50% of energy in mineral processing is consumed in a grinding ball mill. Prolonged ball impacts on mill liner lead to a focused high-impact wear zone, which changes with process parameters. This study presents a sensor fusion unobtrusive system for imaging high-impact regions inside the mill, serving as a precursor to targeted wear quantification and early diagnostics. A novel onset detection algorithm isolates impact transients from 1.5-cm steel shell reverberation, enabling axial and angular localization errors of < 9 degrees (3% of circumference) against physical ground truth and Powell's trajectory model respectively. Unlike standalone wideband beamforming, which fails in this thick-wall environment, the proposed method reduces angular spread by 2.70x versus kurtosis-based preprocessing, enabling predictive liner maintenance without invasive instrumentation.
Retinal disorders are the leading cause of partial and permanent vision loss across the globe, making early diagnosis and timely intervention paramount for effective treatment planning. Recent deep learning methods for retinal disease diagnosis from fundus images acquired using optical imaging sensors primarily focus on single-disease detection, despite the fact that a fundus image may contain multiple diseases. However, the multi-label retinal disease classification task remains challenging due to high inter-class similarity, large variations in lesion shape and size, variations in sensor-acquired image quality, and class imbalance. While limited attempts have been made to this end, they remain limited in effectively capturing the complex relationships among diseases. In addition, they fall short of extracting salient and richer visual features and their relationships, resulting in suboptimal performance. To address these issues, we propose a self-attention guided two-stream graph convolutional network (SAT-GCNet), a unified framework that seamlessly models visual features and disease relationships via graph-based learning for effective multi-label retinal disease diagnosis. Specifically, in the first stream, we employ multiple CNN backbones to capture diverse visual patterns from fundus images and subsequently introduce a self-attention guided adaptive GCN (SAA-GCN) block to model feature correlations. While in the second stream, domain-specific label embeddings are generated and used in the SAA-GCN block to effectively capture disease correlations and incorporate rich semantic information, thereby guiding the model for better prediction. Extensive experimental evaluations on two benchmark datasets, OIA-ODIR and MuReD, demonstrate that SAT-GCNet achieves superior performance compared to current state-of-the-art approaches, while also providing enhanced reliability.
Image-based industrial inspection enables fast detection of cracks, dents and coating defects but often lacks metric quantification, while 3D methods provide spatial information at the cost of fine-scale acquisition and challenging automation. and coating defects but often lacks metric quantification, while 3D methods typically involve greater acquisition complexity and challenging automation. When accurate CAD models are available, quantitative inspection can instead be conducted as high-precision image-to-model registration under imperfect initialization and geometric deviations. This study introduces a raycasting-based iterative registration method that maps stereo image keypoints to a reference mesh and refines pose via triangulation and Iterative Closest Point (ICP) algorithm. To emulate positioning uncertainty, the measurement object is mounted on a three-axis linear stage and translated from its nominal pose. Results show sub-0.2mm translation accuracy for initial offsets up to 2.5mm. A sequential procedure with additional coarse alignment extends the convergence range to 8mm. Conventional registration exhibits systematic residual error along the dominant geometric sliding axis, which is reduced by the proposed refinement strategy.
Radar-based human sensing has emerged as a robust modality for activity monitoring and privacy-preserving healthcare. However, its development is constrained by the limited availability of diverse and annotated radar datasets. In this paper, we propose a physics-grounded cross-modal framework that synthesizes radar echo signals directly from monocular video. By integrating data-driven pose estimation with physics-constrained scale recovery, the proposed method reconstructs metric-consistent human motion in 3D space. Furthermore, a surface-consistent scattering model based on a flattened ellipsoid representation is introduced to capture realistic electromagnetic reflections under millimeter-wave conditions. Experimental results show that the synthesized radar spectrograms reproduce motion dependent Doppler characteristics and achieve over 82% structural similarity with real measurements. The proposed approach provides a low-cost pathway for radar data augmentation by transforming vision data into radar signals.
