Machine Learning and Deep Learning for Manufacturing

Sang Won Lee¹

1 School of Mechanical Engineering, Sungkyunkwan University, Suwon-si, Republic of Korea

E-mail: sangwonl@skku.edu

Status

In the field of Smart Manufacturing, the application of Deep Learning (DL) and Machine Learning (ML) technologies has expanded across various areas such as quality control, machine and device maintenance, and process optimization. Recent advances have primarily focused on predictive maintenance, fault diagnosis, and product inspection.

Predictive maintenance has evolved into sophisticated frameworks which integrate LSTM-based DL for time-series data analysis, ML models such as Random Forest and XGBoost, and hybrid framework leveraging digital twins. For instance, in predictive maintenance research for industrial robots, a model combining temporal features extracted by LSTM and KNN classifier has demonstrated improved equipment uptime and extended failure intervals by incorporating knowledge graph-based maintenance strategies. Additionally, in machining centers, IoT-based data collection paired with Random Forest prediction models has shown high accuracy and real-time performance, effectively supporting operator decision-making [1].

In fault diagnosis, domain adaptation techniques have emerged as a key solution to data scarcity, while semi-supervised learning and federated learning address the need to utilize small amounts of labeled data. A representative study combining Active Learning with Semi-Supervised Learning demonstrated significant cost reduction in data labeling while maintaining fault classification accuracy across various industrial datasets [2].

In the domain of product inspection, real-time YOLO-based detection models, weakly supervised object localization networks, and semantic segmentation approaches have been explored. For instance, the CADN framework proposed a model capable of detecting defects using only image-level labels, while a comparative study between YOLOv4 and RCNN demonstrated their effectiveness in detecting defects during automotive spot-welding processes [3].

Attention-based lightweight networks have also reported for defect detection task [4]. By integrating custom-designed architectures on top of large-scale pretrained backbones, these models achieve superior performance in terms of both accuracy (mAP) and inference speed (FPS) on public datasets such as NEU-DET and GC10-DET.

Current and future challenges

The most demanding challenge in ML/DL-based smart manufacturing is the lack of high-quality labeled data. Due to security, cost, and operational complexity in industrial environments, collecting sufficient labeled datasets—especially for rare failure scenarios—is extremely difficult. While some studies have attempted to use GANs for synthetic fault data generation and LSTM architectures for time-series modeling [5], these approaches are not always applicable in the manufacturing sites where real failure data cannot be collected.

Secondly, limited generalizability and explainability remain critical issues. For instance, predictive models optimized for specific tools often struggle to generalize across varying machines or operational conditions. To address this, explainable AI (XAI) techniques are being increasingly incorporated to improve the interpretability of AI models [6].

Thirdly, achieving real-time performance and lightweight model deployment is an ongoing technical bottleneck. Especially in Edge-AI environments, resource-constrained environments demand efficient yet accurate models. Recent trends include compressing large-scale models via knowledge distillation and fine-tuning using domain-specific data to balance performance and complexity [7].

Lastly, cultural and institutional barriers must not be overlooked. SMEs often face delays or failures in AI adoption due to limited infrastructure, technical capacity, or upfront investment. In documented failures of Industry 4.0 transitions, factors such as unrealistic managerial expectations, infrastructure deficits, and internal resistance have been key contributors [8].

Advances in science and technology to meet challenges

Several technological pathways have been proposed to overcome the challenges as shown in Figure 1:

First, data scarcity and imbalance are addressed using transfer learning, self-supervised learning, GAN-based synthetic data generation, and federated learning. Specifically, domain adaptation has been employed to generalize the performance under limited samples, while GANs have been used to synthesize fault data or exploit unlabeled datasets [9,10,11].

Second, explainable AI (XAI) technologies are increasingly utilized to reduce computational costs during quality inspection while providing interpretable AI predictions for operators. Visualization techniques such as CAM and Grad-CAM help identifying decision rationale, while virtual sensor reconstruction and noise correction methods are developed to enhance data quality and trustworthiness [12,13].

Third, significant efforts have been made toward developing lightweight and real-time ML/DL model architectures. Techniques such as deformable convolution, channel attention, and bidirectional feature fusion have been embedded into modern network designs to achieve inference speeds of 30–100 FPS while maintaining high accuracy, enabling edge deployment in production environments [14,15].

Fourth, the integration of digital twins with hybrid modeling has gained increasing importance. Hybrid frameworks that combine physics-based simulations with sensor-driven data models have shown notable improvements in predicting component lifetimes. These approaches are particularly effective in enhancing prediction accuracy and scalability in real-world operations [16,17].

Lastly, privacy-preserving learning approaches have received considerable attention to address cross-enterprise data sharing concerns. Federated Learning frameworks enable collaborative training of high-performance models without exposing sensitive manufacturing data, and their effectiveness has been demonstrated on real production datasets [18].

Figure 1. Recent challenges and technological advances in ML/DL for smart manufacturing.

Concluding remarks

In conclusion, although ML/DL-based smart manufacturing is rapidly advancing and offers significant potential, it continues to face several foundational challenges. These include difficulties in securing labeled high-quality industrial datasets, limitations in generalization across diverse operational scenarios, low interpretability of complex models, and substantial gaps in infrastructure and readiness—especially for SMEs.

To address the challenges, a comprehensive and layered strategy is required. First, enhanced data acquisition pipelines must incorporate not only sensor-based monitoring but also simulation (digital twin)- driven synthetic augmentation. Integration of heterogeneous structured and unstructured data across domains will facilitate the development of more comprehensive AI models.

