A Computer Science Perspective on Digital Transformation in Production

2.1 Digital Twins vs. Digital Shadow

To realize the IoP, we suggest Digital Twins, which digitally represent material [15, 24] and immaterial [85, 95] objects and processes of the real world. The challenge here is the integration of the different levels of scale (temporal, spatial, etc.) of the numerous underlying processes, yielding large amounts of data, ill-fitted models, and high latencies if data needs to be aggregated and analyzed. There exist various platforms and approaches to realize Digital Twins [8, 77, 94, 106, 143] or to establish the connection between IoT and Digital Twins, e.g., model-driven approaches for interface generation [81] or the H2020 funded IoTwins Innovation Action project [11], which aims to design a reference architecture for distributed and edge-enabled twins and its evaluation in several industrial test beds.

We do not consider a complete Digital Twin to be feasible due to the massive amounts of data that a virtual replica of a product, machine, or production plant would require. Also, the Digital Twins that are used in practice are not complete digital counterparts of physical objects; rather, they are collections of different datasets and models, each representing a particular aspect of the real object. The datasets are collected for a specific purpose, e.g., sensor data for prediction, CAD models for simulation. To model this scenario more exactly, our vision focuses on Digital Shadows, which we consider as task- and context-dependent, purpose-driven, aggregated, and persistent datasets that encompass a complex reality from multiple perspectives in a more compact fashion and with better performance than a fully integrated Digital Twin (cf. Section 4). A Digital Shadow can be compared to a view in database systems: an aggregated subset of the data of the real object, computed by a complex function that might include complex algorithms for data reduction and analysis.

We have already proposed a conceptual model [10] to describe digital shadows and demonstrate it using a concrete example. The conceptual model was established through interdisciplinary research and intensive discussions and was evaluated in various real-world manufacturing scenarios. It is a foundation to manage complexity, automated analyses, and syntheses, and, ultimately, facilitates cross-domain collaboration. For a better understanding on how Digital Shadows could be used within Digital Twins, we refer the reader to a dedicated example [22].

2.2 Comparison to State-of-the-Art

In comparison to existing approaches, the IoP provides a holistic, cross-domain, and collaborative perspective on manufacturing processes. Existing approaches can be categorized into the following areas: Concrete technologies such as the classical Internet, cloud manufacturing and IIoT, business demonstrators and digital transformation strategies such as service-oriented manufacturing approaches, digital manufacturing or the Global Lighthouse Network, and politically and funding-driven approaches such as Industry 4.0 and other national initiatives.

In comparison to the Classical Internet, the IoP offers more functionality to a restricted group of stakeholders. It grants access to world-wide production-focused knowledge bases, provides aggregated and semantically enriched data from production processes, includes intelligent algorithms and functionalities to support question-solving, and enables users to analyze their own data. These functionalities are available to all stakeholders within production processes, such as employees of different companies of the production network within all levels of work, e.g., the workforce, the quality assurance team, marketing experts, production planning, suppliers, or logistics partners.

Certainly, comparing the IoP to the Classical Internet is a fitting analogy given that the Internet itself started as a lab of labs, an aspect that the IoP, in turn, picks up with the concept of the WWL. Moreover, today, the Classical Internet serves as an archetype: It revolutionized networking with its novel idea of packet switching. Other approaches followed up on this idea. For example, in the domain of logistics, the concept of the Physical Internet [7, 109] also relies on the Internet analogy, envisioning to establish a physical form of “packet” switching in logistics. In contrast to the Physical Internet with its cargo and product flows, the IoP itself intends to establish a “knowledge” switching for the manufacturing industry by sourcing information and Digital Shadows from various stakeholders and across domains, with the goal of a sustainable and effective digital transformation.