This article presents a novel urease-PMMA coated paper based disposable sensor and a cubic support vector machine (CSVM) based sensing for detecting urea in a droplet of 0.3 mL artificial urine samples. The proposed CSVM model takes first 9 principal components (PC) of the phase spectrum of sensor's bio-impedance spectroscopy (BIS) data. The PC counts, the phase feature and the CSVM model are all are determined based on the performance index variation in two 32 model×PC Counts grid matrices. The performance index is a 5- fold cross validated training accuracy for classifying 8 concentration classes (0.5$-$ 4.0g/dL) averaged over 15 iterations. Proposed CSVM-PCA-BIS sensing topology achieves an average training accuracy 97.9%$\pm$0.3% on 1656 unaugmented training data and test accuracy 93.7% on 184 held-out test data. The novelty of the work is to design a sensor which is more accurate, faster responsive, covers full range of human urine and requires less analyte than conventional ones. Its selectivity study shows that it is 45$-$57% more selective than any other major interfering compounds. This establish the proposed work as a proof-of-concept for the future development of an inexpensive portable urea detection meter for urine, blood, farm water and other complex media.
Millimeter-wave (mmWave) radar sensing enables perception in visually occluded environments, yet its spatial resolution is fundamentally limited by the small physical aperture of onboard sensors. Although mmWave synthetic aperture radar (SAR) can improve resolution, existing systems typically rely on mechanical slide rails, which restrict mobility and practical deployment. In this letter, we present a rail-free mmWave SAR system that leverages the natural locomotion of a quadruped robot to synthesize virtual apertures without dedicated mechanical actuation. To address locomotion-induced motion errors and trajectory deviations, we introduce a pose graph optimization-based method that constructs a robot pose graph to estimate accurate radar sampling positions for phase error correction. A learning-based enhancement module is incorporated to improve SAR image quality. Experiments demonstrate the effectiveness and robustness of object-level SAR imaging.
The continuous monitoring of biomarkers via wearable electrochemical sensors is often hindered by the power and form-factor constraints of traditional batteries. Passive Near-Field Communication (NFC) technologies offer a highly viable batteryless alternative; however, integrating complex electroanalytical techniques, such as Cyclic Voltammetry (CV), into ultra-low-power System-on-Chips (SoCs) remains a significant challenge due to the lack of dedicated hardware like Digital-to-Analog Converters (DACs). This letter presents the design and in vitro validation of a high-fidelity, ultra-low-power Analog Front-End (AFE) tailored specifically for the severe architectural constraints of passive NFC SoCs. By emulating the dual-supply architecture and the 2 MHz clock limit of the Texas Instruments RF430FRL153H transponder, a continuous staircase voltage sweep is synthesized via Pulse-Width Modulation (PWM) smoothed by a dual-stage, low-pass RC filter ($f_{c} = 10.24$ Hz). Faradaic currents are acquired using a transimpedance amplifier topology constrained to a 1.5 V digital window. Benchtop validation using a 5 mM potassium ferricyanide redox probe on screen-printed electrodes captured distinct faradaic signatures. The proposed AFE was benchmarked against a commercial laboratory-grade potentiostat, demonstrating comparable electrochemical sensitivity and a superior reversibility ratio ($Q_{a}/Q_{c} = 1.039$)± 0.037, n = 3) closer to theoretical unity than the commercial reference (0.766 ± 0.026). These results confirm the architectural feasibility of executing precise CV under the constraints of passive NFC SoCs, with integration onto the target RF-harvested platform reserved for future work.
