Abstract Detecting live underground power cables in complex subsurface environments with ground-penetrating radar (GPR) is hindered by scarce labels, simulation–reality gaps, and weak deep echoes. We present ground-penetrating physics-aware modeling YOLO (GPM-YOLO), a detection framework coupling realistic simulation with physics-aware and depth-sequence fusion. First, we build a realistic GPR dataset by modeling non-uniform layered soils, plastic/concrete conduits, and stones, and by injecting power-frequency excitation into the cable core to emulate “live-cable” signatures, yielding systematic B-scans. Second, a physics feature module separates and adaptively fuses two cues: high-frequency strong reflections from the metal core and low-frequency oscillatory textures from the insulation. Third, a depth-sequence fusion module treats the B-scan vertical axis as an intrinsic temporal sequence and applies interval modeling with attention to enhance the continuity of deep weak signals and suppress false alarms. With db4 wavelet denoising, normalization, and a three-level annotation protocol (boxes, physics masks, depth evolution), GPM-YOLO achieves higher recall for deep cables and stronger robustness than vanilla YOLO on our dataset. The method unifies physical interpretability and data-driven learning for fast, accurate live-cable detection in urban settings.
This study proposes a method for predicting the static calibration (SC) matrix of strain gauge balances during the development stage. Non-linear finite element analysis (FEA) is employed to simulate the complex deformation behavior of a six-component strain gauge balance under both single- and multi-component loads. By correlating the simulated strains with the applied loads, the virtual calibration (VC) matrix is obtained during the design process. The validity of the VC matrix is subsequently corroborated through comparison with the experimentally derived SC matrix. This research enriches balance design methodology, improves the first-attempt success rate of balance design, and enhances both the efficiency and cost-effectiveness of balance development.
Dimensional comparison of micro-computed tomography (micro-CT) reconstructions with CAD geometry requires a precise definition of the evaluated surface points. We apply a statistical and metrological framework to signed CAD–CT deviations at the vertices of three geometrically different titanium dental implant components, D1–D3, representing complex surface geometry, substantial μCT-to-CAD deviations, and ambiguous initial alignment. Complete exports contain 75 812, 46 954, and 16 416 measured-surface vertices. The corresponding RMS deviations are 49.14 μm, 207.72 μm, and 77.07 μm. Empirical cumulative distribution functions, histograms normalized to unit area, and descriptive summaries compare the observed meshes; two-sample Kolmogorov–Smirnov (KS) distances quantify differences between their vertex distributions. D2 has the largest absolute-deviation upper tail. A detailed axial analysis of D1 localizes all 267 vertices with absolute deviation above 200 μm in one of 12 intervals. Separate area-sampled and radial D1 exports provide sensitivity checks for alternative surface and correspondence definitions. Because vertices are spatially dependent and mesh density determines their weighting, conventional KS p-values are only nominal reference calculations. A prospective observation model distinguishes physical deviation from measurement and alignment effects. The workflow provides a reproducible basis for statistical evaluation and measurement-uncertainty assessment of individual components.
Train on-board systems estimate speed using data from multiple sensors, each of which may fail under specific operating conditions. Accurate speed estimation is critical for railway safety. This study considers a hypothetical 60-second braking scenario. Speed is measured by five sensors with different characteristics and measurement models. Sensor data are fused using three methods: mean, median, and weighted. The weights are proportional to the inverse square of the measurement uncertainty. These uncertainties include both random and systematic components. Unlike many existing studies, the proposed approach incorporates systematic effects instead of relying solely on variances. The results show that the mean and median methods are simple to implement but do not provide the most accurate estimates. The mean is highly sensitive to outliers and performs poorly when a sensor malfunctions, for example, during wheel slide. The weighted method accounts for both random and systematic effects and provides the closest agreement with the reference value, although it requires analysis of the underlying measurement models.
A cross-condition fault diagnosis method is developed for multi-channel vibration measurements. For every 0.5-s segment, Short-Time Fourier Transform (STFT) spectrograms and Fast Fourier Transform (FFT) spectral sequences are generated from the same four-channel observation. A convolutional neural network and a temporal convolutional network encode local time–frequency textures and global spectral dependencies, while a sample-adaptive gate fuses the two representations. Multiple Kernel Maximum Mean Discrepancy (MK-MMD) is then applied only in the common fused space to reduce load-induced distribution shift without target labels, pseudo-labels, or adversarial optimization. Experiments cover all three many-to-one transfer directions among 0, 2, and 4 Nm. The proposed model obtains accuracies of 87.62 %, 85.71 %, and 70.59 %, and macro-F1 scores of 89.24 %, 76.00 %, and 72.31 % for T1–T3, respectively. Against four external domain-adaptation baselines, it achieves the highest task-averaged accuracy (81.31 %) and macro-F1 (79.18 %). Five controlled ablations and class-level visualizations further show that fused-space alignment is the most consistent contributor, while the normal–unbalance ambiguity remains the main limitation under the more difficult transfer directions.
