Qianyu Zhou, Ph.D. - University of Connecticut, 2025

Semi-Supervised Pseudo-Labeling with Adaptive Weighting

This study introduces Semi-Supervised Pseudo-Labeling with Adaptive Weighting, a self-evolving learning mechanism that fuses limited labeled and abundant unlabeled data while adaptively regulating pseudo-label confidence. This enables robust convergence with reduced simulation and annotation cost. The framework achieved comparable performance to fully supervised learning but slashed labeling efforts by 72% (saving approximately 487.5 hours in label generation time), demonstrating significant efficiency gains.

Parallel Physics-Informed Representation Learning

This research establishes a Parallel Physics-Informed Representation Learning architecture for vibration-based gearbox fault diagnosis, where signal morphology is encoded through the fusion of temporal and spectral structural priors. The framework integrates domain knowledge and spatially correlated features within a unified network to enhance interpretability and suppress false negatives, establishing a reliable pathway for small-data condition monitoring. This design demonstrates how physical insight can be embedded directly into the representational hierarchy of deep learning models.

This approach enhances diagnostic distinctions even in early-stage fault regimes, allowing for robust and interpretable condition monitoring with limited data. By combining physical principles with deep learning, it provides a powerful solution for improving trustworthiness in safety-critical applications.

Spatially Informed Graph-Neural Surrogate Modeling

This study introduces a graph-based deformation prediction framework for surface quality monitoring in milling operations, wherein structural geometry and spatial correlation among nodes are encoded via dynamic attention modulation. The method integrates physics-derived relationships between local deformation influences and global structural responses, allowing the predictive model to adaptively emphasize salient geometric interactions across machining configurations and process conditions. This ensures efficient and accurate modeling of complex manufacturing processes under constrained computational budgets.

The framework uses a concise, dynamically evolving node graph composed exclusively of critical measurement points, ensuring an effective yet accurate surrogate model. By embedding physics-informed constraints, it accurately predicts post-milling deformation aligned with experimental protocols, supporting rapid surface quality prediction within digital twin environments.

Edge–Cloud Collaborative Compression and Reconstruction Framework

This framework provides a scalable approach for real-time signal analytics, balancing compression efficiency, information fidelity, and system throughput for distributed industrial computing environments. It implements content-aware token selection (sampling) and structure-aware, non-uniform quantization, achieving bit-budgeted signal compression up to 64:1. A physics-guided Transformer regularizes reconstruction to preserve spectral energy, phase relations, and envelope trends. This concurrent and multiplexed approach maintains or improves fault-classification accuracy and RUL prediction relative to representative baselines.

This system delivers edge-viable representations that preserve fault-relevant content under strict compute and bandwidth budgets, crucial for real-time PHM applications. It ensures diagnostic fidelity without sacrificing efficiency, enabling robust performance even under severe compression.

Zero-Shot Vision–Language Reasoning via Retrieval-Augmented Generation

This study extends few-shot visual intelligence to unseen manufacturing scenarios, coupling multimodal embeddings with knowledge retrieval to achieve generalizable understanding across modalities. Our novel visual-text RAG-VLM framework demonstrates the ability to reason about damage type and severity in a flexible, interpretable, and data-efficient manner on wind turbine blade inspection. It achieved up to 98.33% classification accuracy with a minimal number of training samples, significantly outperforming traditional image-based classifiers.

Unlike conventional pipelines, our system does not require task-specific fine-tuning. Instead, it performs in-context few-shot reasoning by retrieving relevant visual-textual examples from a structured knowledge base, enabling robust, interpretable, and scalable damage classification.