How AI Is Reshaping Industrial Automation
By Dr. Tejasri Nampally
AI is changing the way industries understand, monitor and manage their operations. But unlike consumer applications, industrial AI operates in an environment where reliability, safety and consistency are essential. A poor recommendation on a shopping platform may be inconvenient. An incorrect decision in a chemical plant, power station or manufacturing facility could disrupt production, damage equipment or create serious safety risks.
This difference makes industrial automation a particularly demanding area for AI adoption. The goal is not simply to introduce more intelligent algorithms into factories. It is to use AI to improve industrial performance without compromising the control and reliability that these systems depend on.
This shift is already visible in how AI is being applied to industrial monitoring, engineering and process control.
1. From Fixed Rules to Learning from Plant Data
Industrial automation has traditionally relied on established control methods, engineering models and predefined operating rules. For example, a conventional controller continuously adjusts a valve to maintain the desired temperature or pressure. Its behaviour follows an established control strategy. AI brings an additional capability: learning from historical and real-time operating data.
Machine learning is already being applied to industrial problems such as:
Predictive maintenance: Identifying early signs of equipment deterioration before a breakdown occurs.
Anomaly detection: Recognizing unusual process behaviour that conventional alarm settings may not immediately capture.
Soft sensors: Estimating quantities that are difficult or expensive to measure continuously using available sensor readings.
Process optimization: Analyzing relationships between operating conditions, production quality, efficiency and energy consumption.
Rather than replacing existing control systems, these applications complement them by providing engineers with additional information to monitor processes, identify problems and improve operational decisions [1].
2. AI as an Engineering Assistant
A modern factory generates enormous amounts of information beyond sensor measurements. Maintenance records, alarm histories, shift reports, equipment manuals, operating procedures and engineering documentation all contain valuable knowledge about how a plant operates. However, this information is often spread across multiple systems.
Consider an engineer investigating repeated equipment failures. Identifying the cause may involve reviewing sensor trends, maintenance records and similar incidents from the past. AI-powered engineering assistants can help by bringing relevant information together, summarizing equipment histories and highlighting possible causes. Some industrial AI assistants also support engineering tasks such as interpreting technical documentation and generating automation code [2].
Research on industrial operator-support systems has explored how techniques such as deep learning, process mining and graph-based reasoning can help engineers analyze operational information and investigate potential problems. An application evaluated using operational data from an offshore oil field illustrates how AI can support industrial decision-making without replacing human expertise [3].
The important distinction is between assisting a decision and making a control decision. An AI assistant can help an engineer understand why a process is behaving unusually. That does not mean it should automatically change operating conditions without appropriate validation and safeguards. This human-centered approach is particularly important in industrial environments, where engineering experience and accountability remain critical.
3. Making Process Models More Adaptive
Industrial processes rarely operate under exactly the same conditions throughout their lifetime. Raw materials change, equipment ages, production requirements evolve and maintenance activities alter system behaviour. A process model developed using historical data may gradually become less accurate as these changes accumulate. For example, a model trained on the performance of a new industrial pump may become less reliable as the pump experiences wear and changing operating conditions. This gradual decline in model performance, often referred to as model drift, makes continuous monitoring and periodic validation important.
Industrial research is addressing how to detect changes in model performance and maintain reliability as operating conditions evolve. For example, a published patent [4] describes an approach for monitoring the quality of machine-learning models used in industrial applications. Such monitoring can help identify performance degradation and indicate when further investigation or model updates may be necessary. This highlights why industrial AI models need ongoing evaluation rather than relying only on their performance during initial development.
However, models that learn from data also have limitations. A model may perform accurately under normal operating conditions but produce unreliable predictions when it encounters conditions not represented in its training data. This is why high prediction accuracy alone is not enough for industrial AI. Models must also be reliable, adaptable and validated under changing process conditions.
Engineers must understand:
Whether the model is operating within conditions it has been tested on.
How confidently the model can make a prediction.
Whether its predictions agree with known physical behaviour.
How the system responds when the model is incorrect or unavailable.
The value of an adaptive model therefore depends not only on its ability to respond to changing conditions, but also on how reliably it can be monitored, validated and maintained throughout its operating life.
4. Foundation Models Are Entering Industrial Research
Foundation models have attracted significant attention because they can learn patterns from large and diverse datasets and apply that knowledge to different tasks. While language models primarily work with text, researchers are also developing foundation models for scientific and physical systems. One example is Walrus [5], a transformer-based foundation model designed to learn how physical systems evolve over time. It was trained on data from 19 different scenarios, including fluid dynamics, acoustics, plasma physics and geoscience. Its evaluation shows promising results in predicting the behaviour of different physical systems. This research is relevant to industrial automation because many industrial processes involve fluid movement, heat transfer and other physical interactions. However, applying such models to actual industrial plants requires further validation.
However, it is important to distinguish research progress from industrial deployment. A model performing well on research benchmarks has not necessarily demonstrated reliable operation in a real factory. Industrial plants contain equipment-specific configurations, ageing assets, incomplete measurements and operating conditions that may differ significantly from research datasets. Foundation models show promise for industrial applications, but their reliability and ability to work across different plants and operating conditions still need to be demonstrated.
