Hao Zhu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Hao Zhu has a total of 11 co-authors. Here are the top co-authors by number of shared publications:
- Daguang Han
School of Civil Engineering, Southeast University, Nanjing 211189, China
3 shared publications - Yulong Chen
School of Civil Engineering, Chongqing Jiaotong University, Chongqing 400074, China
2 shared publications - Chunli Ying
School of Architecture, Building and Civil Engineering, Loughborough University, Epinal Way, Loughborough LE11 3TU, UK
1 shared publication - Huanhuan Lai
State Grid Wenzhou Electric Power Supply Company, Wenzhou 325000, China
1 shared publication - Jun Chen
School of Civil Engineering, Chongqing Jiaotong University, Chongqing 400074, China
1 shared publication - Liyuan Chen
Pingyang County Power Supply Company, State Grid Zhejiang Electric Power Co., Ltd., Wenzhou 325400, China
1 shared publication
Latest Publications
- Open Access
- 21 October 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (20) , 10430
Near-Duplicate Contamination in a Widely Used Safety Helmet Detection Benchmark: Audit, Controlled Experiments, and a Duplicate-Aware Evaluation Split
Public benchmarks drive research on vision-based safety monitoring, and the Safety Helmet Wearing Dataset (SHWD) has become the de facto standard for helmet-wearing detection. This paper reports a systematic near-duplicate audit of SHWD and measures how the contamination it uncovers affects the accuracy reported on the benchmark. A two-stage matching procedure—64-bit perceptual hashing with complementary thumbnail-correlation verification—shows that 2288 of 7581 images (30.2%) belong to 193 near-duplicate clusters, the largest containing 802 frames from one fixed camera, and that 421 of the 1517 official test images (27.8%; 22–29% across matching thresholds) have a near-duplicate counterpart in the training split, and a blind manual rating of 350 pairs finds that 100% of accepted pairs are frames of the same fixed camera (28% strict near-duplicates). Controlled experiments with YOLOv8n, three seeds each, quantify the consequence. With the test data fixed, removing the 1534 contaminating training images lowers mAP@0.5 on the contaminated test subset by 4.4 points, and a size-matched random control shows that the difference is not a data-volume effect; on the full official test set the inflation is 1.92 points relative to the deduplicated model and 1.35 points relative to the control. A cluster-disjoint experiment that holds the domain composition of the training set fixed shows that frames from the same cameras and sessions as the test frames add a further 0.6 points on those frames, so that most of the gap under the official split reflects the classroom imagery that the contaminating images supply rather than recognition of specific frames. Tracing images to their source shows that 97.1% of the test images imported from the SCUT-HEAD classroom-video dataset are contaminated, against 3.0–5.5% of the genuine helmet imagery, and that the effect is confined to the imported subset (+8.1 points there, −0.4 on helmet imagery). The inflation is reproduced with YOLOv8s (+2.1 points on the contaminated subset) and with RT-DETR-l (+2.8 points). Removing the contaminating training images—25.3% of the training set—leaves accuracy on the clean test subset unchanged within seed variability (−0.14 points). The duplicate lists, a duplicate-aware evaluation split, and the audit code are publicly available.
