Application Case – Energy Forecasting and Management 

Microgrids integrate diverse energy sources, including solar power, wind energy, battery energy storage systems, and diesel generators, and must perform real-time and effective energy dispatch under significant uncertainty. By employing load and renewable energy forecasting models, such as time-series analysis and deep learning techniques, and combining them with optimization or reinforcement learning–based decision strategies, microgrids can determine optimal charging and discharging schedules for energy storage, on/off decisions for diesel generators, and overall energy allocation policies. These data-driven and decision-oriented approaches enable microgrids to reduce fuel consumption and carbon emissions, increase energy self-sufficiency, and maintain stable and reliable operation under highly volatile and uncertain operating conditions.

Representative Publication:
Y.-C. Hsu, Y.-H. Hung, and C.-Y. Lee, “Robust ensemble forecasting and deep reinforcement learning for energy management on islanded microgrids,” International Journal of Electrical Power and Energy Systems, vol. 173, p. 111405, 2025. [PDF]

Application Case – Semi-Automated Production Scheduling 

In highly customized manufacturing environments with significant demand volatility, fully automated scheduling models often fail to meet the flexibility required on the shop floor. As a result, semi-automated production scheduling systems have emerged as a more practical and effective solution. By integrating optimization algorithms, heuristic rules, and human–machine collaborative interfaces, such systems can rapidly generate feasible schedules while supporting subsequent manual refinement.

Based on order characteristics, machine capabilities, and due-date constraints, the system automatically produces multiple candidate scheduling solutions and provides critical decision-support information, including bottleneck machine utilization, estimated completion times, and potential scheduling conflicts. This enables planners to adjust schedules according to their expertise and real-time operational conditions, such as rush orders, machine maintenance, or other unforeseen disruptions.

Representative Publication:
C.-Y. Lee, C.-Y. Ho, Y.-H. Hung, and Y.-W. Deng, “Multi-objective genetic algorithm embedded with reinforcement learning for petrochemical melt-flow-index production scheduling,” Applied Soft Computing, vol. 159, p. 111630, 2024. [PDF]

Application Case – Predictive Maintenance

Predictive maintenance aims to detect equipment abnormalities and anticipate potential failure times in advance by leveraging sensor data and behavioral models, thereby reducing unplanned downtime and maintenance costs. By integrating multi-source data—such as vibration signals, current waveforms, temperature measurements, and production line status—health indicators and degradation models of equipment can be constructed. These models are further combined with time-series analysis, anomaly detection, and deep learning techniques to predict equipment degradation trends.

When abnormal patterns or significant deterioration in health indicators are detected, the system can promptly alert on-site engineers. Moreover, decision-making methods, such as reinforcement learning, can be employed to recommend optimal maintenance timing and repair prioritization. These approaches effectively mitigate equipment failure risks, extend equipment lifespan, and enhance overall production line stability, making predictive maintenance a core enabling technology for building reliable production systems in smart manufacturing.

Representative Publication:
Y.-H. Hung, H.-Y. Shen, and C.-Y. Lee, “Deep reinforcement learning-based preventive maintenance for repairable machines with deterioration in a flow line system,” Annals of Operations Research, pp. 1–21, 2024. [PDF]

Application Case – Mechanical Parameter Estimation and Monitoring

Mechanical parameter estimation and monitoring aim to continuously track variations in critical mechanical parameters using sensor data and system identification techniques, enabling accurate assessment of machine conditions and early detection of abnormalities. By integrating multiple sensing modalities—such as vibration, torque, displacement, and electrical current signals—this approach employs parameter estimation methods, state-space modeling, and statistical machine learning techniques to dynamically estimate key mechanical parameters, including friction coefficients, stiffness, damping, backlash, and external loads.

Long-term monitoring of the temporal evolution of these parameters allows for effective identification of incipient wear and degradation phenomena. The estimated parameters further serve as essential inputs for controller retuning, performance optimization, and predictive maintenance decision-making. Overall, this methodology enhances operational stability, improves control performance, and significantly reduces the risk of unexpected equipment downtime.

Representative Publication:
Y.-H. Hung, C.-Y. Lee, C.-H. Tsai, and Y.-M. Lu, “Constrained particle swarm optimization for health maintenance in three-mass resonant servo control system with LuGre friction model,” Annals of Operations Research, vol. 311, pp. 131–150, 2022. [PDF]

Application Case – Mechanical System Modeling

Mechanical system modeling aims to describe the dynamic behavior of equipment and mechanical systems by integrating physics-based models with data-driven approaches, thereby supporting downstream tasks such as control, prediction, and health management. By combining sensor measurements, operating conditions, and motion characteristics, this approach formulates dynamic equations, friction models, or state-space representations of the system. These models are further enhanced using identification techniques—including parameter estimation, system identification, and deep learning methods—to capture nonlinearities and uncertainties inherent in real-world mechanical systems.

Through this modeling process, system trajectories can be reconstructed, performance degradation can be predicted, and sources of modeling errors can be systematically analyzed. The resulting models provide essential foundations for controller tuning, performance optimization, and predictive maintenance, and play a critical role in enabling digital twins. Mechanical system modeling therefore serves as a cornerstone of intelligent manufacturing, facilitating deeper understanding of equipment behavior and improving operational reliability.

