Search Results (299)

Edge-side acoustic monitoring enables animal sound recognition in remote environments, but microcontroller deployment remains constrained by feature extraction, numerical consistency, memory, latency, and energy consumption. This study presents a sensor-based tiny machine learning (TinyML) acoustic monitoring system on an Arduino Nano 33 BLE Sense Rev2 platform, integrating onboard pulse-density modulation (PDM) microphone acquisition, Mel-frequency cepstral coefficient (MFCC) feature extraction, deployment-side standardization, 8-bit integer (INT8) neural-network inference, and edge-side decision output. To reduce training-to-deployment feature drift, consistent frame parameters, mirrored C++ feature operators, and exported standardization parameters are used to align personal-computer-side and microcontroller-side feature representations. A source-isolated seven-class protocol was constructed for six target animal classes and one compound background-noise class. In the single-run baseline comparison, the proposed multilayer perceptron achieved 98.28% test accuracy and 97.21% test macro-F1, while the ten-seed stability analysis yielded 98.64% ± 0.26% test accuracy and 97.87% ± 0.38% test macro-F1. The deployed INT8 model occupied approximately 26.9 KB, with a post-window latency of about 303 ms. System-level input power was 0.783–0.825 W, corresponding to an estimated autonomy of 7.63–8.03 h under the reference battery setting. Full article

Background/Objectives: Spontaneous intracranial hypotension (SIH) is caused by spinal cerebrospinal fluid (CSF) leakage and is typically diagnosed by clinical presentation and characteristic MRI signs; however, objective tools for monitoring physiological changes and treatment response remain limited. Cine phase-contrast MRI (PC-MRI) enables noninvasive quantification of aqueductal CSF dynamics, yet reliable analysis is challenging since the cerebral aqueduct is extremely small and susceptible to low contrast, partial volume effects, and ROI-dependent measurement variability—particularly in SIH where CSF pulsatility is often reduced. Methods: We propose an end-to-end automated framework that integrates (1) a cascade localization–segmentation strategy, consisting of Tiny YOLOv4 detection followed by MultiResUNet segmentation on a YOLOv4-derived cropped ROI; (2) physiology-informed pulsatility-based segmentation (PUBS) to refine anatomical masks into functional flow ROIs; and (3) one-dimensional convolutional neural networks (1D-CNNs) to extract exploratory waveform morphology features from 32-phase cardiac-cycle velocity waveforms. The study includes 39 participants, yielding 59 cine PC-MRI examinations: 11 controls, 28 Pre-treatment SIH scans and 20 Post-treatment Recovery scans. Results: The cascade model significantly improves segmentation robustness compared with a full-image baseline, achieving higher Dice scores and markedly lower boundary errors across cohorts (e.g., Pre-treatment SIH HD95: 1.66 ± 0.74 px vs. 15.37 ± 44.98 px). PUBS refinement reduces quantification deviation from expert manual references in SIH (mean relative error: 7.4% to 5.6%) and improves diagnostic performance for multiple hemodynamic parameters (e.g., downward mean flow AUC: 0.747 to 0.792). For waveform morphology analysis, the end-to-end 1D-CNN classifier was evaluated using repeated-seed participant-level grouped LOOCV. The repeated-seed ensemble prediction showed modest out-of-sample discrimination between Normal controls and Pre-treatment SIH scans, with an AUC of 0.646, a bootstrap 95% confidence interval of 0.455–0.826, and a permutation-test p-value of 0.072. Separately, exploratory analysis of the final baseline-trained 1D-CNN latent space showed marked, apparent Normal-versus-SIH separability and an intermediate recovery distribution in PCA space, suggesting that aqueductal waveform morphology may encode SIH-related physiological information. Conclusions: These findings suggest that SIH-related information may be reflected not only in flow magnitude but also in aqueductal CSF waveform morphology. However, the modest and statistically non-significant out-of-sample performance of the end-to-end 1D-CNN classifier indicates that morphology-based AI features should currently be regarded as exploratory biomarker candidates rather than validated stand-alone diagnostic tools. Larger independent cohorts are required to confirm their reproducibility, physiological meaning, and clinical utility. Full article

