Keywords

1 Introduction

Today’s customer demands, e.g. high product quality or sustainable production at low cost, impose high requirements on manufacturing companies. Digitalization of processes enables industries to increase process stability, product quality, reduce downtime, and increasing throughput. This leads to highly complex production systems with interconnection of hard- and software. Manufacturing companies must balance supervision and orchestration of resources, executing processes for high quality products.

Big Data is leveraged by Data Analytics, allowing for real-time insight, process optimization, and the prediction of failures before they impact corresponding operations [1]. Data Analytics can be divided into descriptive, diagnostic, predictive and prescriptive approaches. The most advanced form of Data Analytics is prescriptive analytics, which offers the highest degree of decision automation by providing the best possible reaction strategies. Prescriptive Analytics can develop its full potential when applied to the entire production management system [1]. While the combination of production system with data-driven analytics towards smart manufacturing seems very promising, propagation is very low. We do spectate that Data Analytics and AI is almost not used in small and medium-sized enterprises (SME) across the EU [2]. The successful Data Analytics application to solve manufacturing challenges usually gets as far as the predictive level; prescriptive approaches still face manifold challenges.

Companies usually lack a mature data infrastructure to apply AI in production, even though this is an essential part of deployment [3]. Use cases are implemented as lighthouse projects, but a lack of methodology and widespread use is still prevalent [4]. This is due to a lack of transparency on return on investment when implementing such complex solutions [2]. Manufacturing companies unique value proposition revolves around their domain knowledge. Extensive knowledge and capacity for the implementation of complex AI models is often lacking [5].

Prescriptive Analytics needs a full understanding of all elements in the feedback loop of the System of Interest (SoI) [6]. Possible SoIs are resources, processes and products with respective use cases like Prescriptive Maintenance, Autonomous Processes, Prescriptive Quality. Influences on the SoI are derived from inputs and environmental influences, with environmental production resources and processes leading to a complex System of Systems. This creates a highly complex solution space with a lot of factors of uncertainty [6]. Thus, we identify a need to structure use cases to improve a further understanding of the process of developing successful Prescriptive Analytics solutions (in all technical readiness levels). A structuring approach enables data scientist to further standardize and reuse developed artifacts, thus decreasing development time of analytics and increasing acceptance and propagation of use cases.

While approaches like reinforcement learning allow completely data-driven decision making and prescription, it involves knowledge about a SoI in context of its environment and a system-of-systems-understanding. However, current development of Data Analytics solutions is mainly focused on a singular system of interest. A simplified approach is the formalizing of expert knowledge into standard reaction strategies.

Based on the state of the art (Chap. 2), a research gap regarding the characterization of prescriptive analytics use cases in smart manufacturing is identified through literature research. Chapter 3 addresses these issues by presenting an approach to combine both analytics and smart manufacturing driven perspectives on the use case domain. The approach was derived through the extension and refitting of existing approaches. The proposed extended P2PR-model for prescriptive analytics in combination with the provided performance levels enable a new standard in smart factories for reaction strategies and decision making. The summary, future work and further research gaps based on the results are provided in Chap. 4 (conclusion and outlook).

The research goal of this paper is to present approaches to enable SMEs to structure prescriptive analytics use cases in smart factory environments. Our contributions compared with existing work are:

  • We extend the existing PPR-models towards prescriptive analytics and provide a systematic approach to classify use cases and analytics levels.

  • We interlink the degree of decision support and expert knowledge to create performance levels for prescriptive analytics in smart manufacturing systems.

  • We propose a 3-dimensional approach to structure future use cases and their algorithms to find similarities and extend the reusability and applicability of future prescriptive analytics case studies.

