Research

PHM monitors, evaluates, and predicts the health of mechanical, electrical, and industrial systems — so faults are caught early, reliability goes up, and maintenance happens before something breaks. It was the core of my Ph.D., spanning motors, rotating machinery, and thermal power plants.

• Why it matters — predictive failure detection (cost savings), higher reliability, better operational efficiency.
• How I approach it — signal-aware preprocessing, deep representation learning, anomaly detection, and diagnosis that holds up under changing operating conditions.

Rotating machines & motors

1. SCOC: Stator Current Operation Compensation for PMSM fault diagnosis

   Measurement ’23First author

   A preprocessing step that compensates a motor’s stator-current signals for changing speed and load, so a deep model can diagnose permanent-magnet synchronous motor (PMSM) faults reliably under variable operating conditions.

2. MPARN: Multi-scale Path Attention Residual Network

   JCDE ’23First author

   A residual network with multi-scale paths and attention that captures fault signatures across different time scales, improving fault diagnosis of rotating machines.

3. Center-margin loss: uncertainty-aware diagnosis of unseen faults

   JCDE ’26First author

   Adds uncertainty estimation and a center-margin loss so the model can flag novel, never-before-seen fault types instead of forcing them into known classes.

4. Adapter-enhanced Fourier-feature DeepONet for fault-severity estimation

   Eng. Appl. of AI ’26

   A deep operator network with Fourier features and adapters that estimates the severity of stator inter-turn short circuits in induction motors — moving from “which fault” to “how bad”, including severities not seen during training (simulation-assisted).

5. Self-supervised feature learning under various torque conditions

   Knowledge-Based Systems ’24

   Self-supervised pre-training learns motor-fault representations without labels, so diagnosis stays accurate across a wide range of torque conditions with little labeled data.

6. FL-SSDAN: fleet-level semi-supervised domain adaptation

   JCDE ’25Co-first author

   Transfers fault knowledge across many overhead-hoist-transport units in a semiconductor fab using mostly unlabeled data, so a model trained on a few machines generalizes to the whole fleet.

7. Frequency-band graph-based sensor fusion

   Eng. Appl. of AI ’26

   Fuses multiple sensors on a frequency-band graph with a sensitivity-aware “energy assist” network, sharpening fault diagnosis of complex machinery systems.

Thermal power plant

1. Opt-TCAE: boiler-tube leakage detection

   ESWA ’23First author

   An unsupervised temporal convolutional auto-encoder that flags boiler-tube leaks from multi-sensor signals; the latent size is tuned by entropy minimization and the alarm threshold is set automatically via kernel density estimation.

End-to-end deep learning on engineering data often ignores what engineers already know. DINN folds that domain knowledge — physics, signal analysis, expert rules — back into the model.

• Problems it fixes — low generalization (sensitive to noise/conditions), black-box opacity, and physically inconsistent outputs.
• How — training across simulated multi-condition data (generalization), physics-based loss terms / PINNs (physical consistency), and signal-aware inputs such as 2D time–frequency maps (interpretability).

1. A review of PHM methods based on domain-knowledge-informed neural networks

   in progressFirst author

   Defines DINN, organizes the ways physics, signal analysis, and expert rules can be fused into PHM models, analyzes their limits, and maps out open directions.

Using AI across the design loop — generating candidate designs, predicting performance faster than simulation, and optimizing parameters and topology in high-dimensional spaces.

• Design generation — generative models for new design candidates.
• Design evaluation — prediction models, physics-informed networks (PINNs), and explainable AI.
• Design optimization — parameter & topology optimization via reinforcement learning and inverse modeling.

1. Inverse modeling of springback in metal bipolar-plate stamping

   under reviewFirst authorw/ Fraunhofer IPT

   A neural inverse model that predicts the die/stamping parameters needed to hit a target metal bipolar-plate geometry for PEM fuel cells — compensating springback in seconds instead of through iterative simulation. A collaboration with Fraunhofer IPT (Aachen), built on adaptive sampling + a surrogate model + an inverse model.

2. Reinforcement-learning-based optimal parameter estimation of an ultra-precision machine

   ICEAS ’22First author

   Frames process-parameter tuning of an ultra-precision machine as a reinforcement-learning problem to reach optimal settings efficiently.

My current focus at Samsung Research: domain-specialized LLMs, knowledge graphs, and retrieval-augmented generation (RAG) that bring engineering data and physical reasoning into one workflow.

• Where LLMs help — design generation, end-to-end process automation, and integrating human/physics feedback.
• Building a domain LLM — (1) curate engineering datasets for fine-tuning & RAG, (2) handle multi-modal data (CAD, simulation, sensor signals), (3) deploy in workflows that stay aligned with physical principles.

1. LLM-based engineering design framework

   in development

   An LLM workflow that generates designs and folds in feedback to fit user requirements while respecting physical constraints — bridging today’s fragmented, rigid, and slow design-tool chains.

2. Knowledge-graph + RAG + LLM for PHM

   research

   Represents fault-related knowledge as a graph and keeps a PHM assistant accurate and current by combining RAG with LLMs — tackling the black-box reasoning, slow retrieval, and limited access to unstructured maintenance data that plain models struggle with.