD-enhanced spectra and LMA was analyzed. Model performance was assessed using (i) overall prediction accuracy by an 8:2 random split between training and test sets, and (ii) cross-species generalization through leave-one-species-out validation. The results demonstrated that the 1D-CNN using a 1.5-order derivative achieved the best performance (R² = 0.85; RMSE = 11.57 g/m²), outperforming common machine-learning models including partial least squares regression (PLSR), random forest (RF), and support vector regression (SVR). The proposed method also demonstrated great generalization in cross-species prediction. These results indicate that integrating FOD with 1D-CNN effectively enhances LMA-related spectral information and improves LMA prediction across various species. It provides a promising pathway for applying airborne and satellite hyperspectral images in vegetation biochemical parameter mapping, crop monitoring, and ecological assessment. Full article

Accurate upper-limb motor assessment is critical for post-stroke rehabilitation but relies on subjective clinical scales. This study proposes the Action-Aware Multimodal Wavelet Fusion Network (AMWFNet), integrating surface electromyography (sEMG) and robotic kinematics for automated Fugl-Meyer Assessment (FMA-UE)-aligned quantification. Continuous Wavelet Transform (CWT) converts heterogeneous signals into unified time-frequency scalograms. A learnable modality gating mechanism dynamically weights physiological and kinematic features, while action embeddings encode task contexts across 18 standardized reaching tasks. Validated on 40 participants (20 post-stroke, 20 healthy), AMWFNet achieved 94.68% accuracy in six-class classification, outperforming baselines by 9.17% (Random Forest: 85.51%, SVM: 85.30%, 1D-CNN: 91.21%). The lightweight architecture (1.27 M parameters, 922 ms inference) enables real-time assessment-training integration in rehabilitation robots, providing an objective, efficient solution. Full article

Background: Ground contact (GC) detection is essential for sprint performance analysis. Inertial measurement units (IMUs) enable field-based assessment, but their reliability during sprint acceleration remains limited when using heuristic and recently used machine learning algorithms. This study introduces a deep learning one-dimensional convolutional neural network (1D-CNN) to improve GC event and GC times detection in sprint acceleration. Methods: Twelve sprint-trained athletes performed 60 m sprints while bilateral shank-mounted IMUs (1125 Hz) and synchronized high-speed video (250 Hz) captured the first 15 m. Video-derived GC events served as reference labels for model training, validation, and testing, using resultant acceleration and angular velocity as model inputs. Results: The optimized model (18 inception blocks, window = 100, stride = 15) achieved mean Hausdorff distances ≤ 6 ms and 100% precision and recall for both validation and test datasets (Rand Index ≥ 0.977). Agreement with video references was excellent (bias < 1 ms, limits of agreement ± 15 ms, r > 0.90, p < 0.001). Conclusions: The 1D-CNN surpassed heuristic and prior machine learning approaches in the sprint acceleration phase, offering robust, near-perfect GC detection. These findings highlight the promise of deep learning-based time-series models for reliable, real-world biomechanical monitoring in sprint acceleration tasks. Full article

Accurate and timely crop yield prediction is crucial for global food security and modern agricultural management. Traditional methods often lack the scalability and granularity required for precision farming. This paper introduces FARM (Fine-tuning Agricultural Regression Models), a deep learning framework designed for high-resolution, intra-field canola yield prediction. FARM leverages a pre-trained, large-scale geospatial foundation model (Prithvi-EO-2.0-600M) and adapts it for a continuous regression task, transforming multi-temporal satellite imagery into dense, pixel-level (30 m) yield maps. Evaluated on a comprehensive dataset from the Canadian Prairies, FARM achieves a Root Mean Squared Error (RMSE) of $0.44$ and an $R^2$ of $0.81$. Using an independent high-resolution yield monitor dataset, we further show that fine-tuning FARM on limited ground-truth labels outperforms training the same architecture from scratch, confirming the benefit of pre-training on large, upsampled county-level data for data-scarce precision agriculture. These results represent improvement over baseline architectures like 3D-CNN and DeepYield, which highlight the effectiveness of fine-tuning foundation models for specialized agricultural applications. By providing a continuous, high-resolution output, FARM offers a more actionable tool for precision agriculture than conventional classification or county-level aggregation methods. This work validates a novel approach that bridges the gap between large-scale Earth observation and on-farm decision-making, offering a scalable solution for detailed agricultural monitoring. Full article

