Export
Puyuan Hu
· 1 previous affiliation
- Shanghai Jiao Tong UniversitySince 2021
- PhD student · 2025–2030
- Undergrad student · 2021–2025
Research Overview
Puyuan Hu focuses on cross-modal localization for autonomous systems and open-set fault diagnosis in complex industrial processes. Recent collaborative work explores the integration of LiDAR and monocular camera data for mobile robotics, alongside causal counterfactual methods for identifying unknown faults in industrial sensor networks.
Research Trajectory
The available evidence for Puyuan Hu's research trajectory spans from 2024 to 2025, indicating a focus on applied machine learning and sensor data processing across two distinct domains: autonomous navigation and industrial fault diagnosis. Because the corpus of work is relatively sparse, this trajectory reflects participation in specific collaborative projects rather than a decades-long, independent research program.
In 2024, Hu participated as a middle author in research addressing the challenges of cross-modal localization for autonomous driving and mobile robotics. This work introduced LHMap-loc: Cross-Modal Monocular Localization Using LiDAR Point Cloud Heat Map. The project tackled the difficulties of map storage and localization robustness in large scenes by proposing a pipeline that compresses original LiDAR point cloud maps into offline heat point clouds. This allows for more efficient and accurate localization using a monocular camera within a pre-built LiDAR map. As a third author on this paper, Hu contributed to a broader team effort to develop and evaluate this cross-modal feature encoding pipeline.
Moving into 2025, Hu's collaborative focus shifted toward industrial applications, specifically addressing the reliability of diagnostic models in open environments. Serving as the second author, Hu made significant contributions to Causal Counterfactual Faithfulness Generation for Open-Set Fault Diagnosis of Complex Industrial Processes. Traditional intelligent fault diagnosis models often struggle with unknown fault types, frequently misclassifying them into known categories because training samples rarely cover all possible fault conditions. To mitigate this, the research team developed a causal counterfactual faithfulness generation method. The approach processes signal data from fault sensors into graph structures—analyzing nodes and edges to better isolate and identify unknown fault categories in complex industrial processes.
Current Directions
Based on the most recent publications in the evidence, Hu's current research directions center on improving the robustness of machine learning models in open-world and cross-modal settings. The 2025 work on open-set fault diagnosis highlights an active interest in causal inference and graph-based data representations to handle unknown variables and out-of-distribution samples in complex industrial systems. Concurrently, the earlier focus on cross-modal localization demonstrates an engagement with sensor fusion and data compression techniques designed to make autonomous robotics more efficient in real-world environments.
Profile updated
| Title | Venue | Year |
|---|---|---|
| Causal Counterfactual Faithfulness Generation for Open-Set Fault Diagnosis of Complex Industrial Processes | IEEE Trans. Industrial Informatics | 2025 |
| LHMap-loc: Cross-Modal Monocular Localization Using LiDAR Point Cloud Heat MapGitHub resource | arXiv | 2024 |
Loading new papers…
Loading research neighbors…
See something wrong? Report a mistake