Electromyographic (EMG) signals are increasingly used in wearable cyber–physical systems (CPS), where reliable movement recognition must be achieved under limited computational resources. In this study, we present a compact EMG processing framework that integrates signal acquisition, preprocessing, segmentation, and movement classification within a unified pipeline designed for embedded-oriented applications. The proposed approach combines a multi-channel EMG acquisition system with a lightweight one-dimensional convolutional neural network (1D CNN) developed according to TinyML principles, withprocessing input windows of size 32 × 3 and low computational complexity and memory requirements. Experimental evaluation was conducted on a dataset collected from 15 participants performing squat, walking, and running activities under realistic acquisition conditions. The proposed model achieved an accuracy of 0.9135, an F1-score of 0.9124, and a ROC AUC of approximately 0.96, demonstrating reliable classification performance. Following 8-bit quantization, the model size was reduced to approximately 2 KB, supporting deployment on resource-constrained embedded platforms. The results show that compact CNN architectures can effectively classify EMG-based movement patterns while maintaining a small computational footprint, providing a practical foundation for future wearable CPS and TinyML-enabled applications. Full article

The integration of sensor technology and artificial intelligence is revolutionizing athletic training. This paper presents a novel cost-effective smart table tennis racket embedded with a nine-axis inertial measurement unit (IMU) for real-time stroke classification directly on the device. Unlike systems that are reliant on external computation, our approach leverages Tiny Machine Learning (TinyML) to deploy a customized Convolutional Neural Network (CNN) model onto a microcontroller unit (STM32F7), enabling real-time inference at the edge. The system captures accelerometer and gyroscope data, which is automatically segmented via a recursive algorithm and classified into six fundamental strokes (e.g., forehand/backhand stroke, pull, and chop) or a non-swing state. The classified results are wirelessly transmitted to a computer application for real-time feedback. Experimental results with actual players demonstrate that the optimized CNN model achieves an average classification accuracy of 98.3% in controlled tests and over 94% in mixed-stroke scenarios, validating the system’s high accuracy and robustness. This work exemplifies the practical implementation of an end-to-end intelligent sensor system, highlighting the potential of TinyML to enable advanced, low-power motion analysis in sports. Full article

Optimizing crop yield while minimizing energy consumption remains a central challenge in greenhouse horticulture. This study introduces an integrated deep learning framework that couples multi-horizon time-series forecasting with dual-layered explainability to address the critical need for spatiotemporal transparency in optimizing greenhouse crop yield and energy efficiency. Four deep learning architectures, including the One-Dimensional Convolutional Neural Network (1D-CNN), Long Short-Term Memory Network (LSTM), Bidirectional Long Short-Term Memory Network (BiLSTM), and TinyTimeMixer (TTM), were evaluated across two varieties of capsicum. LSTM and BiLSTM achieved the highest accuracy for incremental yield prediction, whereas TTM outperformed other models in forecasting daily energy usage, reflecting the distinct temporal characteristics of biological growth and environment-driven energy demand. To uncover the factors driving these predictions, two complementary explainability methods were applied: Gradient SHapley Additive exPlanations (SHAP) for feature-level attribution and a Temporal Convolutional Network with Convolutional Block Attention Module (TCN–CBAM) attention mechanism for joint temporal-feature interpretation. Radiation and drainage-related variables consistently emerged as the dominant contributors to yield, whereas external temperature, and humidity were the primary determinants of energy usage. Temporal attention further showed that yield is influenced by both recent irrigation responses and longer-term developmental dynamics, while energy consumption is driven mainly by short-term climatic fluctuations. These findings provide actionable insights for irrigation scheduling, climate-control strategies, and energy optimization, supporting more transparent and sustainable greenhouse management. Full article