2 State of the Art

Manufacturing analytics describes the use of data-based algorithms to ensure product quality, reduce maintenance costs, and optimize production processes. Manufacturing analytics methods are divided into four levels, with each level further reducing the need for human interaction [7]. The levels descriptive, diagnostic, predictive, and prescriptive Analytics will be explained in the following:

The most basic form is descriptive analytics, which provides basic insight and visualizations of what has happened in previous production cycles [8, 9]. Diagnostic analytics methods close this gap by mapping data with possible causes to identify correlations between variables, sources of trends, and errors. When it is known what has happened in the past and why it happened, it is possible to predict what is likely to occur in the future. Predictive analytics refers to the use of statistical modeling, data mining techniques, and machine learning to make predictions about future outcomes based on historical and current data, such as predicting the next maintenance date [10, 11]. Prescriptive analytics builds upon the three other types of Data Analytics which describe the present and make predictions about the future [4, 12]. It processes and evaluates predictions as well as detected errors and anomalies and derives reaction strategies or recommendations for actions. In addition, expert knowledge can be introduced to further enhance this process.

Prescriptive Analytics is usually structured into a fully automated support and a human in the loop approach according to Gartner [7]. Vater et al. propose to structure existing approaches by the level of human interaction and the chosen IT infrastructure [4].

2.1 Formalization of Data Analytics Use Cases in Smart Factories

In the domain of Data Analytics, a smart factory specific systematization of use cases for prescriptive analytics has not been discovered via systematic literature research. Authors mainly focus on solving defined problems with narrow boundaries [13,14,15,16,17]. Use cases are either described in a process-related, result-based or artifact-based way. Workflows to recreate analytics solutions (crisp-dm) are process-related examples. Result-based examples are typical of experiment based and centered machine learning papers. Artifact-based papers develop a use case specific solution that evolved around a reusable part like a UI or algorithm. Kühn et al. propose to structure use cases regarding their interaction with algorithms, it-infrastructure, and data sources [18]. This agnostic approach does not consider which the solution has. A canvas to identify analytics use cases regarding smart services was developed by Panzner et al. [19]. The approach focuses on generating use cases.

2.2 Product, Process and Resource in Smart Factories

A “product” is defined as the final or intermediate result of production processes. The flow descriptions necessary for the creation of a product or its individual parts and assemblies are described by the process. The process describes all interacting operations that are necessary for the completion of a product. Resources are the objects necessary for the execution of a process or generally to fulfill a task (e.g., infrastructure and employees), but not the work object or the product to be produced itself. For a more detailed description of PPR, we refer to [20, 21]. The PPR model defines the links between the three entities and is deeply rooted in quality management. In [22], a skill definition of PPR is proposed. A skill is an abstract form of a process that defines the ability of a resource to perform a process. A task is the application of a skill to a defined product with a desired outcome. A similar approach for Machines as a Service is found in [23] where the product and resource are decoupled from each other. A resource has skills that must match the requirements of the product. Only on a match, a process task is executed.

3 Structuring Prescriptive Analytics in a Smart Factory Environment

All analyzed approaches regarding smart factory use cases deal with specific problem-solution-combinations and do not generalize their approach sufficiently. Solutions regarding data governance and integration are missing. Hence, to our knowledge, no method with a holistic approach towards further specifying and categorizing use cases in a smart factory environment could be identified. The general approach for dealing with uncertainty in use case specifics was identified as a research gap.

3.1 Data Analytics View on Use Cases

One possible way to structure analytics approaches is given by the analytics levels according to Gartner (see Chap. 2). The one-dimensional and algorithmic view does not deal well with the complexity and uncertainty of possible use cases in manufacturing. Furthermore, the different approaches and levels of possible prescriptive approaches are not well represented regarding different kinds of prescriptive analytics. Dimensions like decision impact, area of validity and the interconnectivity of different decision-making systems are not represented. We propose a further differentiation between the different possible kinds of prescriptive analytics.