- 21 October 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (20) , 10430
Open AccessNear-Duplicate Contamination in a Widely Used Safety Helmet Detection Benchmark: Audit, Controlled Experiments, and a Duplicate-Aware Evaluation Split
Public benchmarks drive research on vision-based safety monitoring, and the Safety Helmet Wearing Dataset (SHWD) has become the de facto standard for helmet-wearing detection. This paper reports a systematic near-duplicate audit of SHWD and measures how the contamination it uncovers affects the accuracy reported on the benchmark. A two-stage matching procedure—64-bit perceptual hashing with complementary thumbnail-correlation verification—shows that 2288 of 7581 images (30.2%) belong to 193 near-duplicate clusters, the largest containing 802 frames from one fixed camera, and that 421 of the 1517 official test images (27.8%; 22–29% across matching thresholds) have a near-duplicate counterpart in the training split, and a blind manual rating of 350 pairs finds that 100% of accepted pairs are frames of the same fixed camera (28% strict near-duplicates). Controlled experiments with YOLOv8n, three seeds each, quantify the consequence. With the test data fixed, removing the 1534 contaminating training images lowers mAP@0.5 on the contaminated test subset by 4.4 points, and a size-matched random control shows that the difference is not a data-volume effect; on the full official test set the inflation is 1.92 points relative to the deduplicated model and 1.35 points relative to the control. A cluster-disjoint experiment that holds the domain composition of the training set fixed shows that frames from the same cameras and sessions as the test frames add a further 0.6 points on those frames, so that most of the gap under the official split reflects the classroom imagery that the contaminating images supply rather than recognition of specific frames. Tracing images to their source shows that 97.1% of the test images imported from the SCUT-HEAD classroom-video dataset are contaminated, against 3.0–5.5% of the genuine helmet imagery, and that the effect is confined to the imported subset (+8.1 points there, −0.4 on helmet imagery). The inflation is reproduced with YOLOv8s (+2.1 points on the contaminated subset) and with RT-DETR-l (+2.8 points). Removing the contaminating training images—25.3% of the training set—leaves accuracy on the clean test subset unchanged within seed variability (−0.14 points). The duplicate lists, a duplicate-aware evaluation split, and the audit code are publicly available.
- Open Access
- 22 September 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (19) , 9420
An Auditable Gridded Wind–DEM-to-FE Workflow with Source Time, Spatial Contract, Texture Axis, and Solver Balance Gates: A Transmission Tower Case Study
When gridded meteorological data drive finite element (FE) models, the interfaces carrying product identity, coordinates, axis order, time, and units are rarely checked. We built an auditable workflow linking an archived ECMWF gridded wind product, DEM metadata, Unity sampling, node exports, and ANSYS 2025 R2 comparisons, treating each interface as a gate that can fail. Three gates failed. The archived GRIB holds 48 monthly averages by hour of day (stream = mnth), not a continuous 48-h event; the runtime decoded magnitude from a constantly zero channel; and although the inline WGS84-to-UTM formula matches PROJ to 2.41 × 10−5 m, the wider mapping never closed, with none of the 234 towers inside the hard-coded wind subset and two Texture3D axes transposed. We then repaired what we could. Re-cropping the source to the line corridor and declaring the axis and vertical contracts moved tower containment from 0/234 to 234/234 and reproduced the source field to 0.0103 m/s against a 0.0307 m/s quantization step. Repeating the A–D solves with everything else fixed raised group A load from 292.050 to 355.034 N, displacement by 20.9%, and stress by 19.7%. The source time gate still fails, so results remain representative slot computations, not event responses. A synthetic control passes all six gates.
- 22 September 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (19) , 9420
Open AccessAn Auditable Gridded Wind–DEM-to-FE Workflow with Source Time, Spatial Contract, Texture Axis, and Solver Balance Gates: A Transmission Tower Case Study
When gridded meteorological data drive finite element (FE) models, the interfaces carrying product identity, coordinates, axis order, time, and units are rarely checked. We built an auditable workflow linking an archived ECMWF gridded wind product, DEM metadata, Unity sampling, node exports, and ANSYS 2025 R2 comparisons, treating each interface as a gate that can fail. Three gates failed. The archived GRIB holds 48 monthly averages by hour of day (stream = mnth), not a continuous 48-h event; the runtime decoded magnitude from a constantly zero channel; and although the inline WGS84-to-UTM formula matches PROJ to 2.41 × 10−5 m, the wider mapping never closed, with none of the 234 towers inside the hard-coded wind subset and two Texture3D axes transposed. We then repaired what we could. Re-cropping the source to the line corridor and declaring the axis and vertical contracts moved tower containment from 0/234 to 234/234 and reproduced the source field to 0.0103 m/s against a 0.0307 m/s quantization step. Repeating the A–D solves with everything else fixed raised group A load from 292.050 to 355.034 N, displacement by 20.9%, and stress by 19.7%. The source time gate still fails, so results remain representative slot computations, not event responses. A synthetic control passes all six gates.
