
Biological tissues exhibit frequency-dependent nonlinear shear wave dispersion arising from viscoelasticity. This study evaluates an empirically modified power-law dispersion model that incorporates frequency-dependent gradient and concavity in an in vivo animal model to better capture nonlinear dispersive behavior in shear wave dispersion imaging. In vivo imaging was conducted on a fatty duck liver model at four time points during steatosis progression. The nonlinear dispersion gradient and concavity coefficient exhibited an overall increase over the gavage period, with mean values rising from 4.28 to peaking at 12.05 m/s/100 Hz and from 1.09 to 4.47 m/100 Hz, respectively. The observed p-values (< 0.001) between the early pre-feeding and lipid accumulation stages demonstrate that these markers can significantly distinguish between normal and fatty tissue states. These results support in vivo validation of the model and highlight its potential utility for characterizing nonlinear, lossy tissue behavior.
Accurate time synchronization is a fundamental requirement for networked control systems, among which Time-Sensitive Networking (TSN)-enabled Unmanned Ground Vehicles (UGVs), whose timestamped control strategies rely on synchronized clocks to compensate for communication delay and jitter. This paper investigates how clock synchronization uncertainty propagates through a timestamped pursuit controller operating in position open-loop conditions, where tracking accuracy depends entirely on dead-reckoning without corrective feedback. An analytical uncertainty propagation framework based on the Guide to the Expression of Uncertainty in Measurement (GUM) and Jacobian linearization is developed to quantify how synchronization uncertainty contributes to the measurement uncertainty of vehicle pose estimation. The analytical model is complemented by a Monte Carlo simulation that captures the nonlinear trajectory-dependent behavior and provides empirical error statistics. The analytical framework captures the orientation-to-position uncertainty coupling and offers physical insight into the dominant error mechanisms, whereas the Monte Carlo model provides realistic empirical error distributions. Simulations covering synchronization accuracies ranging from sub-microsecond IEEE 802.1AS/gPTP performance to hundreds of microseconds with software timestamping demonstrate that clock synchronization uncertainty can become the dominant contributor to the pose estimation uncertainty and overall measurement uncertainty budget. The proposed framework provides a quantitative methodology for evaluating synchronization-induced measurement uncertainty in networked robotic systems, while offering practical design guidelines for industrial cyber-physical systems requiring reliable time synchronization.
This work presents an integrated plantar pressure measurement system and interpretable deep learning framework for plantar pressure analysis. Even with the advances in plantar pressure sensing, the integration of scalable sensing architectures with interpretable learning models remains limited. A custom high density plantar pressure measurement system based on a multiplexed 32×32 force sensitive resistor (FSR) array is developed to acquire spatially distributed load measurements from 1024 sensing elements using a reduced number of analog channels. The pressure data are separated into independent left and right foot representations and interpolated to 64 × 32 maps to enhance spatial continuity prior to classification. A convolutional neural network is trained for each foot to discriminate between rectus and valgus alignment patterns. To improve interpretability, Gradient weighted Class Activation Mapping (Grad-CAM) is employed to identify the sensing regions that contribute strongly to the classification decisions. Experimental results show consistent and balanced performance across both feet, while intensity maps highlight biomechanically meaningful regions such as the forefoot and longitudinal arch. The proposed approach combines scalable plantar sensing and explainable artificial intelligence, supporting objective and clinically meaningful assessment of foot posture.
Casting micropores in aluminum alloy transmission spacer bars expand under alternating wind-induced vibration and eventually evolve into fatigue cracks, posing severe safety risks to overhead power grids. Portable X-ray nondestructive testing (NDT) enables on-site flaw inspection, yet practical deployment faces three bottlenecks: insufficient real defective radiographs, low micro-pore contrast under low-dose radiation, and high labor cost of pixel-level manual annotation. To address these limitations, this paper proposes a weakly supervised detection framework integrated with X-ray physical prior pseudo-defect generation. First, three physical priors (luminance attenuation, edge gradient distribution, elliptical projection morphology) are established based on the Lambert–Beer X-ray attenuation law. A total of 10 000 physically consistent pseudo-defect samples with pixel masks are automatically synthesized using 500 defect-free spacer radiographs; Kolmogorov–Smirnov test and gradient cosine similarity metrics verify that synthetic flaws match the statistical distribution of real casting pores. Second, a lightweight detection backbone based on MobileNetV3 is designed, combined with a scale-adaptive parameter-free SimAM attention module enhanced by imaging physical constraints. A composite loss function containing classification loss, spatial consistency loss and physical prior regularization loss is constructed, matched with a two-stage training strategy strictly separating full-supervised pre-training (only pseudo-data) and image-level weakly supervised fine-tuning (only real defective images without pixel labels). Experiments are conducted on an in-house dataset consisting of 500 normal samples, 45 binary-labeled defective samples and 40 fully annotated test radiographs. Ablation studies, fair SOTA comparisons with equal synthetic data, anti-interference robustness tests and standardized metrological evaluation are carried out. Quantitative results show that the proposed method achieves 87.3% mAP@0.5 and 84.6% recall, surpassing mainstream weakly supervised detection 7.6 percentage points in mAP and 9.2 percentage points in recall. Metrological analysis confirms the minimum detectable pore size reaches 0.6 mm with average relative sizing error lower than 4.3%. Deployed on NVIDIA Jetson Xavier NX, the model only consumes 87 ms per frame with 1.28 M parameters, satisfying real-time detection requirements of portable field terminals. The proposed framework provides a low-annotation, interpretable quantitative measurement solution for power fitting X-ray NDT.
Measuring elbow movement can support biomechanics research and rehabilitation assessment, but laboratory motion-capture systems are less portable than wearable sensors. To examine how these measurements compare, we provide optical trajectories, wireless inertial-sensor recordings, and videos from one participant during a single recording session. Four pointing trials were recorded; Trial 8 was excluded after observed sensor movement and Trial 11 because no verified IMU–BTS temporal pairing was recorded. We estimated an angle correction from Trial 13 and applied it unchanged to four stationary periods in Trial 14. Compared with the optical reference, the test mean absolute error was 1.92° and root mean square error was 1.98° over reference angles of 93.63°–99.76°. The test recording was aligned in time using its optical trajectory; varying the plausible alignment gave root mean square errors of 0.77°–3.50°. These results describe agreement for this participant and sensor placement during stationary holds, rather than accuracy during movement or across people and sessions. The shared recordings enable independent reanalysis.