Rolling element bearing fault diagnosis (BFD) is fundamental to Predictive Maintenance (PdM) strategies for rotating machinery, as early anomaly detection prevents catastrophic failures, reduces unplanned downtime, and optimizes operational costs. This study introduces an interpretable Deep Learning (DL) framework that rigorously compares the performance of an Artificial Neural Network–Multilayer Perceptron (ANN-MLP), a one-dimensional Convolutional Neural Network (1D-CNN), and a ResNet-1D architecture for classifying seven bearing health states using a compact vector of 15 statistical features extracted from vibration signals. Both baseline models (ANN-MLP and 1D-CNN) failed to detect the critical Abrasive Particles fault (F1 = 0.0000). In contrast, the ResNet-1D architecture achieved statistically superior diagnostic performance, successfully resolving the most challenging class with a perfect F1-score of 1.0000 and an overall macro F1-score of 0.9913. This superiority was confirmed by a paired t-test on 100 bootstrap samples, establishing a highly significant difference in performance against the 1D-CNN ($t=592.702$, $p=0.00000$). To boost transparency and trust, the SHapley Additive exPlanations (SHAP) method was applied to interpret the ResNet-1D’s decisions. The SHAP analysis revealed that the Crest Factor from Sensor 1 (Crest_1) exerts the strongest influence on the critical Abrasive Particles fault predictions, physically validating the model’s intelligence against established domain knowledge of impulsive wear events. These findings support transparent, highly reliable, and evidence-based decision-making in industrial PdM applications within Industry 4.0 environments. Full article

Data loss is a recurring and critical issue in Structural Health Monitoring (SHM) systems, often arising from a range of factors including sensor malfunction, communication breakdown, and exposure to adverse environmental conditions. Such interruptions in data availability can significantly compromise the accuracy and reliability of structural performance assessments, thereby hindering effective decision-making in safety evaluation and maintenance planning. In this study, a novel deep learning-based framework is proposed for data reconstruction in SHM, employing a hybrid architecture that integrates one-dimensional convolutional neural networks (1D-CNNs) with recurrent neural networks (RNNs). By combining these complementary strengths, the hybrid 1D-CNN–RNN model demonstrates superior capacity for accurate signal reconstruction. A real-world case study was conducted using vibration data from the Trai Hut Bridge in Vietnam. Five network configurations with varying depths were examined under single- and multi-channel loss scenarios. The results confirm that the method can accurately reconstruct lost signals. For single-channel loss, the best configuration achieved an MAE = 0.019 m/s² and R² = 0.987, while for multi-channel loss, a deeper network yielded an MAE = 0.044 m/s² and R² = 0.974. Furthermore, the model exhibits robust and stable performance even under more demanding multi-channel data loss conditions, highlighting its resilience to practical operational challenges. The results demonstrate that the proposed CNN–RNN framework is accurate, robust, and adaptable for practical SHM data reconstruction applications. Full article

