Non-Traditional Machine Learning for Highly Connected and Complex Manufacturing Systems
Dai-Yan Ji¹, Takanobu Minami1, Ruoxin Wang¹ and Jay Lee¹
¹Center for Industrial Artificial Intelligence, Department of Mechanical Engineering, University of Maryland, College Park, MD 20742, USA
E-mail: leejay@umd.edu
Status
The development of prognostics and health management (PHM) and its integration into industrial AI has progressed from component-level monitoring toward system-level intelligence [1], [2]. In its early stages, PHM relied on reliability engineering and model-based analysis, later evolving into hybrid AI frameworks that incorporate data-driven methods with expert knowledge [3], [4]. These advances have produced valuable results for predictive maintenance and monitoring, yet the increasing scale and interconnectivity of highly connected and complex manufacturing systems (HC-CMS) continue to challenge traditional machine learning (ML) methods. Today’s industrial systems are no longer limited to single components or individual
units; they have become fleet-based, distributed, and deeply networked across factories, supply chains, and operational domains [5], [6]. Such systems generate vast sensor data and involve heterogeneous assets that must be managed collectively. As illustrated in Figure 1, HC-CMS connect multiple domains of industrial AI, including manufacturing AI, new energy AI, transportation AI, and healthcare AI, all of which share common requirements for resilience, adaptability, and intelligent decision-making. This system-level complexity underscores why PHM remains vital. In aerospace, energy, marine, and mobility, the ability to ensure uptime, optimize maintenance, and extend life cycles remains a cornerstone of reliability [7], [8]. With systems expanding into fleets and multi-plant networks, the importance of scalable PHM continues to grow. Further advances promise significant benefits: improved safety, reduced downtime, enhanced efficiency, and more trustworthy decision support [9]. Non-traditional machine learning is increasingly central to achieving these gains. Rather than focusing narrowly on algorithmic accuracy, the field now emphasizes resilience, interpretability, and enterprise-scale integration, and is supported by methodology platforms such as the continuous machine learning: stream-of-quality (SoQ), 5C-level cyber-physical system, and digital twin framework. Representative non-traditional ML approaches—topological data analysis (TDA) for structural insight, domain adaptation and transfer learning for fleet-wide generalization, similarity-based models for interpretable reasoning, surrogate models for efficient optimization, and industrial large knowledge models (ILKM) for knowledge integration—illustrate how PHM in HC-CMS is being reshaped and extended. These paradigms demonstrate why PHM is not only still important but also positioned to deliver even greater impact as Industrial AI continues to evolve [10].
Current and future challenges
Despite the progress of Industrial AI, scaling PHM in HC-CMS continues to face formidable obstacles. Cross-domain variability remains a critical barrier: models trained on one production line often degrade when deployed across different plants or fleets due to domain shifts in operating conditions. This problem is amplified in industries such as wind energy, mobility batteries, and marine engines, where operational environments change rapidly and sensor distributions are inconsistent [11]. Label scarcity and data imbalance also persist, as failure events are rare, costly to capture, and frequently undocumented, limiting the applicability of supervised deep learning approaches [12]. Equally important is the challenge of system interconnectivity. Reconfigurable and sustainable production paradigms introduce nonlinear couplings, shifting bottlenecks, and complex scheduling dynamics that cannot be reduced to isolated equipment analysis. Interpretability further complicates adoption: practitioners often reject black-box models without transparent reasoning. Studies in aircraft engine prognostics have emphasized the role of aggregated feature importance and interpretable dimensionality reduction to build trust in PHM predictions [8]. Hybrid models face their own limitations. Data-driven frameworks excel when rich signals are available but struggle under distributional drift. Physics-based models are transparent yet require detailed failure physics, which are not always accessible. Hybrids attempt to combine these strengths, but parameter calibration and model updating remain complex [13]. From an organizational standpoint, knowledge fragmentation is a systemic issue. PHM efforts often rely on expertise contained in manuals, reports, and personal experience, making consistent integration into PHM difficult. Smart factory environments add further constraints, including cybersecurity, CPS/IoT integration, and the governance of heterogeneous big data. Recent reviews stress that a persistent gap between algorithmic advances and practical implementation continues to hinder PHM adoption. Based on the above description, these issues underscore the urgency of non-traditional ML techniques that embed robustness, adaptability, and explainability. Without addressing data scarcity, cross-domain adaptation, and interpretability, PHM systems will remain fragile and limited in delivering enterprise-wide value.