In relationship to Industry 4.0, the IoP can be seen as a concrete initiative to realize aspects of these strategies in cooperation between research and industry. This includes the integration of digitized CPPSs with their processes and stakeholders to optimize the complete value-added chain. However, this idea is also relevant internationally within the US Advanced Manufacturing Initiative [108], the Chinese Made in China 2025 strategy [101], the Japanese Industrial Value Chain Initiative [71], the South Korean Manufacturing 3.0 [72], and the UK national Catapult research center on High Value Manufacturing [30].

The Global Lighthouse Network [151] is an initiative of the World Economic Forum. The initiative was launched because of the global manufacturing industry’s lag in adopting Industry 4.0 technologies. It has essentially the same goals as the IoP, i.e., namely a move toward globally networked production. However, it focuses more on the management perspective, whereas the IoP develops the necessary technical foundations to achieve this goal.

Approaches such as Digital Manufacturing [107] as part of the Fourth Industrial Revolution [136] refer to the digital transformation of production processes using smart and agile manufacturing and smart factories together with digital manufacturing technologies such as additive manufacturing (3D printing), laser cutting, and CNC processes. In contrast, the IoP has no limit regarding specific manufacturing technologies, and integrates the whole value chain.

IoP and IIoT. General efforts, such as the IIoT or Industry 4.0, typically focus on enabling communication only within the same company [36]. In contrast, our vision of the IoP does not merely intend to enable communication between different companies, but also aims to realize a new level of cross-domain collaboration by enabling the exchange of semantically adequate and context-aware data whenever and wherever it is needed [121].

Cloud Manufacturing is “a service-oriented business model to share manufacturing capabilities and resources on a cloud platform” [48], which encapsulates distributed resources into cloud services, and allows for their integrated management [153]. According to Siderska and Jadaan [139], cloud manufacturing focuses on inter-factory integration, whereas Industry 4.0 also considers intra-factory integration. The IoP shares the Industry 4.0 idea of intra-factory and inter-factory integration and it does not restrict its technological approach to only one technology such as cloud applications [48], and service-oriented architectures [134]. Cloud manufacturing platforms can be built for small-medium size enterprises or group enterprises [97]. In contrast to that, the IoP is not limited to one enterprise or a group of enterprises.

Service-oriented Manufacturing integrates services and physical products into one product service system and companies involved focus typically on a specific sector [53]. Research in this area focuses on the business perspective and pricing strategies [154], not the technological and computer science perspective.

In summary, there are already approaches towards the digital transformation of production. They either focus on singular aspects of digital production, take a management perspective without sufficiently resolving the technical challenges, or lack strategies to enable collaboration across company boundaries. As outlined in the next sections, the IoP integrates these different approaches holistically and provides concepts and method to achieve this ambitious goal.

General Challenges that the aforementioned global initiatives are proposing to solve by integrating manufacturing and IT are of societal and political nature. The proposed solutions have partially overlapping issues, but each also has its strategic foci and thus associated challenges, some of which we outline here. Digital Manufacturing, which aims to integrate digital manufacturing methods into production processes, leads to typical issues in human-robot collaboration like unforeseen events that robots cannot handle [107]. Data availability for monitoring and control, as well as its management and networking are further challenges. Cloud Manufacturing leads to two major challenges [139]: First, general integration issues of cloud computing, IoT, and high-performance computing exist. Second, technical issues such as cloud management engines and visualization in cloud environments surface. For service-oriented manufacturing, Gao et al. [53] discuss challenges in the cooperation between businesses and adaptations of business models for outsourcing parts of the value chain as services. They see service-oriented manufacturing as “innovation from the perspectives of business model, industry insight, and technology advantages”.

In Section 6, we specifically describe the challenges involved in realizing the IoP. We categorize these into four layers. In the Outlook, we point out challenges that go beyond these layers, such as implications for business models.

6.1 Humane Interfaces for Interacting with Digital Shadows

The new possibilities of the IoP have given rise to new questions in the area of designing the interfaces between the human actors and the IoP that have so far been insufficiently addressed in current research [78]. On the one hand, more and smarter automation raises the question of responsibility and control [93, 133], and on the other hand, new forms of hybrid intelligence as the collaboration between human operators and the IoP must be designed that harness the potentials of both artificial and human intelligence [44, 89].