Human-centered interpretability must be also incorporated from the early design stage. In addition to visualizations for transparency, techniques such as feature importance highlighting, attention heatmaps, and contextual explanations should empower operators to understand and validate AI-based decisions. This fosters trust and improves field-level acceptance.

Furthermore, continuous innovation in lightweight AI model architecture is essential. In edge computing environments, where real-time decision-making is required, new forms of DL/ML methodologies must be developed that balance the trade-off between accuracy and computational efficiency. In particular, there is growing interest in model designs that minimize energy consumption while maintaining high performance, as well as federated fine-tuning techniques that incrementally improve model accuracy using localized field data.

Ultimately, the successful deployment and scaling of AI-enabled manufacturing systems will require coordinated collaboration among academia, industry, and government. When technical innovations are validated through empirical testing and aligned with institutional and workforce strategies, smart manufacturing will evolve beyond automation to become transparent, resilient, and human-aligned intelligent production systems.

Acknowledgements

This article was supported by the Industry Technology Alchemist Project funded by the Ministry of Trade, Industry & Energy (MOTIE, Korea) (No. 20025702), and National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2022R1A2C3012900).

References

[1] Rosati, R., Romeo, L., Cecchini, G., Tonetto, F., Viti, P., Mancini, A., & Frontoni, E., From knowledge-based to big data analytic model: a novel IoT and machine learning based decision support system for predictive maintenance in Industry 4.0. 2023 *Journal of Intelligent Manufacturing,*34, 107–121 [2] Farahani, B., & Monsefi, A. K., Smart and collaborative industrial IoT: A federated learning and data space approach, 2023 *Digital Communications and Networks,*9, 436-447 [3] Zhang, J., Su, H., Zou, W., Gong, X., Zhang, Z., & Shen, F., CADN: A weakly supervised learning-based category-aware object detection network for surface defect detection, 2020 *Pattern Recognition,*109, 107571 [4] Chen, H., Du, Y., Fu, Y., Zhu, J., & Zeng, H., DCAM-Net: A Rapid Detection Network for Strip Steel Surface Defects Based on Deformable Convolution and Attention Mechanism, 2023 *IEEE Transactions on Instrumentation and Measurement,*72, 1- 12 [5] Alzubaidi, L., Bai, J., Al-Sabaawi, A., Santamaría, J., Albahri, A. S., Al-Dabbagh, B. S. N., ... & Gu, Y., A survey on deep learning tools dealing with data scarcity: definitions, challenges, solutions, tips, and applications, 2023 *Journal of Big Data,*10, 46 [6] Lei, Y., Yang, B., Jiang, X., Jia, F., Li, N., & Nandi, A. K., Applications of machine learning to machine fault diagnosis: A review and roadmap, 2020 *Mechanical systems and signal processing,*138, 106587 [7] Zhang, H., Semujju, S. D., Wang, Z., Lv, X., Xu, K., Wu, L., ... & Ma, X., Large scale foundation models for intelligent manufacturing applications: a survey, 2025 *Journal of Intelligent Manufacturing,*1-52 [8] Pfeifer, M. R., SMEs in Failed Transition Towards Industry 4.0: A case study of a Czech SME, 2021 Journal of Innovation & Business Best Practice, 1-16 [9] Jain, S., Seth, G., Paruthi, A., Soni, U., & Kumar, G., Synthetic data augmentation for surface defect detection and classification using deep learning. Journal of Intelligent Manufacturing, 1-14 [10] Zhong, H., Yu, S., Trinh, H., Yuan, R., Lv, Y., & Wang, Y. A time-saving fault diagnosis using simplified fast GAN and triple-type data transfer learning. 2025 Structural Health Monitoring,24, 869-882 [11] Mehta, M., Chen, S., Tang, H., & Shao, C. A federated learning approach to mixed fault diagnosis in rotating machinery. 2023 Journal of Manufacturing Systems,68, 687-694 [12] Bordekar, H., Cersullo, N., Brysch, M., Philipp, J., & Hühne, C. eXplainable artificial intelligence for automatic defect detection in additively manufactured parts using CT scan analysis. 2025 Journal of Intelligent Manufacturing, 36, 957-974. [13] Fabiocchi, D., Giulietti, N., Carnevale, M., & Giberti, H. Ai-driven virtual sensors for real-time dynamic analysis of mechanisms: A feasibility study. 2024 Machines, 12, 257. [14] Wan, D., Lu, R., Shen, S., Xu, T., Lang, X., & Ren, Z. Mixed local channel attention for object detection. 2023 Engineering Applications of Artificial Intelligence,123, 106442 [15] Zhang, X., Song, Y., Song, T., Yang, D., Ye, Y., Zhou, J., & Zhang, L. LDConv: Linear deformable convolution for improving convolutional neural networks. 2024 Image and Vision Computing,149, 105190. [16] Iranshahi, K., Brun, J., Arnold, T., Sergi, T., & Müller, U. C. Digital Twins: Recent Advances and Future Directions in Engineering Fields. 2025 Intelligent Systems with Applications, 200516.

[17] Sahin, T., Wolff, D., von Danwitz, M., & Popp, A. Towards a hybrid digital twin: Physics-informed neural networks as surrogate model of a reinforced concrete beam. 2024 arXiv preprint arXiv:2405.08406. [18] Piran, F. J., Chen, Z., Imani, M., & Imani, F. Privacy-preserving federated learning with differentially private hyperdimensional computing. 2025 Computers and Electrical Engineering,123, 110261.