2025
Chawalitanont, Akarawint; Bashyal, Atit; Wicaksono, Hendro
In: Journal of Manufacturing Systems, vol. 83, pp. 713–735, 2025.
Abstract | Links | BibTeX | Tags: artificial intelligence, deep learning, industry 4.0, industry 5.0, machine learning, manufacturing, sustainability
@article{chawalitanont2025uncertaintyb,
title = {Uncertainty-aware power consumption prediction in customized stainless-steel manufacturing: A comparative study of hierarchical Bayesian and deep neural models},
author = {Akarawint Chawalitanont and Atit Bashyal and Hendro Wicaksono},
doi = {https://doi.org/10.1016/j.jmsy.2025.10.010},
year = {2025},
date = {2025-12-01},
urldate = {2025-12-01},
journal = {Journal of Manufacturing Systems},
volume = {83},
pages = {713–735},
publisher = {Elsevier},
abstract = {Energy-efficient and data-driven decision-making has become a critical priority in modern manufacturing, particularly in customized or make-to-order (MTO) production where product variability causes large fluctuations in power consumption. Existing prediction models in this domain are often deterministic, lacking the ability to quantify uncertainty and capture hierarchical data dependencies, which limits their reliability for operational use. This study addresses this gap by developing a hierarchical Bayesian learning framework for power consumption prediction in customized stainless-steel manufacturing. The objective is to design models that not only achieve high predictive accuracy but also provide calibrated uncertainty estimates to support risk-aware production decisions. Four models, i.e., Hierarchical Bayesian Linear Regression (HBLR), Hierarchical Bayesian Neural Network (HBNN), Fully Connected Neural Network (FCN), and One-Dimensional Convolutional Neural Network (1D-CNN), were implemented and benchmarked using three inference algorithms: No-U-Turn Sampler (NUTS), Automatic Differentiation Variational Inference (ADVI), and Stein Variational Gradient Descent (SVGD). The innovation lies in systematically quantifying uncertainty using coverage probability, sharpness, and calibration error, and in establishing a unified comparison between probabilistic and deterministic models. Results show that the HBLR–NUTS model achieves the best trade-off between accuracy (RMSE = 11.85) and calibration quality (coverage 0.98), while ADVI offers near-equivalent performance with significantly lower computation time. These uncertainty-aware predictions can be directly integrated into Manufacturing Execution System (MES) and Enterprise Resource Planning (ERP) environments for energy-optimized scheduling and cost-aware planning. The proposed framework provides a scalable, interpretable, and statistically reliable foundation for advancing sustainable, data-driven manufacturing analytics.},
keywords = {artificial intelligence, deep learning, industry 4.0, industry 5.0, machine learning, manufacturing, sustainability},
pubstate = {published},
tppubtype = {article}
}
Energy-efficient and data-driven decision-making has become a critical priority in modern manufacturing, particularly in customized or make-to-order (MTO) production where product variability causes large fluctuations in power consumption. Existing prediction models in this domain are often deterministic, lacking the ability to quantify uncertainty and capture hierarchical data dependencies, which limits their reliability for operational use. This study addresses this gap by developing a hierarchical Bayesian learning framework for power consumption prediction in customized stainless-steel manufacturing. The objective is to design models that not only achieve high predictive accuracy but also provide calibrated uncertainty estimates to support risk-aware production decisions. Four models, i.e., Hierarchical Bayesian Linear Regression (HBLR), Hierarchical Bayesian Neural Network (HBNN), Fully Connected Neural Network (FCN), and One-Dimensional Convolutional Neural Network (1D-CNN), were implemented and benchmarked using three inference algorithms: No-U-Turn Sampler (NUTS), Automatic Differentiation Variational Inference (ADVI), and Stein Variational Gradient Descent (SVGD). The innovation lies in systematically quantifying uncertainty using coverage probability, sharpness, and calibration error, and in establishing a unified comparison between probabilistic and deterministic models. Results show that the HBLR–NUTS model achieves the best trade-off between accuracy (RMSE = 11.85) and calibration quality (coverage 0.98), while ADVI offers near-equivalent performance with significantly lower computation time. These uncertainty-aware predictions can be directly integrated into Manufacturing Execution System (MES) and Enterprise Resource Planning (ERP) environments for energy-optimized scheduling and cost-aware planning. The proposed framework provides a scalable, interpretable, and statistically reliable foundation for advancing sustainable, data-driven manufacturing analytics.