Atrial fibrillation (AF) is the most common type of heart rhythm disorder worldwide. Early recognition of brief episodes of atrial fibrillation can provide important diagnostic information and lead to prompt treatment. AF is mainly characterized by an irregular heartbeat. Today, many personal devices such as smartphones, smartwatches, smart rings, or small wearable medical devices can detect heart rhythm. Sensors can acquire different types of heart-related signals and extract the sequence of inter-beat intervals, i.e., the instantaneous heart rate. Various algorithms, some of which are very complex and require significant computational resources, are used to recognize AF based on inter-beat intervals (RR). This study aims to verify the possibility of using neural networks algorithms directly on a microcontroller connected to sensors for AF detection. Sequences of 25, 50, and 100 RR were extracted from a public database of electrocardiographic signals with annotated episodes of atrial fibrillation. A custom 1D convolutional neural network (1D-CNN) was designed and then validated via a 5-fold subject-wise split cross-validation scheme. In each fold, the model was tested on a set of 3 randomly selected subjects, which had not previously been used for training, to ensure a subject-independent evaluation of model performance. Across all folds, all models achieved high and stable performance, with test accuracies of 0.963 ± 0.031, 0.976 ± 0.022, and 0.980 ± 0.023, respectively, for models using 25 RR, 50 RR, and 100 RR sequences. Precision, recall, F1-score, and AUC-ROC exhibited similarly high performance, confirming robust generalization across unseen subjects. Performance systematically improved with longer RR windows, indicating that richer temporal context enhances discrimination of AF rhythm irregularities. A complete Edge AI prototype integrating a low-power ECG analog front-end, an ARM Cortex M7 microcontroller and an IoT transmitting module was utilized for realistic tests. Inferencing time, peak RAM usage, flash usage and current absorption were measured. The results obtained show the possibility of using neural network algorithms directly on microcontrollers for real-time AF recognition with very low power consumption. The prototype is also capable of sending the suspicious ECG trace to the cloud for final validation by a physician. The proposed methodology can be used for personal screening not only with ECG signals but with any other signal that reproduces the sequence of heartbeats (e.g., photoplethysmographic, pulse oximetric, pressure, accelerometric, etc.). Full article

Urban railway bridges are critical components of modern transportation networks. Dynamic loads and harsh environments put urban railway bridges at high risk of damage. Conventional vibration-based damage detection approaches often fail to provide sufficient accuracy and robustness under complex urban conditions. To address this limitation, this study introduces a hybrid deep learning framework that integrates a one-dimensional convolutional neural network (1D-CNN) and a recurrent neural network (RNN) for automatic damage detection using Linear Variable Differential Transformer (LVDT) displacement data. Start with the calibration of a finite element model (FEM) of the target bridge, achieved through updating the model parameters to align with field-acquired LVDT data, thereby establishing a robust and reliable baseline representation of the structure’s behaviour. Subsequently, a series of failure and damage scenarios is introduced within the FEM, and the associated dynamic displacement responses are generated to construct a comprehensive synthetic training dataset. These time-series responses serve as input for training a hybrid deep learning architecture, which integrates a one-dimensional convolutional neural network (1D-CNN) for automated feature extraction with a recurrent neural network (RNN) designed to capture the temporal dependencies inherent in the structural response data. Results show rapid convergence and minimal error in single-damage cases, and robust performance in multi-damage conditions on a dataset exceeding 5 million samples; the model attains a mean absolute error of ≈3.2% for damage severity and an average localisation error of <0.7 m. The findings highlight the effectiveness of combining numerical simulation with advanced data-driven approaches to provide a practical, data-efficient, and scalable solution for structural health monitoring in the urban railway context. Full article

The combined application of machine or deep learning algorithms and hyperspectral imagery for bathymetry estimation is currently an emerging field with widespread uses and applications. This research topic still requires further investigation to achieve methodological robustness and accuracy. In this study, we introduce a novel methodology for shallow bathymetric mapping using a one-dimensional convolutional neural network (1D-CNN) applied to PRISMA hyperspectral images, including refinements to enhance mapping accuracy, together with the optimization of computational efficiency. Four different 1D-CNN models were developed, incorporating pansharpening and spectral band optimization. Model performance was rigorously evaluated against reference bathymetric data obtained from official nautical charts provided by the Servicio de Hidrografía Naval (Argentina). The BoPsCNN model achieved the best testing accuracy with a coefficient of determination of 0.96 and a root mean square error of 0.65 m for a depth range of 0–15 m. The implementation of band optimization significantly reduced computational overhead, yielding a time-saving efficiency of 31–38%. The resulting bathymetric maps exhibited a coherent depth gradient from nearshore to offshore zones, with enhanced seabed morphology representation, particularly in models using pansharpened data. Full article