Despite the obvious potential of increasingly automated production control through IoP-based Digital Shadows, people will remain an integral component of socio-technical production systems (STPS) [50, 78]: Either as certain tasks cannot be fully automated for technical, legal, or ethical constraints, because of a shift from manual activities to monitoring and planning tasks, or as a final arbitrator when automated systems fail or come into conflict [150].

However, reliable automation leads to the automation conundrum [46]: The more systems are automated, and the higher the performance of the automation, the lower the supervisors’ situational awareness and the more difficult supervision, intervention, or manual control becomes. Thus, several challenges need to be addressed to support operators’ interaction with Digital Shadow-based automation at all company levels (e.g., shop floor operation, factory planning, supply chain management, and strategic planning).

Transparent Automation and Meaningful Control. Operators’ and decision makers’ process knowledge and understanding deteriorate through abstraction and automation [6, 9, 46, 150]. However, this situational awareness is crucial should automation fail, to evaluate the functioning of an automated system, or to handle unmodeled situations (out-of-the-loop loss of situational awareness). Consequently, a challenge is to design simple interfaces to automated processes and decision-support systems that are accessible, transparent, and easy to learn and use.

Although one of our intended interfaces is as simple as an Internet search engine (cf. Figure 1), processing these queries is more sophisticated than a simple keyword search. As users should not need deep technical knowledge about the underlying system, autonomous agents interacting on the WWL can provide this semantic information implicitly. For example, when inquiring about the ideal type of an electric car battery, the planned production method or driving safety is considered implicitly. Other forms of automation can draw on transparent explanatory approaches so that the basis for the system’s decisions can be interpreted, understood, and corrected if necessary. An emerging research field to increase usability and comprehensibility of models is Explainable AI [4], whose approaches and methods must be adapted to the specific use cases of the production domain.

Modeling of Tacit Knowledge. A further challenge is the utilization of human expertise through automated systems. Most machine learning approaches require representations of human knowledge to create digital models of product and production planning, as well as production and usage. In some cases, the description of this knowledge is simple and often already exists (e.g., image classification for online quality control, if quality can be measured easily). In other cases, a representation of the expert knowledge is necessary, but this tacit knowledge is difficult to verbalize and hidden in unconscious evaluations and motor memory [126]. Consequently, it is difficult to communicate this knowledge and expertise to others and other domains, to record and describe this knowledge digitally, and to use sparse data to train AI algorithms [89].

Bias-Free Interaction with Automated Systems. Trust, reliance, and trustworthiness is a crucial prerequisite for acceptance and use of automation in production and other domains [65]. Adequate and meaningful use of automation by operators must be carefully balanced between disuse (intentional neglect of decision aids, either due to missing trust or missing perceived benefits) and misuse (over-utilization of automation by over-trust and neglecting to check its results) [65, 114]. This fine balance relates to automation biases and automation complacency, and sound systems design helps [19, 58]: if automated systems are designed right, operators have more capacity to detect malfunctioning automation and to handle exceptions.

Context- and User-Centered Interfaces. Third, a challenge is to make the vast amount of heterogeneous information from the IoP transparent and accessible through user-, context-, and task-dependent interfaces [1, 29]. Again, the design of STPSs and interface usability are important factors, as good interfaces facilitate the understanding of the systems’ status and functioning, lower cognitive load, enable successful operation, and offer insights on further optimizations. Furthermore, we need a “natural and trusted” communicative etiquette for enabling a close and trusted collaboration between the operators and the AI-based systems. For this, understanding the operators’ basic emotional needs as well as their mental models of and general attitudes towards these highly complex systems is of importance. Furthermore, the demographic shift and changes in the workforce pose further challenges, as user interfaces and support systems must take older workers and their specific requirements, different skill-sets, interests, and abilities into account [34, 47].