Boroukhian, Tina; Supyen, Kritkorn; Samson, Jhealyn Bautista; Bashyal, Atit; Wicaksono, Hendro
Integrating 3D object detection with ontologies for accurate digital twin creation in manufacturing systems Journal Article
In: The International Journal of Advanced Manufacturing Technology, vol. 140, no. 9, pp. 4679–4711, 2025.
Abstract | Links | BibTeX | Tags: data science, deep learning, digital twins, image processing, industry 4.0, machine learning, manufacturing, object detection, ontologies, semantic web
@article{boroukhian2025integrating,
title = {Integrating 3D object detection with ontologies for accurate digital twin creation in manufacturing systems},
author = {Tina Boroukhian and Kritkorn Supyen and Jhealyn Bautista Samson and Atit Bashyal and Hendro Wicaksono},
doi = {https://doi.org/10.1007/s00170-025-16548-x},
year = {2025},
date = {2025-10-01},
urldate = {2025-10-01},
journal = {The International Journal of Advanced Manufacturing Technology},
volume = {140},
number = {9},
pages = {4679–4711},
publisher = {Springer London},
abstract = {The digitization of manufacturing resources through digital twins (DTs) enhances operational efficiency and resource management. Ontologies play a key role in maintaining semantic consistency within DT systems. However, existing ontology-based approaches face challenges, including limited adaptability, integration of heterogeneous data—such as 3D images—and high manual effort in ontology development. These limitations hinder the scalability of DT implementations. Traditional 2D imaging often lacks spatial accuracy in complex manufacturing environments, causing inefficiencies and higher costs. Integrating richer data with intelligent frameworks is crucial for improving production and adaptability. The proposed study addresses these challenges by introducing a methodology that integrates existing ontologies with advanced 3D object detection models. The proposed approach employs two fully automated pipelines: one for detecting manufacturing resources from 3D images and another for mapping them into ontologies, ensuring seamless integration into DT frameworks. By leveraging established ontologies, the methodology enhances interoperability, reduces implementation complexity, and facilitates scalable deployment of DT systems across various industrial applications. Additionally, a comparative analysis of multiple advanced 3D detection models provides valuable insights to guide the selection of optimal solutions for diverse industrial settings. Experimental results show that YOLOv8 achieved the highest performance, with 91% classification accuracy, 86% precision, 81% recall, and the fastest inference time of 0.66 s. For ontology population, four machine labels—Robot, MillingMachine, BandSaw, and Lathe—were successfully integrated using a semantic similarity-based mapping strategy, enabling automated class creation and merging. This innovative framework sets a new benchmark for DT implementations, offering enhanced accuracy, efficiency, and semantic coherence in modern manufacturing.},
keywords = {data science, deep learning, digital twins, image processing, industry 4.0, machine learning, manufacturing, object detection, ontologies, semantic web},
pubstate = {published},
tppubtype = {article}
}
The digitization of manufacturing resources through digital twins (DTs) enhances operational efficiency and resource management. Ontologies play a key role in maintaining semantic consistency within DT systems. However, existing ontology-based approaches face challenges, including limited adaptability, integration of heterogeneous data—such as 3D images—and high manual effort in ontology development. These limitations hinder the scalability of DT implementations. Traditional 2D imaging often lacks spatial accuracy in complex manufacturing environments, causing inefficiencies and higher costs. Integrating richer data with intelligent frameworks is crucial for improving production and adaptability. The proposed study addresses these challenges by introducing a methodology that integrates existing ontologies with advanced 3D object detection models. The proposed approach employs two fully automated pipelines: one for detecting manufacturing resources from 3D images and another for mapping them into ontologies, ensuring seamless integration into DT frameworks. By leveraging established ontologies, the methodology enhances interoperability, reduces implementation complexity, and facilitates scalable deployment of DT systems across various industrial applications. Additionally, a comparative analysis of multiple advanced 3D detection models provides valuable insights to guide the selection of optimal solutions for diverse industrial settings. Experimental results show that YOLOv8 achieved the highest performance, with 91% classification accuracy, 86% precision, 81% recall, and the fastest inference time of 0.66 s. For ontology population, four machine labels—Robot, MillingMachine, BandSaw, and Lathe—were successfully integrated using a semantic similarity-based mapping strategy, enabling automated class creation and merging. This innovative framework sets a new benchmark for DT implementations, offering enhanced accuracy, efficiency, and semantic coherence in modern manufacturing.