While predictive maintenance increasingly relies on artificial intelligence, strict dependence on cloud computing introduces network latency and demands continuous connectivity, creating critical bottlenecks for time-sensitive industrial applications. To overcome this, we introduce a novel hybrid edge-cloud architecture, which allows deploying an ultra-low-power microcontroller (ESP32-S3) without dedicated AI acceleration hardware to perform complete, operational, predictive maintenance on ultra-constrained embedded hardware. The edge model is optimized to be very small to ensure that increasing model complexity does not cause inference latency to exceed 100 ms or make real-time operation infeasible. We created a very compact INT8-quantized neural network to perform the simultaneous classification of faults and estimation of Time-to-Failure (TTF) with a deterministic mean inference time of 42.3 ms. It dynamically estimates prediction confidence, processes high-confidence predictions locally, and offloads uncertain predictions to a higher-capacity cloud model, and recovers 97.3% of the cloud accuracy gain at 92% of the cloud latency budget. An asymmetric loss function penalizes over-prediction of the remaining useful life, and thus it provides conservative and safe warnings of fault. Operators’ interpretability is improved with Shapley Additive exPlanations (SHAP) and natural-language recommendations. Network outages of up to 50% have not influenced the safety-critical fault recall (above 0.924), so graceful degradation is reached when the network is used in real time in industrial applications. The edge-first with adaptive cloud fallback approach is demonstrated to be technically feasible for a full predictive maintenance workflow—including inference, confidence fusion, and explainability on a low-cost commercial microcontroller. Full article

Background/Objectives: The aim of this study is to test different convolutional neural network (CNN) and Transformer-based models to detect and classify vertical misfit at the abutment-prosthesis interface on panoramic radiographs, and to develop a hybrid deep learning model enhanced with attention mechanisms. Methods: A dataset consisting of a total of 566 images, manually classified as 249 ‘fit’ and 317 ‘misfit’ cases by two experts, was created. Images were resized to 224 × 224 and divided into training, validation, and test groups. The deep learning model yielding the most successful results was determined as the backbone; a hybrid model was developed by integrating three different attention modules (SE, CBAM, and ECA) into this structure. Model performance was evaluated using accuracy, precision, sensitivity, and F1 score metrics. Results: CNN-based models (RegNetY-800MF, ConvNeXt-Tiny, EfficientNetV2-S, ResNet50) performed better than Transformer-based models (DeiT, Swin-Tiny) in all metrics. The proposed hybrid model exhibited the highest success among all tested models with a 99.12% accuracy rate. This model reached a 100% precision value in the misfit group and yielded no false positive results. The F1 scores of the hybrid model were recorded as 99.01% for the fit group and 99.21% for the misfit group. Conclusions: The findings of this study demonstrate that attention-enhancing deep learning frameworks have the potential to significantly improve the diagnostic utility of routine panoramic radiographs. It shows that panoramic imaging, when supported by advanced artificial intelligence, can provide valuable diagnostic support in detecting vertical misfit. The developed model has the potential to become a reliable clinical decision support system. Full article

Background: Solar energy provides a sustainable and environmentally friendly alternative to fossil fuels, and photovoltaic (PV) systems are increasingly deployed worldwide. However, their operational reliability is often compromised by various fault conditions, which reduce power output and shorten system lifespan. Although automated image-based deep learning methods have shown promise for PV fault classification, their performance is often limited by severe class imbalance and subtle, low-contrast defect patterns. This study aims to address these challenges by proposing an improved deep learning framework for robust PV fault classification. Method: An attention-enhanced convolutional neural network framework, termed AEConvNeXt, is proposed for PV fault classification. The model is built on a ConvNeXt-Tiny backbone and incorporates a dropout-regularized Convolutional Block Attention Module (CBAM) to enhance localized feature refinement. To further improve learning under imbalanced data conditions, a hybrid loss function combining Cross-Entropy Loss and Focal Loss is employed. Results: Experimental evaluations demonstrate that AEConvNeXt achieves an overall accuracy of 94.37% and a macro F1-score of 94.43%, outperforming the strongest baseline model, ResNet-50, by more than 3%. Grad-CAM visualizations further confirm that the model effectively focuses on fault-relevant regions, improving interpretability. The proposed framework also shows consistent and robust performance across all six PV fault categories under varying conditions. Conclusions: The proposed AEConvNeXt framework provides an accurate and explainable solution for real-time PV fault detection, effectively addressing class imbalance and improving minority fault recognition. Full article