In prior work, we have shown that good user interface design is crucial for successful and trustful interaction with automated production systems [125] and robots interacting closely with humans [14]. Well-designed user interfaces mitigate automation biases by enabling operators to intervene should automation fail [19].

Example. Taking the production of an e-vehicle as an example, the IoP results in changes for workers along the value chain. Through human-centered design of decision dashboards and approaches such as Explainable AI, decision support systems can provide transparent suggestions for improving the performance of production processes, quality insurance, or the supply chain, thus reducing human errors in decision making [19]. Also, by continuously capturing the interactions of experienced workers with the production systems, their knowledge can be integrated to improve future recommendations for novices. Furthermore, digital images of the workers’ capabilities and requirements can be generated and used to orchestrate the collaboration between production systems and workers, for example, by adopting the speed of production process and human-robot collaboration to the workers’ needs [102].

Takeaway. Overall, the adequacy of these systems and their design should not be determined by experts from engineering, computer science, or ethics alone, but rather in partnership with the employees. A participatory design ensures that the technological advances and implementations of STPSs are harmonized with people’s capabilities, norms, and values.

6.2 Model-Integrated Artificial Intelligence with Autonomous Agents in the IoP

Within the IoP, the goal is to create a synergy between data-driven AI methods and state-of-the-art model order reduction techniques from engineering mathematics across disciplinary boundaries. This enables a high level of automation to realize real-time decisions, through built and shared Digital Shadows. E.g., in e-vehicle manufacturing, sheet metal is still a fundamental material whose processing consumes large amounts of energy, while deviations are safety-critical. Therefore, integrating reduced engineering models of material properties with machine learning in the hot rolling process to inform artificial networks as Digital Shadows enables real-time compensations for deviations during the process [103]. This adjustment allows significant energy savings. For this degree of automation, which is needed to gain all data and knowledge from different sources and domains worldwide to build Digital Shadows and realize model-integrated AI, we need autonomous agents [131] based on various AI techniques like knowledge-based systems or machine learning to name but a few. Manual data queries that would otherwise be required would be infeasible given the level of cross-domain collaboration and networking.

Another example in e-vehicle manufacturing covers sophisticated logistics robots used in modern modularized factories without assembly lines [25]. In such complex settings, sophisticated logistics robots controlled by autonomous agents can help to assemble products, integrating different AI methods [66, 67]. Similar methods were used for autonomous agents communicating as programs with the WWW to realize Semantic Web applications [100, 127]. In addition, further examples of successful applications of agent technologies in industrial settings exist [90, 91].

In the IoP, we develop autonomous agents, called WWL Agents [21], in a multi-agent network connected to information sources for semantic information, e.g., ontologies [99] and knowledge graphs providing provenance information [57]. With the latter, it is possible to get, e.g., the origin of data used for training an artificial neural network representing a particular Digital Shadow or the usage history of mathematical models comprised within the Digital Shadow to solve difficult production steps.

Comparable agent systems in the literature are very often only used to support manufacturing processes within a single production facility or company [75, 90, 91]. The novelty of our approach is that the purpose of WWL Agents is realizing interoperability in the WWL and breaking data silos enabling data-intensive AI approaches, such as machine learning, to generate specific Digital Shadows. For instance, communicating with agents from other companies, apply different AI methods for sharing, generating, and using their data and Digital Shadows from different production domains is their main function. Furthermore, with semantic information about the origin of the data, they can provide detailed information of solutions found by a WWL Agent to users around the world. In addition, this approach can incorporate existing local multi-agent production systems if they provide an interface to the WWL.

Explainable AI. To be able to present comprehensible results for humans as discussed above, algorithms need to be able to explain why their results are reasonable and accurate. Such explainable AI methods [2, 45] can be provided by knowledge-based systems because they represent the knowledge in a human-understandable way [17]. However, one of the challenges is to find the right explanations and their representation for production processes. Furthermore, other AI approaches as, e.g., machine learning or process mining [3], need to be included in trustworthy explanations.

