Production of the Future

Principles of Software-Defined Automation

Given the growing complexity of production processes and labor shortages (in Germany alone, shortages of engineers and IT specialists cost an estimated 13 billion Euro annually [2]), it is increasingly important to efficiently network machines and systems, integrate AI-supported services, and create transparency across all corporate systems. Traditional automation technologies are reaching their limits, even when hardware, such as programmable logic controllers (PLC), is virtualized [5].

Decoupling Hardware and Process Control

As with software-defined networks (SDN), software-defined automation abstracts hardware, and production capabilities are controlled exclusively through software [4]. Decoupling hardware and software significantly increases flexibility. For instance, a welding robot only needs to know how to weld; the exact place of the welding point can be determined dynamically via IT services. Tasks such as welding, screwing, or transporting can be orchestrated flexibly – independent of the specific machine and without specialized on-site personnel [5].

Implementing IT Security

Software-defined automation creates a flexible network of IT services, replacing the rigid automation pyramid. When implementing this, the necessary cybersecurity protections must be addressed rigorously. Traditionally, production environments have been largely self-contained, with IT security applied selectively. Thus, additional exposure, such as that caused by introducing software-defined automation, increases vulnerability within production [6].

Enforcing an organization‘s security policies centrally is a fundamental requirement for operational resilience [7]. The convergence of IT and operational technology (OT) that comes with software-defined automation also presents the opportunity to unify and standardize security measures across organizations and beyond [5]. This approach supports unified system governance with significantly lower system administration costs [8]. The EU’s NIS2 (Network and Information Security) directive will play an increasingly important role here [9]. Although it is primarily aimed at IT operations and management, it can also be used to improve security measures in companies in general. Thus, it can be assumed that the NIS2 will seriously impact the OT in production too.

Digital Twins at the Core

Digital twins form the backbone of software-defined automation. They are virtual representations of real objects, maintaining live synchronization with their physical counterparts (Figure 1). This bidirectional connection allows real-time monitoring and control of production processes [10]. Digital twins are indispensable for applications like condition monitoring, prescriptive maintenance, as well as anomaly detection and safety assurance, enabling efficient production control, whether via cloud or edge computing.

Figure 1

How ChatGPT might envision itself within a production setting if it had a physical form

Digital twins offer the capability to interact directly with AI. They can provide precise, context-based data from the production process, enabling AI to make accurate decisions [5]. Suggested process changes and optimizations can then be implemented in actual production through digital twins [4].

AI in Production

Introducing AI into production promises significant efficiency gains and innovation (Figure 1). However, successful AI integration requires a deep understanding of the technology and a targeted strategy that considers economic, technical, organizational, and cultural aspects. The following outline key requirements and steps for companies starting AI implementations.

Categorization and Fundamentals

AI refers to systems that independently learn from data through algorithms, handling cognitive tasks such as problem-solving and decision-making. Unlike Large Language Models (LLMs), which primarily focus on language-based tasks [11], a realization of an AI application essentially comprises three steps: Systematic data collection, model training, and integration into the production environment. You can find a more detailed breakdown in [12].

Implementing AI requires technical and domain-specific knowledge to adapt AI solutions to production requirements. Additionally, it should be considered to provide structured data in the process context rather than the mere provision of semantically poor data [13]. The „Periodic Table of AI“ offers helpful guidance on industry application areas [14].

Successful AI Implementation

For companies beginning with AI, proof of concepts (PoC) can be used to test specific applications. The PoC approach aims to develop use cases with minimal effort and evaluate data availability and potential business benefits. Furthermore, the PoC phase reveals whether the project team possesses the necessary skills and how employees can be integrated into the process. Employee involvement is key, as close collaboration between the AI team and specialist departments is a foundation for success.

After a successful PoC phase, companies need to plan for the operation and scalability of AI solutions. Important questions include whether to use cloud-based or on-premises solutions, open or closed source, and whether to develop in-house or rely on external expertise [15]. Cost management is also crucial, especially with cloud solutions, where ongoing costs vary. In the long term, it is essential to establish mechanisms for maintaining and updating machine learning models to ensure consistent performance. Figure 2 presents this process in a flowchart.

Figure 2

Flowchart for successful AI project implementation

Technical, Organizational, and Legal Constraints

Implementing AI in production is complex, involving numerous technical and organizational requirements. Verifying data availability and quality is essential. High data quality is crucial, as poor data can undermine model accuracy and jeopardize AI project success. Transparent data management and maintenance standards are needed to ensure a stable and reliable system. Data Mesh is a particularly effective standard for integrating AI systems into companies [16]. Risk management strategies also help to identify and address common pitfalls such as bias or overfitting.

Another key aspect is data protection, compliance with legal regulations, and the previously discussed cybersecurity measures. When handling personal or sensitive data, complying with the EU’s GDPR (General Data Protection Regulation) is crucial. Wherever possible, data should be hosted within the EU. The EU AI Act also provides valuable guidance on legal risks [17].

Finally, transparent internal decisionmaking in AI models is essential to ensure traceability of errors and foster employee acceptance. Considerations around sustainability and ethics, such as efficient resource use and privacy protection, are also gaining importance as they reflect the company’s social responsibility.

Quality Control and Process Agility

AI visual inspection and quality control applications reduce human errors and improve precision [18]. Image processing algorithms quickly and reliably detect defects, reducing waste and production costs – a considerable advantage in precision-focused industries such as automotive or electronics and pharmaceutical or food industries. Automated and digitally available results shorten processing times and increase efficiency.