In scenarios where global anchoring information is unavailable, how to correctly and robustly inject relative observational information into the collective state estimate constitutes a core challenge in multi-robot cooperative navigation. Conventional nonlinear filtering methods have difficulty preserving the observability consistency of the system, and are prone to producing unreliable estimates when process anomalies and measurement outliers coexist. To address this problem, this paper proposes a robust maximum a posteriori update framework based on the invariant extended Kalman filter. The proposed method constructs an invariant error propagation model on the joint state manifold and formulates relative observations as invariant residuals in the Lie-group error space, thereby preserving the consistency between the true geometric relationships of the system and its unobservable subspace. Meanwhile, within the same invariant tangent space, Huber-type robust constraints are jointly introduced for both the prior prediction term and the measurement observation term, and the resulting problem is solved via iterative reweighted least squares, so as to jointly suppress process mismatch and measurement outliers. This dual synergy between structure preservation and robust suppression enables the proposed method to reduce the influence of compound anomalies on the filtering recursion while maintaining observability consistency, thereby achieving more accurate, robust, and reliable multi-robot cooperative navigation. Simulation results and experiments on the UTIAS MR-CLAM real-world dataset demonstrate that the proposed method achieves favorable localization accuracy and estimation stability in compound anomalous scenarios.
Real-time SONAR image classification on embedded platforms is critical for autonomous underwater monitoring and inspection. However, developing reliable classifiers for such deployments remains challenging due to severe speckle noise, limited annotated training data, and strict computational constraints on edge devices. To address these challenges, we propose a noise-aware training strategy for SONAR image classification that generates additional training samples by jointly exploiting noisy and despeckled image representations. By exposing the classifier to intermediate noise conditions between noisy and despeckled observations, the proposed strategy effectively enlarges the training distribution and promotes stable feature learning across varying noise levels. This formulation improves robustness in data-scarce SONAR imaging scenarios without increasing inference complexity. Experimental results on the 18-class Marine Debris Turntable Dataset improve MobileNet accuracy from 72.5% to 89.8% and SqueezeNet accuracy from 72.4% to 86.4%. The resulting lightweight classification pipeline is deployed on a Raspberry Pi 5 using ONNX Runtime and INT8 quantization, achieving inference latencies as low as 2.29 ms while reducing model size by up to 74.8%. The proposed method further achieves 100% classification accuracy on the 12-class Marine Debris configuration, demonstrating its effectiveness for robust SONAR image classification on resource-constrained edge devices.
Early and automated identification of valvular heart diseases (VHDs) using phonocardiogram (PCG) signals provides a cost-effective solution for developing intelligent healthcare applications. In this letter, a lightweight deep convolutional neural network, LWHSNet, implemented on a field-programmable gate array (FPGA)-based edge computing device, is proposed to identify VHDs via time-frequency domain (TFD) analysis of PCG signals. The TFD representation (TFDR) of the PCG signal is computed using the continuous wavelet transform (CWT). The LWHSNet model consists of 13 layers and is trained using TFD images of the PCG signals. Pruning and fixed-point (FxP) precision-based quantization are used to reduce the size of the LWHSNet model for VHD identification. The FPGA implementation of the proposed TFD-based LWHSNet model is performed using a high-level synthesis (HLS) framework. The performance of the proposed FPGA-based TFD-based LWHSNet model is evaluated using PCG signals from a public database. The proposed LWHSNet model achieves an overall accuracy of 95% with a power consumption of 1.88 watts, a latency of 0.96 seconds, and a throughput of 207 instances per second on the PYNQ-Z2-based FPGA for VHD identification during inference. The proposed embedded healthcare system is well-suited for resource-efficient edge computing applications that utilize PCG signals to identify VHDs.