Furthermore, in the IoP, semantic information about the data in the WWL is available to calculate appropriate answers. Given that the IoP presents a well-understood domain [20], the semantic information is already represented in existing models and methods. However, it is often not machine-readable and not yet linked to the raw machine-produced sensor data, which need to be exploited by software agents. Open questions are how to gain machine-readable semantic information from the existing models and how to link the semantic information from the engineering models to the data to build trustworthy software agents [141] dealing with the semantics in Digital Shadows and supporting the decision process of engineers in the WWL.

WWL Agent Dialogs. In the vision of the IoP, WWL Agents realize user dialogs via interfaces like shown, e.g., in Figure 1, and help the users to get the appropriate answers to their problem in the sense that it improves the product and the production processes. These agents need to be interconnected within the WWL to enable them to integrate information from external data sources as well. With this information, the agents are able to compute solutions or suggestions and generate answers to the user requests. Here, the challenges lie in realizing a human-machine communication that is understandable for the humans working in the production. Additionally, the relevant information from the WWL has to be identified to give adequate answers, which really lead to process improvements.

Multi-Agent Network. We claim that only the WWL delivers the amount and variation of data that is necessary to build Digital Shadows as, e.g., trained artificial neural networks, so that they can provide aids for specific purposes in a production process. Therefore, WWL Agents in a world-wide multi-agent network are a key factor in reaching this level of interoperability.

By communicating via the WWL, the agents share services, data, and knowledge, so that other agents can support their local clients with their production. In that way, it is possible, for instance, to realize a fully automated on-demand pull production [70], with implications along the whole supply chain, as long as every participant is connected to the WWL. Thus, production can dynamically adapt to local or global changes, such as product design modifications, local production failures, or supply chain variances in case of a strike, natural disaster, or pandemic. The challenges range from finding the appropriate network structure for the agents to how requests to agents are processed.

Standardized and Open Communication Protocols. Similar to the WWW, the WWL can only unfold its full function if a crucial number of participants is able to share their information in the network. The success of the WWW was only possible because there was an open access to all its protocols as HTTP. Therefore, for the IoP we intend to let the autonomous agents use protocols based on HTTP and other open standards. Furthermore, new protocols which are needed for the communication between the autonomous agents have to be freely available and standardized in the long run to let everyone participate in the WWL with their own agent.

Connection to Versioned Data Storages and Ontologies. The agents in the WWL have to be connected to various information sources to gain the knowledge they need to give proper answers to the user. For doing so, the agents need semantic information about the posed queries. Therefore, we want to use semantic techniques, for example, ontologies [116] and versioned knowledge graphs providing provenance information about data, models, and knowledge from production processes [57]. With the former, the agents gain semantic information about the terms which are used in user requests. With the latter, the agent can get provenance information about the origin of production data or the history of a product part. Here, challenges are, for instance, how the semantic information can be used and how the information from different data sources can be combined to improve production processes.

Example. Assume a setting where an e-vehicle manufacturer wants to improve the material choice for the main car body. Then the engineer, e.g., instructs a WWL Agent to first collect information about available steel composites from different steel producers and then analyze them according to the characteristics of the local production process as described by a Digital Shadow. This Digital Shadow could be created by another agent in the WWL analyzing the processes in the car factory and using AI methods such as machine learning. Provided with semantic information about the processes and the materials which are planned to be used, a WWL Agent can give the user a detailed answer with explanations and links to the provenance of the agent’s information.

Takeaway. The integration of (mathematical) models from engineering with AI models enables new opportunities, and makes WWL Agents the enabling factor in the IoP. To this end, they rely on the connection to various data sources from different production companies, and standardized protocols. Existing knowledge-based agent technology needs further extensions to meet the demands of the manufacturing industry in the setting of the IoP.