Figure 1. Highly connected and complex manufacturing systems
Advances in science and technology to meet challenges
Addressing these challenges requires a series of focused scientific and technological advances. Building on this foundation, TDA has gained prominence as a streaming-compatible tool. Persistence-based descriptors allow for the detection of subtle distributional changes in production data, providing robust health indicators even under noisy and high-dimensional conditions [14]. TDA is particularly useful for long-term monitoring in HC-CMS since it not only makes anomaly detection possible but also clusters operational states and makes it easier to identify system transitions. In parallel, domain adaptation and transfer learning have been extended with ensemble, meta-learning, and continual strategies that enable models to adapt to evolving conditions in fleets of assets, such as wind farms and marine systems [15]. These methods not only reduce the cost of retraining but also support rapid deployment when labeled target data are scarce, ensuring greater generalization across diverse industrial scenarios. Extending from these developments, Similarity-based models have emerged as a crucial non-traditional method. By leveraging case libraries and distance metrics, similarity-based approaches provide transparent reasoning, enabling interpretable RUL estimation and diagnostics that practitioners can validate against historical precedents [16]. When enriched with retrieval-augmented embeddings, these models enhance both efficiency and explainability, supporting decision-making in complex operational contexts. Moreover, similarity-based reasoning fosters knowledge reuse across assets, allowing engineers to justify decisions with concrete historical references. The use of surrogate models has grown quickly. From energy systems to compact lens assemblies, neural operators, Bayesian surrogates, and differentiable simulators are being used more and more in industrial design and optimization to provide real-time decision support with quantified uncertainty [17]. Surrogate models not only speed up computationally costly simulations but also offer a way to combine physics-informed constraints with machine learning models to produce hybrid solutions that maintain a balance between interpretability and accuracy. The most transformative development is the ILKM framework. By constructing structured knowledge libraries, aligning them with industrial workflows, and coupling them with instruction-tuned large models, ILKMs enable retrieval-augmented, auditable decision-making across smart factories [18]. This approach allows PHM systems to incorporate human expertise, domain knowledge, and large-scale analytics within a unified platform, directly addressing the issue of fragmented expertise. ILKMs also create opportunities for cross-domain reasoning, linking maintenance records, quality data, and operational best practices into an integrated knowledge ecosystem. Recent reviews across Industrial AI applications consistently emphasize the importance of explainable AI, uncertainty quantification, and enterprise-level integration [19], highlighting the crucial role of non-traditional ML. Recent progress in
foundation models further extends these non-traditional approaches toward system-level industrial intelligence. Under the SoQ paradigm (Figure 2), foundation model-based SoQ can be structured into four research thrusts: representation, efficient adaptation, dynamic learning, and cognitive reasoning. This emerging framework enables continuous modeling of quality propagation across stages, scalable deployment in distributed environments, real-time adaptation to evolving processes, and lifecycle-aware knowledge-driven reasoning. Such developments mark a transition from data-centric predictive maintenance toward integrated industrial cognition.
Figure 2. Foundation model-based SoQ for HC-CMS
Concluding remarks
The direction of PHM research shows a clear shift from algorithm-focused investigations toward a broader emphasis on system-level intelligence in HC-CMS. Traditional ML methods, while successful in controlled settings, struggle with distributional shift, data scarcity, and limited interpretability. Non-traditional ML techniques—TDA for robust structure discovery, domain adaptation for generalization, similarity-based reasoning for transparency, surrogates for cost-efficient optimization, and ILKM for enterprise-wide integration—together provide a coherent roadmap not only for the next phase of PHM but also for the advancement of Industrial AI applications. Future PHM within Industrial AI must place emphasis on adaptability, interpretability, and scalability. In order to convert predictive accuracy into actionable intelligence, these strategies—such as ILKM frameworks that integrate disparate areas of expertise and non-traditional ML methods grounded in domain knowledge—will be crucial. Higher uptime, safer operations,
lower maintenance costs, and more robust production ecosystems are expected advantages of these advances. The main challenge [20] for the community is not the pursuit of algorithmic novelty, but rather the successful integration of these techniques into enterprise-scale, auditable, and reliable industrial systems that can deliver lasting impact.
Acknowledgements
This work was supported by the U.S. Department of Education through the Fund for the Improvement of Postsecondary Education (FIPSE) under Grant No. P116S230014.
References
[1] D.-Y. Ji, H. Su, T. Minami, and J. Lee, “Introduction to Industrial Artificial Intelligence,” Advances in Artificial Intelligence Applications in Industrial and Systems Engineering, pp. 1–18, 2025.