Grounding systems are critical to the safe and reliable operation of electrical power systems, especially in high-voltage transmission networks, where their performance directly influences fault current dissipation, equipment protection, and overall system reliability and security. To address the need for continuous field assessment of these structures, this work presents a complete measurement system for remote online monitoring of electrical quantities associated with transmission tower grounding systems. The proposed platform combines Rogowski-coil current sensing at the four tower legs, voltage measurements around the tower base, analog signal conditioning, embedded waveform acquisition and processing, and remote communication. The system was implemented in a 525 kV transmission tower and validated by comparison with an oscilloscope and commercial instruments, showing consistent waveform, RMS, and harmonic measurements. Field tests under different counterpoise connection conditions further demonstrated that the measured electrical variables can support diagnostic assessment of the grounding structure. The results indicate that the proposed system provides reliable field acquisition and enables the development of methods for online supervision of transmission tower grounding behavior
This article presents a novel method to detect the change in papaya ripe stages starting from an unripe harvested state. First, papaya bio-impedance spectroscopy (BIS) data through its ripening time-line are collected for up to 10 days over 18 samples and then the data are labelled into four ripe classes based on ethylene emission pattern and various other physical factors. This original training dataset is class-balanced using a typical SMOTE algorithm resulting in total 1232 training data. From a hold-out samples of 4 papayas 69 test data are also obtained. The work investigates three different dimension reduction techniques LDA, PCA and t-SNE, to reduce the order and find the axes most sensitive to papaya ripening. Although both PCA and t-SNE shows comparable accuracy, t-SNE requires less number of components. With first 5 t-SNE components of phase spectrum 95.2% accuracy (5-fold cross validated and averaged over 40 iteration) is achieved with a typical Ensemble Subspace $k$-nearest neighbor (ElSuKNN) model. The paper includes detailed study results, RoC plot and stage wise precision, recall and F1 score while step-by-step developing the proof-of-concept of ML-BIS based sensing mechanism for the ripe class identification of papaya.
Infrared imaging sensors play a crucial role in unmanned aerial vehicle imagery due to their passive imaging capability and robustness under low light conditions. The small pixels of distant targets result in sparse feature representation, which can easily lead to low detection rates. This study proposes a small target detection network, namely IST-RTDETR. A novel attention mechanism namely Frequency Mamba- like Linear Attention (FMLLA) has been designed, which can capture local detailed features in detail and fully model global dependencies at long distances, thereby significantly enhancing the network's ability to extract key information of small targets in complex backgrounds. A dynamic adaptive feature fusion module (DAFF) is proposed to achieve adaptive integration of multi-scale local features and global features, further enhancing the feature expression and fusion effect of the network, and further improving the overall detection performance. The experimental results on the HIT-UAV dataset show that the detection accuracy of IST-RTDETR is superior to mainstream object detection methods. Compared with RT-DETR, IST-RTDETR increased $mAP_{50}$ and $mAP_{s}$ by 1.5% and 2.3%, respectively. This network also achieves an effective balance between computational complexity and detection accuracy.
Garment-integrated wearable sensor devices are becoming essential for daily motion monitoring. The direct ink writing (DIW) technique offers a promising solution by enabling direct printing onto flexible substrates that can be seamlessly integrated into garments. Unlike conventional printing methods that require multiple layers or masking steps, the proposed approach enables rapid, precise dispensing through controlled deposition of conductive inks onto the substrate. This letter presents a TPU-based silver (Ag) ink formulation for DIW fabrication of a flexible capacitive sensor, ensuring strong interfacial bonding with the TPU substrate and enabling seamless integration into garments. The developed sensor demonstrated stable and reproducible capacitance responses under repeated mechanical loads, achieving a sensitivity of 0.343 $\mathrm{kPa^{-1}}$ and pressure sensing up to 20 $kPa$ and a high durability of 5,500 loading/unloading cycles. As a proof of concept, the sensor was heat-transferred onto a textile sleeve for elbow motion detection. It functioned as a pressure sensor by detecting touch through capacitance changes. Tests in two wearable scenarios with five separate attempts, with 20 repetitions each showed consistent, reliable performance (p = 0.357 > 0.05).
To reduce the development cost and hardware consumption associated with repeated Captive Flight Tests (CFTs), reusable infrared (IR) seeker cooling systems have been widely adopted in place of one-shot pyrotechnic cooling devices. To ensure flight-equivalent verification, the reusable cooling system must preserve the certified Guidance Control Unit (GCU) software, ignition sequence, and operational timing of the original pyrotechnic cooling system. Conventional verification methods typically replace one-shot pyrotechnic initiators with reusable solenoid valves; however, continuous valve actuation generally requires modification of the GCU software, increasing software recertification effort and preventing flight-equivalent verification under the certified operational configuration. This paper presents a software-independent verification interface architecture that enables reusable IR seeker cooling without modification to the GCU software. The proposed architecture detects the original pyrotechnic ignition signal through a Detection Stage and Differential Sensing Stage and autonomously converts the transient ignition pulse into a sustained solenoid drive signal using an SCR-based Trigger Stage, thereby reproducing the original pyrotechnic cooling sequence while preserving the certified operational environment. Experimental results demonstrate a hardware activation delay of approximately 12 μs, successful Electromagnetic Compatibility (EMC) evaluation in accordance with mandatory airworthiness certification requirements, and reliable cooling verification using an operational IR seeker. The proposed interface was further validated through more than twenty ground verification procedures performed in support of more than five CFTs, demonstrating its operational reliability and practical applicability. The proposed software-independent verification architecture provides an effective solution for reusable cooling verification while preserving the certified operational configuration of the IR seeker system.