6.3 Model-Driven Digital Shadows

The term model-driven refers to development methodologies that rely on abstract models of systems as central development artifacts [148]. These models carry explicit domain expertise and serve as a foundation for communication, documentation, analysis, and synthesis in agile development projects [129]. They can be systematically transformed into concrete implementations [49] such as Digital Twins [13, 43], privacy-preserving IoT systems [104], information systems [54], or assistive systems [105]. Digital shadows [88, 128, 135] relate to models, can carry models themselves, and serve as (aggregated) abstractions of models for automated processing [13, 43].

Cross-Domain Collaboration. Interdisciplinary teams consisting of experts from the production domain, computer science, automation, and many more develop a new generation of CPPSs. All of them contribute individual expertise, perspectives, paradigms, technologies, and solutions to the IoP. And often, this expertise is encoded in different kinds of models [69, 105, 152]. The successful and efficient integration of domain-specific knowledge into the IoP is crucial to construct the multi-perspective data and models at design time, simulation time, and run time. In our vision, Digital Shadows also serve to semantically enrich process data to enable (automated) decision making in (domain-specific) real time. To this end, they must be semantically integrated with data and models engineered during design and simulation [80]. This need demands a modeling of detailed aspects (from manufacturing system details to factory behavior, to strategic goals, to interface descriptions) in sufficiently formal languages [129, 130].

Systems Engineering. Model-driven systems engineering [12, 42] lifts models to primary development artifacts that increase abstraction and engineering efficiency in the interdisciplinary engineering of cyber-physical (production) systems. These models usually conform to (domain-specific) modeling languages (such as Simulink [37], SysML [51], or AutomationML [98]), that provide experts with required functionality and facilitate describing and integrating systems engineering concerns. Consequently, the system description is distributed over several models and tools that currently are not syntactically and semantically integrated. Designing and engineering the systems of the IoP, therefore, demands novel solutions for the automated, ad-hoc integration of modeling languages and their tools, e.g., as presented by Dalibor et al. [41], such that experts of the different domains can leverage modeling views tailored to their desired level of abstraction across domain boundaries and optimized for analysis. Software Language Engineering (SLE) [68] is a discipline that investigates the efficient engineering and integration of heterogeneous modeling languages; hence SLE is a crucial prerequisite for providing domain-specific representations and integrating knowledge from various domains within the IoP.

Integrating Modeling Languages and Tools. This model and language diversity [152] will also be reflected in Digital Shadows that need to provide optimized structures for handling large amounts of data, selected engineering models, formalized knowledge about data, models, and their context. To this end, Digital Shadows need to be able to integrate high-volume structured and unstructured data with semantically rich, detailed engineering models, and knowledge bases. Thus, we need techniques capable of linking the different underlying modeling languages [27, 138] at system design time as well as ad-hoc at runtime. These techniques consider syntactic [62] and semantic integration [33] to bridge semantic gaps between the languages used to express parts of the Digital Shadows, as, e.g., realized for the design of experiments in injection molding [13]. A first concept on how Digital Shadows can be created and are handled during runtime together with process mining techniques has been derived [22]. Efficiency is crucial to also enable an efficient application to very large models. Therefore, research should leverage techniques from database schema modeling [76] and artificial intelligence to enable compositional mechanisms [28] for syntactic and semantic abstraction, aggregation, and integration of data and models.

Example. Domain experts from e-vehicle production define one or more Digital Shadow Types [22] based on a conceptual model [10] for purposes related to production, e.g., quality monitoring or predictive maintenance. These types can be used as blueprints for concrete Digital Shadows recorded from production data at runtime, e.g., a Digital Shadow Type might serve the purpose to minimize the product rejection rate of the grinding process of front window panes and capture the related information accordingly. During runtime, the Digital Shadows are created according to their types and populated with models, data, and metadata. Using such Digital Shadows, the rejection rates from every window pane are aggregated to every job on a grinding machine. Periodically, a new Digital Shadow is created that aggregates again the rejection rates based on the new time slice.