[2] J. Lee, J. Ni, J. Singh, B. Jiang, M. Azamfar, and J. Feng, “Intelligent maintenance systems and predictive manufacturing,” Journal of Manufacturing Science and Engineering, vol. 142, no. 11, p. 110805, 2020.
[3] S. Ochella, M. Shafiee, and F. Dinmohammadi, “Artificial intelligence in prognostics and health management of engineering systems,” Engineering Applications of Artificial Intelligence, vol. 108, p. 104552, 2022.
[4] K. T. Nguyen, K. Medjaher, and D. T. Tran, “A review of artificial intelligence methods for engineering prognostics and health management with implementation guidelines,” Artificial Intelligence Review, vol. 56, no. 4, pp. 3659–3709, 2023.
[5] V. Atamuradov, K. Medjaher, P. Dersin, B. Lamoureux, and N. Zerhouni, “Prognostics and health management for maintenance practitioners-Review, implementation and tools evaluation.,” International Journal of Prognostics and Health Management, vol. 8, no. 3, pp. 1–31, 2017.
[6] H. Gharib and G. Kovács, “A review of prognostic and health management (PHM) methods and limitations for marine diesel engines: new research directions,” Machines, vol. 11, no. 7, p. 695, 2023.
[7] M. Yan, S. C. Hui, N. Jiang, and N. Li, “A review on data-driven prognostics and health management for wind turbine systems,” Engineering Applications of Artificial Intelligence, vol. 159, p. 111484, 2025.
[8] Y. Alomari, M. Andó, and M. L. Baptista, “Advancing aircraft engine RUL predictions: an interpretable integrated approach of feature engineering and aggregated feature importance,” Scientific Reports, vol. 13, no. 1, p. 13466, 2023.
[9] S. Han, F. Li, X. Han, and S. Zhang, “A novel feature engineering method for network anomaly detection,” Computers and Electrical Engineering, vol. 127, p. 110627, 2025.
[10] C. Huang, S. Bu, H. H. Lee, C. H. Chan, S. W. Kong, and W. K. Yung, “Prognostics and health management for predictive maintenance: A review,” Journal of manufacturing systems, vol. 75, pp. 78–101, 2024.
[11] P. Kumar, M. Tanveer, K. Park, I. Raouf, and H. S. Kim, “A Systematic Review of Prognostics and Health Management in Mobility Batteries,” International Journal of Precision Engineering and Manufacturing-Green Technology, pp. 1–24, 2025.
[12] X. Lai, H. Shui, D. Ding, and J. Ni, “Data-driven dynamic bottleneck detection in complex manufacturing systems,” Journal of Manufacturing Systems, vol. 60, pp. 662–675, 2021.
[13] P. Kundu, A. K. Darpe, and M. S. Kulkarni, “Development of data-driven, physics-based, and hybrid prognosis frameworks: a case study for gear remaining useful life prediction,” Journal of Intelligent Manufacturing, pp. 1–21, 2024.
[14] J. Lee, D.-Y. Ji, and Y.-M. Hsu, “Novel topological machine learning methodology for stream-of-quality modeling in smart manufacturing,” Manufacturing Letters, vol. 43, pp. 60–63, 2025.
[15] T. Minami, A. Suer, P. Kundu, S. Siahpour, and J. Lee, “Novel ensemble domain adaptation methodology for enhanced multi-class fault diagnosis of highly-connected fleet of assets,” in Phm society asia-pacific conference, 2023.
[16] T. Minami, D.-Y. Ji, and J. Lee, “PHM for Spacecraft Propulsion Systems: Developing Resilient Models for Real-World Challenges,” in PHM Society European Conference, 2024, pp. 7–7.
[17] W. Li et al., “A novel methodology for lens matching in compact lens module assembly,” IEEE Transactions on Automation Science and Engineering, vol. 20, no. 2, pp. 741–750, 2022.
[18] Lee J and Su H, A unified industrial large knowledge model framework in Industry 4.0 and smart manufacturing Int. J. AI *Mater. Des.*1 (2) 41–47, 2025
[19] T. Xia, Y. Dong, L. Xiao, S. Du, E. Pan, and L. Xi, “Recent advances in prognostics and health management for advanced manufacturing paradigms,” Reliability Engineering & System Safety, vol. 178, pp. 255–268, 2018.
[20] P. Kumar, I. Raouf, and H. S. Kim, “Review on prognostics and health management in smart factory: From conventional to deep learning perspectives,” Engineering Applications of Artificial Intelligence, vol. 126, p. 107126, 2023.