Representative Publication:
R.-Q. Hong, Y.-H. Hung, and C.-Y. Lee, “Correcting model misspecification by domain-adaptive physics-informed neural networks with interpretable auto-differentiation-based correction,” working paper.

Application Case – Process Parameter Optimization 

Process parameter optimization aims to identify optimal operating conditions that simultaneously improve yield, efficiency, and process stability by leveraging statistical and machine learning methods. By integrating historical process records, sensor data, and quality measurement results, process behavior models can be constructed to analyze the influence of key parameters on product quality and system performance. These models are then combined with optimization techniques or design of experiments (DoE) to identify efficient and feasible parameter configurations.

This data-driven optimization approach significantly reduces trial-and-error costs, shortens process tuning cycles, and enhances process consistency and product quality. As such, process parameter optimization plays a crucial role in enabling high-yield and stable production in intelligent manufacturing systems.

Representative Publication:
B.-R. Chen, Y.-H. Hung, and C.-Y. Lee, “Causal inference of policy learning for manufacturing process parameters,” INFORMS Journal on Data Science, under review.

Application Case – In-line Process Parameter Prediction and Monitoring 

In-line process parameter prediction and monitoring aims to continuously track critical operating conditions during production and provide early warnings before deviations occur. By integrating sensor signals, historical process records, and quality measurement data, adaptive predictive models can be developed to dynamically estimate key process parameters such as temperature, pressure, flow rate, and other operational variables in real time.

By combining time-series analysis, deep learning, and anomaly detection techniques, the system is able to identify abnormal trends, monitor process stability, and determine whether operating conditions are drifting away from their optimal ranges. This approach enables timely adjustments of process settings, helping to prevent quality degradation and equipment risks while improving production efficiency and process consistency. As a result, in-line predictive monitoring serves as a foundational capability for real-time quality assurance in intelligent manufacturing systems.

Representative Publication:
C.-Y. Lee, C.-S. Wu, and Y.-H. Hung, “In-line predictive monitoring framework,” IEEE Transactions on Automation Science and Engineering, vol. 18, no. 4, pp. 1669–1678, 2020. [PDF]

Application Case – Defect Detection 

Defect detection focuses on the automatic identification of product or process defects using imaging, sensor data, and machine learning techniques, with the goals of improving quality consistency and reducing reliance on manual inspection. By integrating computer vision, deep learning models, and statistical feature analysis, defect detection systems are capable of identifying surface scratches, voids, cracks, foreign objects, and process-induced anomalies.

These systems operate in-line and in real time, enabling rapid localization of suspicious regions, accurate defect classification, and seamless linkage to process parameters to support downstream root cause analysis. As a result, defect detection techniques significantly improve yield, shorten inspection cycles, and strengthen process quality control, forming a critical foundation for quality assurance in intelligent manufacturing systems.

Representative Publication:
T.-T. Hsieh, C.-Y. Lee, Y.-H. Hung, P.-C. Shen, and T. Yang, “Unpaired image denoising and fusion with adaptive multi-branch task UNet for semiconductor packaging defect recognition,” IEEE Transactions on Automation Science and Engineering, 2025.

Research – BMB-LIME

Many existing Explainable Artificial Intelligence (XAI) methods assume that a model exhibits local linear decision boundaries or regression behavior. However, in real-world data, local nonlinearity is often pronounced, causing traditional explanation methods to produce substantial approximation errors when modeling local classification boundaries or regression functions. In addition, perturbation-based explanation techniques typically yield single-point feature contribution estimates, failing to reflect the inherent uncertainty in explanations.

To address these limitations, we propose BMB-LIME, an explanation method designed to capture nonlinear local decision boundaries or regression structures while explicitly quantifying uncertainty in feature contributions. BMB-LIME constructs local surrogate models using weighted Multivariate Adaptive Regression Splines (MARS) and integrates bootstrap aggregating with Bayesian uncertainty estimation. This design enables a more accurate and robust characterization of local decision behavior, providing both high-fidelity explanations and uncertainty-aware feature attributions.

Representative Publication:
Y.-H. Hung and C.-Y. Lee, “BMB-LIME: LIME with modeling local nonlinearity and uncertainty in explainability,” Knowledge-Based Systems, vol. 294, p. 111732, 2024. [PDF]

Research – KDLIME

As complex black-box machine learning models are increasingly deployed in high-risk domains such as healthcare and criminal justice, the reliability and stability of model explanations have become critically important. While Local Interpretable Model-Agnostic Explanations (LIME) is a widely adopted local explanation method, it often suffers from instability due to randomly generated out-of-distribution perturbations. These perturbations can lead to poor local fidelity, inconsistent explanations, and vulnerability to adversarial manipulation.

To address these limitations, we propose KDLIME, an improved local explanation method that extends LIME through K-nearest neighbor (KNN)–based kernel density estimation. Instead of generating perturbations arbitrarily, KDLIME samples perturbation points from the in-distribution neighborhood of the target instance. This distribution-aware perturbation strategy enables the local surrogate model to fit explanations in a more representative and less biased region of the data space.

Representative Publication:
Y.-H. Hung and C.-Y. Lee, “KDLIME: KNN-kernel density-based perturbation for local interpretability,” in Proceedings of the Workshops of the European Conference on Machine Learning and Knowledge Discovery in Databases (ECML PKDD 2024): XKDD 2024, Vilnius, Lithuania, Sep. 9–13, 2024.