Takeaway. Purpose-driven Digital Shadows can be created and provided at runtime using design time models; thus, model-driven development supports and simplifies the automation of production.

6.4 Interconnected and Industry-Capable Infrastructure

The concept of Digital Shadows is based on the notion that a problem-specific view on the overall process can be derived from a sufficient amount of process-related data. To this end, process data needs to be recorded and collected, ideally in a fine-grained manner and by a variety of different sensors, to provide a comprehensive description of the process. Subsequently, this description can be scaled down to match the concrete requirements of a specific problem or task. In general, the quality of the Digital Shadows correlates with the quality and the richness of the available data, i.e., larger amounts of data are generally favorable. Companies are thus incentivized to collect, process, and store huge volumes and varieties of data, which consequently requires a capable infrastructure. Setting up this infrastructure and enabling the global WWL and its models introduce several obstacles.

Data Integration. Availability and accessibility of information with high data quality is an important issue in many production and business processes. For example, the quality assurance of production companies could require access to detailed process data a long time after the production is finished, which cannot be realized by traditional data integration approaches that use carefully engineered data processing workflows to extract, transform, and load data into an integrated data store. For the collection of data, we envision a data lake platform in which data is stored in its raw format without prior integration or aggregation [61, 74]. Compression techniques on sensor data could be applied to address the real-time requirements by reducing the amount of data, but a lossless compression should be guaranteed. For example, in a use case of Laser Powder Bed Fusion, in which high power-density lasers are used to melt and fuse metallic powders, we apply dynamic compression techniques to reduce the data volume, but to maintain the information content.

The data lake stores raw data to avoid restricting the data analysis to a predefined integrated schema. The data in the lake should be enriched with semantic metadata and data quality information (e.g., source, accuracy, and time) to make it interpretable and usable in various applications.

Data Collection. As the data lake is intended to collect information from many sources in the WWL, the underlying infrastructure must be able to transfer very large amounts of data. This requirement is independent of whether the data lake is deployed on-premise or in the cloud.

We already identified that cumulative data rates for a single production cell can easily be in the range of giga- to peta-bytes for settings where several machines are interconnected [55]. Directly transmitting all data is thus often infeasible as available data rates are too low. The current bandwidth limitations dictate pre-processing and aggregation to make the WWL possible at all, although this form of data reduction techniques technically contradicts our previously stated requirement that data lakes should obtain all information. Thus, it is vitally important to devise domain knowledge-based methods which can first reduce the amount of data that needs to be transferred without loss of information, e.g., if some values can be derived from others. In this context, Lipp et al. [96] propose a process-driven data collection that allows to statically configure which data needs to be collected in which phase of the process at what granularity. This high level of control allows to precisely adjust the amount of generated data to the required signal accuracy as well as the available bandwidths.

Similarly, compute capabilities in the network can also be used to first dynamically detect the current process phase (opposed to the static definition by Lipp et al.) and then scale the generated data volume as needed, e.g., ensuring high data quality in times of interesting process behavior while reducing the load in idle times [87]. Additionally, these in-network processing techniques allow for handling data at line-rate and can thus more easily further reduce the load by removing (presumed) irrelevant information or by performing pre-computation steps [55]. Finding the right trade-off between storing as much raw data as possible while also adhering to infrastructure limitations by reducing the transmitted data is currently a predominant challenge. In the aforementioned use case of Laser Powder Bed Fusion, for example, the compression techniques automatically adapt their configuration to the network bandwidth, processing capabilities, and data structures. Provisioning a suitable infrastructure with sufficient bandwidth and storage capabilities is certainly a long-term goal, enhanced by carefully placed and designed in-network compute functions.