Coating degradation monitoring is increasingly important in offshore industries, where protective layers ensure corrosion prevention and structural integrity. In this context, coating thickness estimation provides critical information. The ultrasound pulse-echo technique is widely used for non-destructive testing (NDT), but closely spaced acoustic interfaces often produce overlapping echoes, which complicates detection and accurate isolation of each layer’s thickness. In this study, analysis of the pulse-echo signal from a coated sample has shown that the front-coating reflection affects each main backwall echo differently; by comparing two consecutive backwall echoes, we can cancel the acquisition system’s impulse response and isolate the propagation path-related information between the echoes. This work introduces an ultrasound echo-based methodology for estimating coating thickness by first obtaining the impulse response of the test medium (reflectivity sequence) through a deconvolution model, developed using two consecutive backwall echoes. This is followed by an enhanced detection of coating layer thickness in the reflectivity function using a 1D convolutional neural network (1D-CNN) trained with synthetic signals obtained from finite-difference time-domain (FDTD) simulations with k-Wave MATLAB toolbox (v1.4.0). The proposed approach estimates the front-side coating thickness in steel samples coated on both sides, with coating layers ranging from $60\mathsf{\mu }$m to $740\mathsf{\mu }$m applied over 5 mm substrates and under varying coating and steel properties. The minimum detectable thickness corresponds to approximately $\mathit{\lambda }/5$ for an 8 MHz ultrasonic transducer. On synthetic signals, where the true coating thickness and speed of sound are known, the model achieves an accuracy of approximately $8\mathsf{\mu }$m. These findings highlight the strong potential of the model for reliably monitoring relative thickness changes across a wide range of coatings in real samples. Full article

This study proposes Dual-Stream Former, a novel architecture that integrates a Video Swin Transformer and Conformer designed to address the challenges of visual speech recognition (VSR). The model captures spatiotemporal dependencies, achieving a state-of-the-art character error rate (CER) of 3.46%, surpassing traditional convolutional neural network (CNN)-based models, such as 3D-CNN + DenseNet-121 (CER: 5.31%), and transformer-based alternatives, such as vision transformers (CER: 4.05%). The Video Swin Transformer captures multiscale spatial representations with high computational efficiency, whereas the Conformer back-end enhances temporal modeling across diverse phoneme categories. Evaluation of a high-resolution dataset comprising 740,000 utterances across 185 classes highlighted the effectiveness of the model in addressing visually confusing phonemes, such as diphthongs (/ai/, /au/) and labio-dental sounds (/f/, /v/). Dual-Stream Former achieved phoneme recognition error rates of 10.39% for diphthongs and 9.25% for labiodental sounds, surpassing those of CNN-based architectures by more than 6%. Although the model’s large parameter count (168.6 M) poses resource challenges, its hierarchical design ensures scalability. Future work will explore lightweight adaptations and multimodal extensions to increase deployment feasibility. These findings underscore the transformative potential of Dual-Stream Former for advancing VSR applications such as silent communication and assistive technologies by achieving unparalleled precision and robustness in diverse settings. Full article

Distributed denial-of-service (DDoS) attacks are a prevalent threat to resource-constrained IoT deployments. We present an edge-based detection and mitigation system integrated with the oneM2M architecture. By using a Raspberry Pi 4 client and five Raspberry Pi 3 attack nodes in a smart-home testbed, we collected 200,000 packets with 19 features across four traffic states (normal, SYN/UDP/ICMP floods), trained Decision Tree, 2D-CNN, and LSTM models, and deployed the best model on an edge computer for real-time inference. The edge node classifies traffic and triggers per-attack defenses on the device (SYN cookies, UDP/ICMP iptables rules). On a held-out test set, the 2D-CNN achieved 98.45% accuracy, outperforming the LSTM (96.14%) and Decision Tree (93.77%). In end-to-end trials, the system sustained service during SYN floods (time to capture 200 packets increased from 5.05 s to 5.51 s after enabling SYN cookies), mitigated ICMP floods via rate limiting, and flagged UDP floods for administrator intervention due to residual performance degradation. These results show that lightweight, edge-deployed learning with targeted controls can harden oneM2M-based IoT systems against common DDoS vectors. Full article