This is motivated from a technical point of view. Autonomy levels and other Industry 4.0 concepts work in parallel to the proposed approach. The different levels are visualized in Fig. 1. They are needed to make prescriptive analytics more applicable regarding their impact on smart factory production systems. A structuring approach assists by further specifying the solution space for prescriptive analytics and the validity of the decision under observance. The degree of decision support and the degree of integrated expert knowledge in the system create the X and Y axis. The subsystem-specific strategy selection is already complex, but the first level of prescriptive analytics. Based on that, system specific reaction strategy selection exists. The highest form of prescriptive analytics makes decisions that are taken with regard to other systems’ decisions as well (outside the SoI). The model enables a more differentiated discussion about different solutions for prescriptive analytics use cases. The goal is to later enable a mapping between the technical solution (algorithm), the prescriptive analytics level (Fig. 1) and use case specifics (e.g., the PPR model). It mainly focuses on the value proposition of the analytics level “prescriptive”. It is contributing towards the technical feasibility of the automation of actionable decisions, by making them compatible with other concepts such as autonomy, reconfiguration, online learning (update mechanisms), self-healing mechanisms, human in the loop, and other approaches.

Fig. 1
A graphical representation of the degree of interconnectivity for decision-making versus the size of the area of validity. Three regions of use cases 1, 2, and n are presented.

Vision of smart factory use cases in prescriptive analytics (based on [24])

Full size image

3.2 Smart Manufacturing View on Use Cases

A second dimension is needed to connect smart factory and Data Analytics views on use cases. The second dimension can be built based on the existing PPR model (product, process and resource) as presented in Chap. 2. The model was initially developed to structure workflows in car manufacturing companies. The existing model does not focus well on the differentiation between different kinds of processes. Manufacturing processes are processes like stamping or welding. Other processes like production planning, scheduling, logistical orchestration or the allocation of resources in production have completely different characteristics. Organizational processes and manufacturing processes share the same name but demand different analytical tools and algorithms to handle them (compare [12]). The goal is to ensure a maximum difference between different classes for analytics use cases (and used algorithms). Thus, we propose to differentiate the “process” into a supervisory and a manufacturing related one. This way, business and organizational processes and their use cases are strictly divided and differentiated from production processes like welding and stamping. Based on Stanev [21] and Haasis et al. [20], we propose the following definitions for product, process, and resource: The definition is given to further standardize the usage of all terms and lays a foundation for the extended PPR (P2PR) model:

  • Process (supervisory): The process is no physical entity. It maps purpose (task) and realizer (resource) to create products. Thus, it defines the order in which resources are invested to create a product and which state of the product is regarded as a finished product. The realizer and the purpose do not have to be tied directly to a product. A Data Analytics process use case optimizes the orchestration and behavior behind the allocation of resources which are embedded in the process.

  • Process (Manufacturing): Manufacturing Processes describe operations like welding, stamping or glueing. A data analytics manufacturing process use case optimizes their parameters, regulation or reconfiguration.

  • Resource: Resources represent the entities, that are used to generate products with defined characteristics. Resources are orchestrated by a process and always exist in a logical relation to other resources and steps in the production value chain. allocated to a task by a process. A Data Analytics resource use case deals with maintenance, health status and behavior of the resource regarding its ability to do its intended job (ensure ability to produce under given KPIs).

  • Product: Products in a smart factory environment are physical entities with characteristics (planed and realized). They are instantiated because a process demands it, and a resource fulfills the demand. A Data Analytics product use case analyzes and optimizes the quality, integrity and characteristics of a product instance or class.

Additionally, an environment can be added to integrate non-manufacturing, but use case related objects into the model. E.g., the delivery of parts from a partner can take longer, thus having an influence on the existing product, process, and resource. The new proposed model is summarized in Fig. 2. The surrounding text boxes are there to further enhance the model understanding and represent examples. Different authors [22, 23] have altered the PPR model as well but have not focused on its application in the smart factory domain (see Chap. 2). The definition is meant as a reference for the design and assignment of future use cases into a category. The model needs to be expanded in the future for non-smart factory (but enterprise) related advanced analytics use cases. The model explicitly does not include the manufacturing environment (e.g. order and purchase processes).

Fig. 2
A circular P P R model of a smart factory environment and its components. 3 parts of the model are process, resource, and product. The environment consists of service provider, partner, ecosystem, and user.