Low-Latency Guarantees. Apart from high data rates, the IoP-enabling infrastructure must also satisfy tight latency bounds, which, for example, are needed for process control [132]. This constraint is particularly relevant if decisions are to be made by a remote system or individual based on broader information from the data lake, rather than by process-near controllers solely based on local knowledge. In this case, physical latencies between the processes and remote systems are often already too high for very time-critical applications (sometimes with requirements in the one-digit millisecond range), rendering pure remote solutions, e.g., over the Internet, infeasible. In-network processing again offers a solution as control programs can be deployed in the network and thus significantly reduce the inherent latencies. These programs currently range from simple LQR controllers [132] to basic line detection mechanisms [56] and can thus cover a variety of simple control tasks. Additionally, complementary safety measures, such as emergency stops, can also be realized in networking hardware, as is demonstrated by Cesen et al. [31]. The accuracy and computing speed of such approaches is generally capable of reaching levels similar to userspace applications, while networking devices are especially capable of processing significantly higher packet rates [86]. Yet, implementing the required functionality on the current generation of programmable networking hardware is still challenging [86].

Control Loops. Thus, in the context of the IoP, we envision that critical decisions for cyber-physical control loops are made quickly within the production cell, e.g., emergency shutoffs for safety reasons. More complex controls are then implemented at the edge, within the network (in-network processing), or in the cloud, where they can source additional data sources. Consequently, the (local) decision quality improves with an IoP. Due to the interconnected nature of manufacturing processes, companies even have the chance to account for issues in subsequent process steps. For example, they can react to slight deviations or inaccuracies with control loop adjustments in the next production cell, effectively implementing a real-time control loop for the shop floor. Here, companies can rely on well-known approaches, or, in line with the IoP, they can utilize control loop adjustments that originate from model-integrated artificial intelligence and the knowledge base in form of the WWL.

Device Heterogeneity. The immense heterogeneity regarding the involved sensors and machinery, which often characterizes industrial production settings, is another challenge that needs to be addressed [1, 35]. For example, depending on the vendor, sensors might differ in their expressiveness or in the protocols that they support. Addressing this challenge on a system user’s level, Bodenbenner et al. [16] propose a domain-specific language that abstracts from the sensor-specific details and allows a unified access to the sensor information. However, their solution does not directly address how such sensors can actually be integrated and interconnected on a technical level. Additionally, devices can also change dynamically, e.g., if a process is reconfigured to allow for the production of different products. High flexibility is thus one of the most important infrastructure requirements. Especially when considering the long lifetime of industrial devices, protocols capable of providing security even in these heterogeneous settings are needed [40]. Paniagua et al. [113] provide a survey of different architectural frameworks for Industry 4.0, such as FIWARE, IDS [112], or BaSys4.0. Although these frameworks also address device heterogeneity, they are not specific on the level of communication protocols. More specific digital industrial platforms are provided by different Industry 4.0 key players such as Siemens (MindSphere) or Bosch (IoT Suite) [115]. However, these platforms are often customer-specific solutions, i.e., they claim to address heterogeneous data and devices, but often require a significant customization effort to fit the needs of specific use cases. Additionally, we note that despite the trends of security-by-design and privacy-by-design, devices and protocols must be configured correctly to benefit from these, which is, however, a frequently neglected aspect [38, 39].

Example. Overall, an interconnected and industry-capable infrastructure will enable the e-vehicle manufacturer to move away from a traditional assembly line towards a more dynamically reconfigurable sequence of production steps [25]. In this context, the flexible data collection and integration centered around the data lake is key for ensuring a sufficient richness of data even in such dynamic settings; thus, allowing access to detailed process information at all times, e.g., for quality assurance purposes. While technical advancements, e.g., regarding storage and data rates, are important to steadily increase the amount of processible data, layered control loops will enable a fine-grained control of all running processes to improve the overall efficiency of the system and allow for quick responses to system changes.

Takeaway. The specific characteristics of industrial environments pose challenges that cannot be addressed by traditional data management and networking solutions, as some aspects require the inclusion of remote services or edge computing while other aspects simply do not support such approaches. Instead, a concept is needed, which carefully includes mechanisms to satisfy all of the mentioned needs: high-data rates, support for heterogeneous data structures, low-latency data processing, and flexibility.