Investigating and categorizing muscle-generated bioelectrical activity specifically EMG recordings associated with extraocular muscles (EOM) is fundamental for building advanced assistive systems.The dynamic and time-varying nature of these physiological waveforms demands analytical techniques capable of capturing temporal dependencies for accurate interpretation and classification. In this study, we propose a compact graph-based representation for EMG of EOM signal classification that transforms one-dimensional temporal waveforms into structured relational graphs. Specifically, each signal instance is segmented into temporal windows that are modeled as graph nodes, with edges encoding both local temporal continuity and feature-level similarity, enabling the learning of non-local dependencies. A Graph Attention Network is employed to adaptively weight inter-node relationships and extract discriminative representations without reliance on frequency-domain decomposition or signal reconstruction, replacing complex domain-specific feature engineering with lightweight statistical descriptors computed directly on raw temporal windows. The proposed approach achieves a peak classification accuracy of 99.17% and mean classification accuracy of 98.33% (95% CI: [96.79%, 99.55%]) across a 5-fold cross-validation with 80-20 split and 98.17% (95% CI: [96.99%, 99.35%]) across a 10-fold cross-validation with 90-10 split, demonstrating competitive performance against strong classical baselines. The results highlight the importance of appropriate graph construction in leveraging attention-based graph models for time-dependent biomedical signal analysis.
This letter presents a flexible, battery-free wireless electrochemical pH sensing platform toward intraoral health monitoring. The device integrates a Nafion-functionalized Au sensing electrode, a solid-state Ag/AgCl reference electrode, and a miniaturized LC resonator on a polyimide substrate. Variations in pH are transduced into open-circuit potential (OCP) signals, which modulate the resonance frequency via a back-to-back varactor pair. Experimental results show a monotonic response over pH 4–10 with a sensitivity of -47.8 mV/pH. The corresponding wireless readout exhibits a resonance shift from 339.50 MHz to 333.00 MHz with a average quality factor of about 42. The consistent relationship between OCP and resonance frequency validates the electrochemical-to-electromagnetic transduction mechanism. The proposed platform offers a compact, passive, and flexible proof-of-concept solution for minimally invasive oral pH monitoring, providing a foundation for future long-term intraoral applications.
In visual scene matching navigation, environmental differences, seasonal variations, and noise can lead to numerous inaccurate matching results, negatively impacting final navigation accuracy. To address the heterogeneous image matching problem, we propose a scene matching and localization algorithm based on an improved ViT model. For feature extraction, we employ the ViT model as the global feature extraction framework and introduce K-means clustering to aggregate local features. Combined with the global features extracted by the ViT main framework, we generate a robust local-global feature representation vector. For feature matching, we use incremental principal component analysis (IPCA) to reduce the dimensionality of the high-dimensional feature space and construct a KD-tree structure for fast feature retrieval, improving matching efficiency. We validate our algorithm on the University-1652 UAV multi-view geolocation dataset and real-world satellite-aerial image datasets. Results show that our method outperforms other models in both Recall and AP metrics. We also conducted flight experiments in real-world scenarios. The results show that, at a flight altitude of 350 meters, our algorithm achieves an average absolute value of 6.21m for latitude, 6.76m for longitude, and 10.20m for horizontal error. Therefore, our algorithm demonstrates ideal overall positioning accuracy.
