Reference Model for Machine Learning Operations in Manufacturing: A Scalable Blueprint for Industrial AI
In today’s data-rich industrial world, manufacturers are collecting more machine, process, and quality data than ever before. Yet, the potential of this data remains largely untapped. The vision of smart manufacturing showcases that Machine Learning (ML) should do more than run pilots – it should deliver real value at scale.
To realize that vision, manufacturers need an MLOps foundation tailored to their unique requirements. Generic tools built for digital-native companies often fail to meet the demands of industrial manufacturing. That’s where a domain-specific reference model for MLOps in manufacturing can help – scalable, modular, and ready for industrial deployment.
This blog post outlines the key components of a manufacturing-specific MLOps reference model, explaining how it addresses industrial challenges, and showing how it enables scalable, efficient deployment of AI on the shopfloor.
The need for a domain-specific MLOps reference model
Contrary to traditional software engineering, ML-driven applications are usually developed in an explorative and iterative manner. With the proliferation of ML, specialized lifecycle models were developed. Despite nuanced differences between the workflows, a common view on the ML lifecycle is emerging. The typical ML lifecycle consists of the following four stages, while a continuous feedback loop allows data or model design to be refined based on operational insights and performance assessments [1]:
- Requirement stage: Defines objectives and problems, setting the foundation for data selection, model design, and system goals.
- Data-oriented stage: Involves collecting, cleaning, and preparing data, including feature engineering, annotation, and train-test-validation splitting.
- Model-oriented stage: Covers model selection, training, optimization (using validation data), and final evaluation on test data to assess performance.
- Operations stage: Focuses on deploying the model into production, integrating it into its environment, and continuously monitoring performance to address issues like model drift.
Implementing machine learning (ML) in manufacturing is a complex endeavor influenced by a range of technical and non-technical factors. From an integration standpoint, manufacturing environments are characterized by spatio-temporally distributed assets, heterogeneous communication protocols, diverse manufacturing processes, and the need to comply with potentially short takt times. These factors create significant challenges in terms of data collection, storage, and architectural integration, particularly when embedding ML models into existing systems.
On the modeling side, the diversity of use cases in smart manufacturing systems – which are expected to monitor, adapt, optimize, and operate autonomously – significantly increases the modeling effort required. This complexity is further exacerbated by issues related to data quality, quantity, and imbalance, as well as the prevalence of multimodal sensory data such as scalar values, process curves, time-series, images, and point clouds. Ensuring that this data is available in sufficient quantity and quality is critical to the performance of ML models. In addition, privacy concerns must be addressed throughout the data handling and modeling processes.
Once ML models are deployed in production, maintaining their performance becomes an ongoing challenge. Observability of the manufacturing environment is essential to detect changes caused by concept drift, process drift, human interventions, or the high variety and small batch sizes typical in modern manufacturing. These factors often necessitate the retraining and redeployment of models to prevent performance degradation.
Beyond technical challenges, non-technical issues also play a significant role. There is a notable shortage of skilled personnel capable of developing and managing ML systems, and the lack of interpretability in probabilistic, black-box models hinders trust and acceptance among stakeholders. Addressing these challenges requires a robust and scalable MLOps platform tailored to the specific demands of manufacturing systems. [2]
So far, most MLOps platforms are designed for use cases in finance, e-commerce, or SaaS-domains where systems are largely cloud-based, data is homogeneous, and retraining is straightforward. As depicted, manufacturing, on the other hand, presents entirely different challenges.
Industrial environments demand:
This complexity means manufacturers need a tailored reference model that considers these constraints and enabling to scale AI efficiently.
Key characteristics of a manufacturing-ready MLOps reference model:
Manufacturers need different tools than tech companies.
A modular architecture built for manufacturing
To achieve scalability, flexibility, and maintainability, the reference model is based on a modular architecture [3].
Each part of the MLOps lifecycle – data processing, training, inference, deployment, monitoring – is abstracted into components that can be reused and adapted for different use cases.
This architectural strategy includes:
This approach not only enables scaling across multiple factories or machines but also simplifies governance, auditability, and performance monitoring.
Architecture highlights:
The reference model is like a toolbox –
easy to combine, change, and reuse across machines and factories.
Enabling scalable AI in manufacturing
Our tensoryze core solution was created according to this MLOps reference model, empowering manufacturers to scale AI efficiently – beyond PoCs, beyond pilots, and into production.
tensoryze core adopts an engineering-first approach, leveraging established software architecture patterns such as modularity, separation of concerns, and service-oriented design to ensure robustness and scalability. Its architecture is highly configurable and modular, enabling seamless adaptation to heterogeneous manufacturing environments with varying infrastructure constraints. Built for automation, it supports declarative paradigms for deployment, configuration, and monitoring – facilitating reproducibility and integration with CI/CD pipelines. Designed with usability in mind, it minimizes the operational burden through intuitive interfaces and abstractions, reducing the need for deep MLOps expertise during deployment and maintenance phases.
What you get with tensoryze core:
Whether you’re starting with one use case or aiming to scale ML across multiple manufacturing lines,
our platform gives you the tools to succeed – faster and more reliably.
References
[1] M. Schlegel & K.-U. Sattler: “Management of Machine Learning Lifecycle Artifacts: A Survey” (2023)
[2] T. Raffin, T. Reichenstein, D. Klier, A. Kühl, J. Franke: “Qualitative assessment of the impact of manufacturing-specific influences on Machine Learning Operations” (2022)
[3] T. Raffin, T. Reichenstein, J. Werner, A. Kühl, J. Franke: “A reference architecture for the operationalization of machine learning models in manufacturing” (2022)
Want to start deploying your ML models?
Contact us now and get a demo!
Click here to change this text. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut elit tellus, luctus nec ullamcorper mattis, pulvinar dapibus leo.