With the continuous growth of global population and increasing pressure on food security, the transformation toward precise and intelligent agricultural production has become an inevitable trend. In this context, accurate identification of field weeds is crucial for improving crop yields and reducing agricultural inputs. However, agricultural Internet of Things (IoT) edge devices are generally subject to strict constraints in terms of power consumption, storage, and real-time performance. Existing lightweight convolutional neural networks often struggle to simultaneously achieve high accuracy and low resource consumption for fine-grained weed identification tasks. To address this challenge, this paper proposes a hardware aware lightweight convolutional neural network named Weed-Aware Ghost Enhanced Network (WAGENet) for microcontroller deployment. The network synergistically integrates Ghost low-cost feature generation, Mobile Inverted Bottleneck Convolution (MBConv) for deep semantic extraction, Squeeze and Excitation (SE) and Coordinate Attention (CA) dual attention mechanisms for channel space joint calibration, and Atrous Spatial Pyramid Pooling (ASPP) for multi-scale context fusion. It constructs a progressive feature abstraction system from shallow textures to high-level semantics. On the public DeepWeeds dataset, WAGENet achieves 95.71% classification accuracy and 93.80% F1 score with only 0.163 M parameters and 2.43 × 10⁸ multiply accumulate operations (MACC), attaining a parameter efficiency of 587.19%/M and significantly outperforming existing mainstream lightweight models. The model has been successfully deployed on the STM32H7B3I microcontroller development board, achieving a single inference latency of 94.63 ms, an internal Flash footprint of only 686.95 KiB, and a single inference energy consumption of 41.45 mJ. Experimental results demonstrate that WAGENet achieves a trade off among accuracy, latency, and energy consumption under strict resource constraints, providing a reproducible microcontroller deployment paradigm for battery powered field robots, drones, and other agricultural IoT edge devices. Full article

This paper investigates the design and evaluation of compact convolutional neural networks (CNNs) for keyword spotting (KWS) and acoustic event detection under the stringent constraints of the TinyML paradigm. The research expands upon traditional binary classification approaches by addressing a multi-class acoustic scenario encompassing eight distinct categories: stop, no, go, yes, unknown, silence, noise_ambient, and noise_sudden. The primary objective is to evaluate the feasibility of deploying reliable acoustic detection systems on ultra-low-power microcontrollers for edge computing applications. To this end, five lightweight architectures were developed and benchmarked: AlexNet-Tiny, LeNet-Tiny, MobileNet-Tiny, VGG-Tiny, and CustomCNN-Tiny. The models were trained using Mel-spectrogram features and optimized through INT8 post-training quantization to facilitate embedded deployment. Hardware simulation was conducted targeting the XIAO nRF52840 Sense microcontroller (64 MHz, 256 KB RAM). Experimental results demonstrate that the Gold VGG-Tiny architecture achieves the highest classification accuracy ($89.81\%$), while Silver MobileNet-Tiny provides the superior operational efficiency with the lowest inference latency ($0.88$ ms) and minimal energy consumption ($14.4$ µJ). Furthermore, the Bronze CustomCNN-Tiny model achieves the most reduced memory footprint ($42.9$ KB), highlighting its suitability for memory-constrained environments. Statistical validation using Cohen’s Kappa, Matthews Correlation Coefficient (MCC), and Area Under the Curve (AUC) confirms the robustness and reliability of the proposed models. The potential application of this system is motivated by acoustic monitoring for the early detection of high-risk situations, such as gender-based violence. Future work will focus on on-device physical validation and real-world deployment in wearable safety electronics. Full article