5. AI is Becoming Part of Existing Automation Systems
An important development in industrial AI is how it is being integrated into existing automation architectures. Factories already depend on multiple layers of control, monitoring and protection. These include controllers that maintain process conditions, optimization systems that improve performance, alarms that alert operators and safety systems that protect equipment and personnel. AI is increasingly being developed to work within or alongside these established systems.
For example, a machine-learning model can estimate a process variable that is difficult to measure directly, identify early signs of equipment failure or recommend adjustments to improve performance. These predictions or recommendations can then be used within existing control and monitoring systems, subject to operational and safety constraints. Research is also exploring how reinforcement learning, where a model learns decision-making strategies through interaction with an environment, can support industrial process optimization and control [6]. However, safely integrating such methods into real industrial operations remains an important engineering challenge.
The direction is nevertheless clear: industrial AI is evolving as part of the automation ecosystem, not as a complete replacement for it. This means successful industrial AI must be measurable, maintainable, understandable and compatible with the engineering practices already used to operate complex facilities.
6. The Bigger Question: Who Owns the Data and Intelligence?
As AI becomes more closely integrated with industrial operations, an important question emerges: Who owns the knowledge that AI learns from a factory? Industrial AI models can learn from more than sensor measurements. Depending on the application, they may use equipment histories, maintenance records, process settings, production constraints and engineering decisions. Together, these sources represent valuable operational knowledge accumulated over years of running a factory.
Equipment manufacturers and automation suppliers have an important role in developing industrial AI. They bring valuable domain expertise, specialized engineering knowledge and experience across different industrial applications. However, when suppliers develop AI models using a factory's operational data but retain exclusive control over those models, the factory may become dependent on the supplier for maintaining, updating or transferring the intelligence learned from its own operations. [7].
For example, a supplier-controlled model may become central to predicting equipment failures or optimizing production. If the factory cannot independently access, validate, maintain or transfer that model, changing suppliers may become increasingly difficult. The concern is not simply about who stores the data. It is about who controls the intelligence developed from that data. A factory may retain its raw sensor measurements while losing practical control over the models, derived information and operational insights generated from them.
This creates several important questions for industrial organizations:
Who has access to operational data, and how can it be used?
Who owns models developed using plant-specific information?
Can the factory independently validate and maintain these models?
Can models and relevant information be transferred when changing suppliers?
Can a factory retain control over its operating knowledge while benefiting from external AI expertise?
Research on privacy-preserving machine learning and industrial data spaces is exploring ways for organizations to collaborate without giving up unrestricted access to their data [7, 8]. However, protecting data during sharing does not automatically resolve questions about who owns the resulting models, who can use them or whether they can be transferred to another platform. For factories, data governance is therefore not just an IT or contractual concern. It is also an important part of how industrial AI systems are designed, maintained and controlled.
Conclusion
AI is already contributing to industrial automation through predictive maintenance, anomaly detection, process optimization and engineering assistance. At the same time, research into adaptive models and foundation models is expanding the possibilities for industrial AI. However, the success of industrial AI should not be measured only by the sophistication or accuracy of its models. Reliability, transparency, maintainability and integration with existing engineering practices are equally important.
As equipment manufacturers and technology providers develop AI solutions using industrial data, factories must also consider who controls the models and knowledge created from their operations. Research on industrial data sovereignty highlights the importance of secure collaboration, clear access rights and the ability to exchange information and models across organizational boundaries. Ultimately, the goal is not simply to make factories more intelligent, but to make them more capable, resilient and independent.
As AI learns more about how factories operate, the intelligence it creates should remain under the meaningful control of the organizations that generate and depend on that knowledge.
What do you think should belong to a factory when an AI model learns from its operations: the data, the model, or both?
References
Lawrence, Nathan P., et al. "Machine learning for industrial sensing and control: A survey and practical perspective." Control Engineering Practice 145 (2024): 105841.
Moenks, Norbert, Pascal Penava, and Ricardo Buettner. "A systematic literature review of large language model applications in industry." IEEE Access (2025).
Plant Operator Support using Industrial Artificial Intelligence, ATP Magazine, 2023.
Method and Apparatus for Monitoring Machine Learning Models, US20230016668A1, 2023.
McCabe, Michael, et al. "Walrus: A cross-domain foundation model for continuum dynamics.", Proceedings of the 43rd International Conference on Machine Learning, PMLR, 2026.
Dogru, Oguzhan, et al. "Reinforcement learning in process industries: Review and perspective." IEEE/CAA Journal of Automatica Sinica 11.2 (2024): 283-300.
Ranathunga, Tharindu, Alan McGibney, and Sourabh Bharti. "Enabling secure and self-sovereign machine learning model exchange in manufacturing data spaces." Journal of Industrial Information Integration 42 (2024): 100733.
Möller, Frederik, et al. "Industrial data ecosystems and data spaces." Electronic Markets 34.1 (2024): 41.
Acknowledgements
Generative-AI tool use
Generative AI tools were used solely for language editing and text refinement. The ideas and perspectives presented in this blog are the author’s own and are informed by the references cited in the text.
Disclaimer
This blog is a contribution of expertise from our volunteers. It is not a reflection of any opinion or roadmap of their employers. All blogs from WAI Labs go through a review and/or editing as needed, and are vetted for veracity. For questions or comments, write to wailabs@womeninai.co .
Comments