In addition to visual inspection, supplementary sensors can be used to monitor processes more comprehensively. By combining data sources, AI gains more profound insights into process quality. For instance, inaccuracies from earlier process steps can be compensated for in subsequent steps [19].

Regarding this, software-defined automation provides the mechanisms to gather and communicate current and historical production-related data to AI. It provides the mechanisms to take the related feedback from AI for reconfiguring processes or providing actionable recommendations to personnel. In [5], a related application is described, where quality-related process information is gathered from a Body-in-White line and fed to the associated algorithms. The results are then applied at the line.

Anomaly Detection and Safety

AI will play an increasingly crucial role in anomaly detection in the highly regulated process industry to prevent production disruptions and accidents. As process complexity grows and unforeseen events become more frequent, AI algorithms demonstrate advantages over conventional analytics [4, 20]. This is particularly important in sectors with stringent safety and environmental regulations, where accidents can endanger lives and cause significant environmental and financial damage.

Real-time sensor data from software-defined automation enables AI to identify non-linear patterns indicating potential leaks, pressure spikes, or other hazardous states and recommend appropriate countermeasures. Software-defined automation can then directly implement these measures. In [21], more details on AI-based anomaly detection are given, and a case study is introduced. Meaningful publications from the industry currently seem to be kept secret to protect competitive advantage.

A further relevant aspect is „industrial aging“: Aging equipment poses additional challenges, which AI addresses by predicting aging behavior and potential partial failures [22]. Software-defined automation can communicate context-specific data (current and historical sensor data, maintenance information, etc.) to the AI and execute the optimizations suggested by the AI.

Prescriptive Maintenance and Self-Healing

Predictive and prescriptive maintenance are essential for the manufacturing and process industries [5]. Unplanned downtime can be costly in continuous production environments, such as refineries or pharmaceutical plants. A single minute of downtime in the automotive industry can cost up to 20,000 euros. Combining AI and software-defined automation provides substantial advantages over traditional automation technologies.

As depicted in Figure 3, AI and software-defined automation go far beyond traditional automation technologies’ possibilities. Implementation requires deep expertise, a trained AI model, data from current and past production runs, and the integration of relevant systems (e. g., maintenance, ERP, MES). It also requires the ability to influence the production process actively.

Figure 3

Unlocking new possibilities in maintenance through AI

There are several stages of implementation. AI generally analyzes sensor data from machines and systems, e. g., to predict failures or wear before they occur. The stages differ in terms of what additional information is thereby considered by the AI and how the results are used in production actively.

Prescriptive maintenance incorporates dynamic strategies for specific aging and degradation behaviors. Based on this information, AI suggests targeted maintenance actions to ensure system reliability, allowing for more flexible and efficient maintenance planning than traditional static strategies [23]. Whereas related realizations can be found frequently in industrial applications (e.g. [5]), the following two can be seen as promising future trends. A review and a case study regarding the necessary self-reconfiguration capabilities are given in [24].

Proactive maintenance integrates maintenance and logistics knowledge, dynamically steering production to optimize maintenance timing. Self-healing advances this process by automatically detecting and resolving disruptions through corrective actions, process adjustments, or alternative logistics without manual intervention. These approaches require a bi-directional interaction between AI and software-defined automation.

Process Optimization, Supply Chains, and Sustainability

Machine-learning algorithms within software-defined automation enable real-time adjustments to process parameters to improve efficiency, product quality, and energy savings. In the process industry, AI-supported systems help reduce raw material consumption and maximize energy efficiency, reducing operational costs and the ecological footprint [25]. Similar benefits are observed in the manufacturing sector. A notable example is the publicly funded project E-KISS, which achieved nearly 30 percent savings in resources and energy [4].

AI can also optimize complex supply chains and enhance production processes. It improves demand forecasting, optimizes inventory, and enables flexible production adjustments. Software-defined automation provides the mechanisms for creating agile production and logistics environments.

New regulatory requirements further increase the need for real-time data availability. Real-time data, made available through software-defined automation services, can populate data spaces like Catena-X or Manufacturing-X in a structured and production-accompanying manner. In the future, such services and appropriately developed AI may also be used to achieve a global minimum of CO₂ emissions for specific products while simultaneously providing data relevant to future product passports.

Copilots in Production Planning

The application of AI delivers operational advantages in production and substantial benefits in production planning. AI-based copilots support planners by accessing experiential knowledge from past projects, providing helpful recommendations and handling repetitive tasks, allowing planners to focus on strategic decisions.

Production planning is a complex process that cannot be fully captured solely through the reasoning capabilities of a single language model [26]. However, it benefits significantly from enhanced reasoning abilities, enabling precise and efficient planning. The close integration of AI with cross-domain models creates a foundation on which changes in one domain can be automatically reflected in others. Figure 4 illustrates an example architecture for an agent-based logic system in production planning. This architecture incorporates vector embeddings to capture the nuanced relationships between different planning concepts and uses multiple specialized agents to handle various aspects of the complex planning process. It also includes a feedback loop using Reinforcement Learning from Artificial Intelligence Feedback (RLAIF) [27], which further sharpens the reasoning of the agent-based logic. The architecture was developed and successfully validated in a prototype, and an industry-ready application is in progress.

Figure 4

Architecture of an LLM-based agent system in production planning

Active AI involvement can be structured so that a plan serves as a final output and seamless input for subsequent phases. When results—from engineering outputs, through production planning and virtual commissioning, to the services of software-defined automation—are bidirectionally interconnected, a well-trained AI system can ensure that changes in one domain are automatically updated in others. Such a solution could lead to an unprecedented efficiency boost in production.