- Open Access
- 28 August 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (17) , 8582
End-to-End Automated Wind-Induced Stress Simulation of Lattice Transmission Towers in Complex Terrain via Physics-Conserving PINN Wind-Field Reconstruction and Graph-Theory-Based DXF Parsing
Assessing whether existing lattice towers can survive extreme wind when they are located on ridgelines or at saddle points requires three questions to be answered simultaneously: how the local wind field is modified by the surrounding topography, how the structural geometry recorded in legacy computer-aided design (CAD) drawings can be recovered accurately, and which member fails first and by what mechanism. This paper couples a Physics-Informed Neural Network (PINN) wind solver, jointly constrained by mass and momentum conservation, with a graph-theory-based Drawing Exchange Format (DXF) parser and a closed-loop vulnerability screening module, so that all three questions are answered in a single automated pass. The PINN reconstructs the three-dimensional steady-state wind field over irregular topography in approximately 0.12 s, holding the root-mean-square (RMS) velocity divergence below 2.1 × 10−3 (more than two orders of magnitude lower than that of linear interpolation) while recovering the pressure-gradient-driven acceleration that mass-consistent variational solvers cannot represent. On the CAD side, k-dimensional tree (KD-Tree) spatial indexing combined with breadth-first search (BFS) connected-component analysis resolves the pseudo-disconnections, spurious intersections, and multi-level nested block references that are common in production DXF files, achieving 100% node-merging accuracy across fifteen tower drawings. A unified Vulnerability Index (VI) that combines strength, member stability, and plate buckling into a single scalar, updated through Sherman–Morrison rank-one perturbation at a millisecond cost, closes the diagnose–strengthen–verify loop without re-solving the full stiffness system. Applied to a 220 kV line struck by Super Typhoon Meranti, the pipeline identified seven at-risk members that code-based checking had missed, a result consistent with the recorded field damage, and completed the full assessment in 34.4 s, over three orders of magnitude faster than conventional practice.
- 28 August 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (17) , 8582
Open AccessEnd-to-End Automated Wind-Induced Stress Simulation of Lattice Transmission Towers in Complex Terrain via Physics-Conserving PINN Wind-Field Reconstruction and Graph-Theory-Based DXF Parsing
Assessing whether existing lattice towers can survive extreme wind when they are located on ridgelines or at saddle points requires three questions to be answered simultaneously: how the local wind field is modified by the surrounding topography, how the structural geometry recorded in legacy computer-aided design (CAD) drawings can be recovered accurately, and which member fails first and by what mechanism. This paper couples a Physics-Informed Neural Network (PINN) wind solver, jointly constrained by mass and momentum conservation, with a graph-theory-based Drawing Exchange Format (DXF) parser and a closed-loop vulnerability screening module, so that all three questions are answered in a single automated pass. The PINN reconstructs the three-dimensional steady-state wind field over irregular topography in approximately 0.12 s, holding the root-mean-square (RMS) velocity divergence below 2.1 × 10−3 (more than two orders of magnitude lower than that of linear interpolation) while recovering the pressure-gradient-driven acceleration that mass-consistent variational solvers cannot represent. On the CAD side, k-dimensional tree (KD-Tree) spatial indexing combined with breadth-first search (BFS) connected-component analysis resolves the pseudo-disconnections, spurious intersections, and multi-level nested block references that are common in production DXF files, achieving 100% node-merging accuracy across fifteen tower drawings. A unified Vulnerability Index (VI) that combines strength, member stability, and plate buckling into a single scalar, updated through Sherman–Morrison rank-one perturbation at a millisecond cost, closes the diagnose–strengthen–verify loop without re-solving the full stiffness system. Applied to a 220 kV line struck by Super Typhoon Meranti, the pipeline identified seven at-risk members that code-based checking had missed, a result consistent with the recorded field damage, and completed the full assessment in 34.4 s, over three orders of magnitude faster than conventional practice.