Bashyal, Atit; Boroukhian, Tina; Veerachanchai, Pakin; Naransukh, Myanganbayar; Wicaksono, Hendro
Multi-agent deep reinforcement learning based demand response and energy management for heavy industries with discrete manufacturing systems Journal Article
In: Applied Energy, vol. 392, pp. 125990, 2025.
Abstract | Links | BibTeX | Tags: artificial intelligence, data science, deep learning, demand response system, energy management, green energy, machine learning, manufacturing, operation research, reinforcement learning, sustainability
@article{bashyal2025multi,
title = {Multi-agent deep reinforcement learning based demand response and energy management for heavy industries with discrete manufacturing systems},
author = {Atit Bashyal and Tina Boroukhian and Pakin Veerachanchai and Myanganbayar Naransukh and Hendro Wicaksono},
doi = {https://doi.org/10.1016/j.apenergy.2025.125990},
year = {2025},
date = {2025-08-15},
urldate = {2025-01-01},
journal = {Applied Energy},
volume = {392},
pages = {125990},
publisher = {Elsevier},
abstract = {Energy-centric decarbonization of heavy industries, such as steel and cement, necessitates their participation in integrating Renewable Energy Sources (RES) and effective Demand Response (DR) programs. This situation has created the opportunities to research control algorithms in diverse DR scenarios. Further, the industrial sector’s unique challenges, including the diversity of operations and the need for uninterrupted production, bring unique challenges in designing and implementing control algorithms. Reinforcement learning (RL) methods are practical solutions to the unique challenges faced by the industrial sector. Nevertheless, research in RL for industrial demand response has not yet achieved the level of standardization seen in other areas of RL research, hindering broader progress. To propel the research progress, we propose a multi-agent reinforcement learning (MARL)-based energy management system designed to optimize energy consumption in energy-intensive industrial settings by leveraging dynamic pricing DR schemes. The study highlights the creation of a MARL environment and addresses these challenges by designing a general framework that allows researchers to replicate and implement MARL environments for industrial sectors. The proposed framework incorporates a Partially Observable Markov Decision Process (POMDP) to model energy consumption and production processes while introducing buffer storage constraints and a flexible reward function that balances production efficiency and cost reduction. The paper evaluates the framework through experimental validation within a steel powder manufacturing facility. The experimental results validate our framework and also demonstrate the effectiveness of the MARL-based energy management system.},
keywords = {artificial intelligence, data science, deep learning, demand response system, energy management, green energy, machine learning, manufacturing, operation research, reinforcement learning, sustainability},
pubstate = {published},
tppubtype = {article}
}
Energy-centric decarbonization of heavy industries, such as steel and cement, necessitates their participation in integrating Renewable Energy Sources (RES) and effective Demand Response (DR) programs. This situation has created the opportunities to research control algorithms in diverse DR scenarios. Further, the industrial sector’s unique challenges, including the diversity of operations and the need for uninterrupted production, bring unique challenges in designing and implementing control algorithms. Reinforcement learning (RL) methods are practical solutions to the unique challenges faced by the industrial sector. Nevertheless, research in RL for industrial demand response has not yet achieved the level of standardization seen in other areas of RL research, hindering broader progress. To propel the research progress, we propose a multi-agent reinforcement learning (MARL)-based energy management system designed to optimize energy consumption in energy-intensive industrial settings by leveraging dynamic pricing DR schemes. The study highlights the creation of a MARL environment and addresses these challenges by designing a general framework that allows researchers to replicate and implement MARL environments for industrial sectors. The proposed framework incorporates a Partially Observable Markov Decision Process (POMDP) to model energy consumption and production processes while introducing buffer storage constraints and a flexible reward function that balances production efficiency and cost reduction. The paper evaluates the framework through experimental validation within a steel powder manufacturing facility. The experimental results validate our framework and also demonstrate the effectiveness of the MARL-based energy management system.