The rapid proliferation of Internet of Things (IoT) devices has significantly expanded the attack surface of modern networks leading to a surge in IoT-based botnet attacks. Detecting such attacks remains challenging due to the high dimensionality and heterogeneity of IoT network traffic. This study proposes and evaluates three hybrid deep learning architectures for IoT botnet detection that combine representation learning with supervised classification: VAE-encoder-MLP, VAE-encoder-GAT, and VAE-encoder-MoTE. A Variational Autoencoder is initially trained to learn a compact latent representation of the high-dimensional traffic features. Subsequently, the pretrained VAE-encoder component is employed to project the data into a lower-dimensional embedding space. These embeddings are then used to train three different downstream classifiers: a multilayer perceptron (MLP), a graph attention network (GAT), and a mixture of tiny experts (MoTE) model. To further enhance representation discriminability, supervised contrastive learning is incorporated to encourage intra-class compactness and inter-class separability. The proposed architectures are evaluated on two widely studied benchmark datasets—the CICIoT2022 and N-BaIoT dataset—under both binary and multiclass classification settings. Experimental results demonstrate that all three models achieve near-perfect performance in binary attack detection, with accuracy exceeding 99.8%. In the more challenging multiclass scenario, the VAE-encoder-MLP model achieves the best overall performance, reaching accuracies of 98.55% on CICIoT2022 and 99.75% on N-BaIoT. These findings provide insights into the design of efficient and scalable deep learning architectures for IoT intrusion detection. Full article

Object detection in unmanned aerial vehicle (UAV) remote sensing imagery remains a formidable challenge due to the diminutive scale of targets, complex background clutter, and extreme variability in target morphology. Standard convolutional neural networks typically suffer from irreversible fine-grained information loss during downsampling, as strided operations discard critical spatial details essential for the localization of tiny objects. To address these issues, we propose RFE-YOLO, a lightweight receptive field-enhanced network specifically tailored for high-precision small object detection in UAV scenarios. First, the Cross-Scale Receptive Field Enhancement (CSRE) module is designed to mitigate intrinsic information loss by integrating space-to-depth convolution (SPD-Conv), which preserves spatial details by migrating them into the channel dimension. This module further employs an energy-based adaptive weight generation mechanism to distinguish target signals from environmental noise. Second, this paper proposes the C3k2-Dynamic Inception Mixer Block (C3k2-DIMB), which adaptively captures anisotropic features—such as slender vehicles—via dynamic kernel weighting and multi-shape inception kernels. Third, the Shuffled Upsampling for Resolution Enhancement (SURE) module is introduced to maintain spatial fidelity during resolution recovery, utilizing a channel shuffle mechanism to overcome information isolation. Finally, the Multi-feature Fusion Module (MFM) replaces conventional static concatenation with a dynamic softmax-based competition mechanism, effectively bridging the semantic gap between multi-level features while suppressing background distractors. Experimental results on the VisDrone dataset demonstrate that RFE-YOLO significantly enhances the representation capability for small objects. Specifically, the proposed model achieves a state-of-the-art $mA{P}_{50}$ of 42.70%, representing a substantial 9.3% improvement over the baseline YOLO11n. Furthermore, our architecture maintains an exceptionally lightweight profile with only 1.91 M parameters, demonstrating that high-precision detection can be achieved through structural intelligence rather than excessive parameter scaling. This makes RFE-YOLO highly suitable for real-time inference on edge-deployed UAV platforms. Full article

Recent advances in edge computing and Tiny Machine Learning (TinyML) have enabled the deployment of artificial intelligence models directly on microcontrollers with extremely limited computational and memory resources. In this context, this work presents the design, implementation, and validation of a real-time cardiac arrhythmia classification system based on a quantized one-dimensional convolutional neural network (1D-CNN), deployed on an 8-bit Arduino UNO microcontroller. The proposed system integrates end-to-end processing, including ECG signal acquisition using a low-cost AD8232 analog front-end, signal preprocessing, heartbeat segmentation, classification, and real-time visualization on an OLED display. The model was trained and evaluated using the MIT-BIH Arrhythmia Database, considering a reduced three-class problem (Normal, Ventricular, and Supraventricular) to meet the constraints of ultra-low-cost hardware deployment. Under benchmark conditions, the quantized model achieved an accuracy of 97.6%, with a memory footprint below 24 KB and an average inference time of 200 ms per heartbeat, enabling real-time operation on a resource-constrained microcontroller. Real-time experiments were conducted using signals acquired from healthy volunteers to validate system functionality, although no annotated ground truth was available for these recordings, and therefore no diagnostic performance was derived from them. The results demonstrate the feasibility of deploying lightweight deep learning models on ultra-constrained embedded systems using the TinyML paradigm, implemented using TensorFlow 2.15 and TensorFlow Lite. This work should be interpreted as a proof-of-concept platform that highlights the trade-off between classification performance and hardware limitations, providing a foundation for future development of low-cost cardiac monitoring technologies in resource-limited environments. Full article