PPR-Model and its elements in a smart factory environment (extended and based on [20])

Full size image

Our solution to combine the Data analytics domain and the smart factory environment is to merge both structuring views to generate a general use case structuring approach for smart manufacturing Data analytics related use cases. The goal is to group use cases according to their key characteristics and reference objects (product, process, resource). A taxonomy can be built based on this approach. The main goal of combining both approaches is to bridge the gap between the domains of manufacturing and Data Analytics. The main effect is a common baseline among stakeholders, including both data scientists and those responsible for production (see Fig. 3).

Fig. 3
A framework diagram of prescriptive analytics and updated P P R model for a smart factory. The P P R model consists of product, process, and resources.

Combination of Data Analytics and manufacturing domain to find a common approach for smart factory use cases

Full size image

A combination of both models results in a two-view-based classification of smart factory use cases. The generic description of the use case categories is a first step to help categorize use cases into a use case matrix. The categorization can be used to group algorithm classes in use cases according to their potential. To further explain the given characterization of use cases, the following real-world examples are given:

  • Resource, cross-system knowledge-based reactions strategy selection: The maintenance algorithm of the scrap belt conveyor predicts a system failure in two weeks. It automatically schedules a maintenance shift and gives a recommendation how the problem can be avoided or fixed (e.g. based on a reaction strategy selection from a ontology like database).

  • Process Supervising, System specific knowledge-based reaction strategy: selection: Based on changes in product demands, the production schedule is changed accordingly based on different input target values. The schedule is automatically changed.

The standardized approach to defining use cases builds the foundation for further specifying a taxonomy for smart factory use cases. They provide the basis for finding common building blocks in each use case. Efforts can be reduced heavily if the degree of reusability of the applied solution is high. The key to its success is the integration of expert knowledge and different model types. The models can be structured according to the proposed categories in the matrix.

Future work needs to focus on how well both the descriptive and prescriptive approaches to a smart factory use case work together. Based on Lepenioti et al. [12] (SLR to prescriptive Analytics Approaches) and the further implementation of use cases, one could add the dimension “algorithm” to further map which technical solution works best for which customer need (prescriptive level) based on which subdomain of the smart factory (ppr-model). A multi-dimensional space to describe use cases in (prescriptive) analytics for smart manufacturing results (see Fig. 4). The goal is to use the structuring approach to find common solution elements that can be reused or resampled. The approach generates a baseline to further reduce the need for expert knowledge during production. It frontloads the effort by implementing expert knowledge into the system. The overall goal of the approach is to reduce the investment of expert knowledge and time into the overall lifecycle of production system. The proposed approach is complementary to the approaches of Kühn et al. and Panzner et al. and can be used in a later stage of the use case development and specification.

Fig. 4
A graphical representation of the degree of interconnectivity for decision-making versus the size of the area of validity. The use cases from 1 to n of the domains are marked.

Vision of use cases in all 4 domains, based on the two differentiation schemes

Full size image

4 Conclusion

The presented systematization scheme for (prescriptive) analytics use cases in a smart factory environment enables companies and researchers to find common attributes between different prescriptive smart factory use cases within one production ecosystem. A data-centric and a manufacturing-based view (PPR) were presented, reconfigured, and combined. The combination is derived from the Gartner analytics levels and the PPR model according to Haasis et al. [20]. The combination (P2PR and 2-dimensional characterization of prescriptive analytics) ensures a holistic view of prescriptive analytics use cases in manufacturing. Their combination creates a complete view from both analytics need and manufacturing prerequisites site. The resulting schema creates a first step towards a standardized taxonomy for prescriptive analytics use cases.

Further research questions arise in the context of how to combine the different prescriptive analytics approaches. Effective prescriptive analytics demands an efficient integration of expert knowledge and a high degree of reusability between use cases. The implementation of future prescriptive use cases will create a baseline to contribute to finding links between the P2PR characterization and different ML models or ML classifications of models. This approach can be embedded in the presented characterization scheme. Additionally, one needs to create a framework (technical and logical) to integrate the IT infrastructure int a shared approach to create synergies in smart manufacturing companies.