Knowledge Graphs (KGs) have a large potential to support analytics and AI applications in the manufacturing domain, which in the age of Industry 4.0 is increasingly driven by data. Compared to other domains, however, KG techniques have seen limited adoption in this eld so far. We argue that the construction of KGs in the context of Cyber-physical Production Systems (CPPSs) requires a systematic methodology grounded in domain-speci c abstractions. Consequently, we introduce a KG modularization framework based on the well-established RAMI 4.0 architecture model. A key bene t of the proposed approach is that the resulting KGs support navigation across abstraction hierarchies, enabling bottom-up contextualization of raw data on the one hand, and top-down explanations by linking to lower levels of granularity on the other hand. We motivate the proposed approach and illustrate its application with a real-world use case from the automotive sector.
Industrial production is currently undergoing a major paradigm shift that is
often described as Industry 4.0 (I4.0). This development at the con uence of
advances in digital and manufacturing technologies [
Motivation and challenges. Our motivation stems from industrial applications such as the ones within the H2020 Teaming.ai project5. In particular, our real-world motivating scenario focuses on human-AI cooperation in quality management and optimization of injection molding processes in the automotive industry, with a KG at its core. This production process is composed of the following steps: (i) initial parameter setup: setup of the injection molding machine and production of trial parts; iterative parameter adjustment until quality issues are resolved; (ii) production start; (iii) automated quality inspections during production, i.e., a computer vision system inspects each produced part; (iv) parameter optimization is triggered when faults are detected. This scenario raises a number of more general archetypal challenges for CPPSs: { (C1) It requires collaboration of AI agents and human stakeholders, such as quality controllers and process engineers, across multiple organizational, functional, and temporal levels of granularity. { (C2) Reliable quality control and parameter optimization requires contextualization, e.g. of sensor data with higher-level operational information and quality requirements (\navigating upwards"). { (C3) Diagnosis (e.g., of increased NOK rates for a particular part), by contrast, requires explanations, which necessitates the ability to navigate to lower levels of granularity (\navigating downwards"). { (C4) Finally, a key challenge is to link decisions and goals on the production (e.g., line control) and operational levels (e.g., classi cation of defects) to higher-level business goals (e.g., increase Overall Labor E ectiveness (OLE) and guarantee the ful llment of quality requirements).
To address these challenges, we propose a KG layering approach that facilitates purpose-driven, agile construction of reusable KGs across multiple layers of abstraction and perspectives.
Opportunities and Impact. KGs have countless demonstrated applications
across industries including nance, technology, and many other sectors. However,
this potential of KG-based systems is still largely unrealized in manufacturing,
where the concept has seen limited adoption and integration into real-world
manufacturing systems so far [
With the approach demonstrated in this paper, we are building KGs for manufacturing companies in the injection molding and milling sectors, but also for higher-level applications such as manufacturing and use of replacement parts. The most important concrete impact of KGs in these use cases are (i) delivering a trustworthy, auditable, integrated datastore that preserves the inherent heterogeneity of manufacturing data, (ii) the combination of di erent types of AI and 5 http://teamingai-project.eu human reasoning in an integrated platform, and (iii) improved decision-quality in the manufacturing and selection of parts by taking all available information into account. 2
Related Work
Domain-speci c layering models are prevalent in the manufacturing
domain. The ANSI-ISA-95 standard [
E orts to formalize some aspects of these models have focused either on
speci c subsets, such as the semantic description of CPPS components' asset
administration shell [
Apart from using KGs to relate di erent layering models to each other, the
semantic web community also put e ort into making KGs more modular. This
can, e.g., be accomplished by organizing ontologies based on their abstraction
or by proposing proven design principles. For the former, Guarino [
Layered CPPS Knowledge Graph Construction
As a foundation to organize industrial KG conceptualization and
implementation, addressing challenge (C1) raised in the introduction, we focus on the
vertical layer axis in RAMI 4.0 [
Process Engineering Business Functional Information Communication Integration Asset …
C ilitttzxeaauno liitxaannopE o n
KG views { In the conceptualization phase, they support the scoping of vocabularies and ontologies used to represent various aspects in the KG; Mapping concepts to layers facilitates a modularization of the KG that is consistent with the layering of the manufacturing system.
Example. As discussed before, the injection molding use case is composed of the steps: (i) setup of the injection process, (ii) production, and (iii) quality inspection. The layering allows to interface with the domain experts appropriately: the process engineer, e.g., acts at the communication and integration layers (typically represented by Supervisory Control and Data Acquisition (SCADA) and/or Programmable Logic Controller (PLC) systems).
{ At production run time, the layers serve as a guideline to structure the interactions between components and stakeholders with the KG, based on the modularization principles of the framework.
Example. For instance, classi cations made by automated visual quality inspection components are assigned to the integration layer and may be propagated from there through higher-level events to the upper layers. In this context, the KG is a digital twin of the real-world system that links relevant aspects for a given perspective. This allows to overcome challenge (C4) from the introduction. { Methodologically, we propose a modular and incremental approach to develop KG-based applications in I4.0; rather than developing a single, monolithic KG, which requires a large up-front investment and bears considerable risk, our approach aims to incrementally build a CPPS-wide KG as an aggregation of more granular per-use case views, while ensuring their conceptual alignment through a set of shared core concepts. This also enables cross-view linking and navigation across multiple perspectives (dotted lines in Figure 1). Example. In our use case we have discussed (i) setup of the injection process, as well as (iii) quality inspection. In Figure 1, we see the foreground view \Process Engineering" corresponding to (i) and \Quality Control" corresponding to (iii). This however, can be the results of an incremental construction. { Critical to our approach are contextualization (C2) and explanation (C3), illustrated by the \upwards" and \downwards" arrows illustrated on the right side of Figure 1. The proposed modularized KG approach allows one to naturally addresses these challenges.
Example. A process engineer sets a number of initial parameters for the injection molding machine, e.g., injection speed. The number of parameters vary based on machine type (20+); choosing appropriate values requires substantial expert knowledge From the view of the process engineer, contextualization is of utmost importance - what e ect these adjustments have on the overall outcome of production (presented at higher RAMI 4.0 levels). From the view of the factory manager (typically at the three top-most layers), explaining why the production outcome is how it is, e.g., based on parameter settings on the injection molding machines (presented at lower RAMI 4.0 levels) is key. A modularized KG approach is a natural t for this.
Conclusions. In this paper, we introduce a KG modularization framework in the context of CPPSs and illustrate it in a real-world industrial use case. Our layered CPPS KG provides navigation across di erent abstraction hierarchies (i.e., bottom-up contextualization and top-down explanation) but also linking across multiple views. For future work, we plan to demonstrate the application of the framework to a wide set of use cases. Our goal is both to demonstrate the concrete impact of the KG platform across our manufacturing use cases, including our industrial partners in automotive injection molding, milling, aircraft manufacturing, and consumer appliance manufacturing.