Ghribi, Youssef; Graha, Ega Rudy; Wicaksono, Hendro
Comparative Analysis of Statistical and Machine Learning Models for Enhancing Demand Forecasting Accuracy in the Medical Device Industry Journal Article
In: Procedia CIRP, vol. 134, pp. 849–854, 2025.
Abstract | Links | BibTeX | Tags: artificial intelligence, data science, deep learning, demand forecasting, healthcare, machine learning, manufacturing, supply chain management
@article{ghribi2025comparative,
title = {Comparative Analysis of Statistical and Machine Learning Models for Enhancing Demand Forecasting Accuracy in the Medical Device Industry},
author = {Youssef Ghribi and Ega Rudy Graha and Hendro Wicaksono},
doi = {https://doi.org/10.1016/j.procir.2025.02.209},
year = {2025},
date = {2025-01-01},
urldate = {2025-01-01},
journal = {Procedia CIRP},
volume = {134},
pages = {849–854},
publisher = {Elsevier},
abstract = {Demand forecasting is a crucial instrument in the business strategy. The medical devices in the healthcare system are further significant as critical roles. Multiple businesses rely on traditional forecasting techniques due to their simplicity and understandable algorithm’s easy-to-use nature characteristics. The research conducted for each model analyzes how traditional statistical, Machine Learning (ML), and Deep Learning (DL) models can be used to make demand forecasting more accurate and valuable in the medical device industry. The work expands beyond prior research to demonstrate the enhanced effectiveness of DL models compared to statistical and ML models across multiple areas. However, research still needs to identify studies that adopt a business-centric perspective on the practical applicability of these models. Research utilizing SARIMAX, Exponential Smoothing, Linear Regression, Average, Support Vector Regression (SVR), Random Forest (RF), K-Nearest Neighbour Regression (KNR), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Convolution 1D (CONV1D) models to forecast what people demand to order. The data comes from a German medical device manufacturer’s past sales record. We evaluated the model’s performance using the weighted Mean Absolute Percentage Error (wMAPE) method. These showed that DL models needed a lot of knowledge and preprocessing, but they were the most accurate at predicting what would happen. The LSTM model exhibited outstanding performance, achieving an average wMAPE of 0.3102, surpassing all other models. The research results for more sophisticated models surpass traditional statistical models despite limited datasets, recommending that medical device businesses consider investing in advanced demand forecasting models.
publisher={Elsevier}
}},
keywords = {artificial intelligence, data science, deep learning, demand forecasting, healthcare, machine learning, manufacturing, supply chain management},
pubstate = {published},
tppubtype = {article}
}
Demand forecasting is a crucial instrument in the business strategy. The medical devices in the healthcare system are further significant as critical roles. Multiple businesses rely on traditional forecasting techniques due to their simplicity and understandable algorithm’s easy-to-use nature characteristics. The research conducted for each model analyzes how traditional statistical, Machine Learning (ML), and Deep Learning (DL) models can be used to make demand forecasting more accurate and valuable in the medical device industry. The work expands beyond prior research to demonstrate the enhanced effectiveness of DL models compared to statistical and ML models across multiple areas. However, research still needs to identify studies that adopt a business-centric perspective on the practical applicability of these models. Research utilizing SARIMAX, Exponential Smoothing, Linear Regression, Average, Support Vector Regression (SVR), Random Forest (RF), K-Nearest Neighbour Regression (KNR), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Convolution 1D (CONV1D) models to forecast what people demand to order. The data comes from a German medical device manufacturer’s past sales record. We evaluated the model’s performance using the weighted Mean Absolute Percentage Error (wMAPE) method. These showed that DL models needed a lot of knowledge and preprocessing, but they were the most accurate at predicting what would happen. The LSTM model exhibited outstanding performance, achieving an average wMAPE of 0.3102, surpassing all other models. The research results for more sophisticated models surpass traditional statistical models despite limited datasets, recommending that medical device businesses consider investing in advanced demand forecasting models.
publisher={Elsevier}
}
publisher={Elsevier}
}