Background/Objectives: Osteonecrosis of the femoral head (ONFH) is difficult to diagnose, particularly in the early stages, because radiological findings may be subtle. Delayed or inaccurate staging may increase the risk of femoral head collapse and functional loss. Although magnetic resonance imaging is highly sensitive for early-stage lesion detection, interpretation may vary depending on observer experience. Therefore, reliable and explainable automated decision support approaches are needed. Methods: In this study, a deep learning-based approach was proposed to classify ONFH into early and late stages according to the Ficat–Arlet staging system. Stage I–II cases were defined as early-stage, whereas Stage III–IV cases were defined as late-stage. Axial and coronal MR images were evaluated separately to investigate plane-dependent classification performance. The images were converted into a three-channel format, resized to a common spatial resolution, normalized, and augmented during training. Feature extraction was performed using transfer learning with modern convolutional neural network architectures. ConvNeXt Tiny was used as the main classification backbone. Weighted loss was applied to reduce the effect of class imbalance, and the decision threshold was optimized on validation data to reduce missed clinically critical late-stage cases. Results: A dataset collected from the Orthopedics and Traumatology Department of Firat University Hospital was used in the experimental evaluation. The dataset was divided into training and test sets using an 80:20 split, and 10-fold cross-validation was additionally performed to assess model stability. In the hold-out test, the axial plane model achieved 94.51% accuracy, 96.80% sensitivity, 93.49% specificity, 0.9162 F1-score, and 0.981 AUC. In the coronal plane model, 92.84% accuracy, 96.13% sensitivity, 90.96% specificity, 0.9072 F1-score, and 0.988 AUC were obtained. The 10-fold cross-validation results provided a more conservative estimate of generalization performance. Conclusions: The findings indicate that deep learning-based plane-wise analysis of MR images can distinguish early- and late-stage ONFH with high performance. Grad-CAM-based visual explanations showed that the model focused mainly on clinically relevant subchondral and weight-bearing regions of the femoral head. The proposed approach may serve as an explainable decision support tool for reducing observer-dependent variability in clinical staging. Future studies should validate the method using external, multicenter datasets and paired patient-level axial–coronal images. Full article

Blood pressure is a central marker of cardiovascular risk, but continuous monitoring remains difficult because cuff-based measurements are intermittent and uncomfortable. Photoplethysmography (PPG) is already ubiquitous in wearables and can, in principle, enable cuffless blood pressure estimation from a single optical signal. However, many deep learning approaches that perform well in floating-point are impractical for microcontroller-class devices, where memory budgets, latency, and integer-only arithmetic constrain what can be deployed. A key open question is which neural architectures retain accuracy after full-integer quantization, rather than only under desktop inference. Here, we show an end-to-end, microcontroller-oriented evaluation framework that benchmarks multiple 1D convolutional models for cuffless systolic and diastolic pressure estimation from single-channel PPG, jointly optimizing estimation error, model footprint, and quantization robustness. We find that floating-point accuracy alone is a poor predictor of deployability: some lightweight CNNs exhibit substantial performance drift after INT8 conversion, whereas a compact residual 1D CNN preserves its predictions with near-identical error statistics after integer quantization. We then deploy the selected integer-only model on an STM32N6 microcontroller using an industrial toolchain and confirm that on-device inference maintains low bias and limited error dispersion while meeting real-time constraints for continuous operation. These results highlight architecture-dependent quantization stability as a critical design dimension for sensor-edge intelligence and support the feasibility of fully on-device cuffless blood pressure monitoring without multimodal sensing or cloud processing. Full article