1 Introduction

Industry 4.0 (I4.0) [13] denotes the technological change towards intelligent production in which Artificial Intelligence (AI) methods, data analysis techniques, the Internet of Things (IoT) and distributed systems are integrated into industrial processes. In this context, optimization and efficiency-increase of production processes is investigated [20, 22, 38]. The growing automation and interconnection of factories offers new opportunities, such as the promising use of edge nodes [46]. These are miniaturized computing devices located directly in the production environment. The exploration of the industrial potential of these edge nodes is addressed by the EASY projectFootnote 1 running from 2022 until the end of 2025. The acronym stands for “Energy-Efficient Analysis and Control Processes in the Dynamic Edge-Cloud Continuum for Industrial Manufacturing” and is the name of a German consortium funded by the Federal Ministry for Economic Affairs and Climate Action. In addition, an affiliated project financed by the Austrian Research Promotion Agency contributes toward this research. Namely, the consortium consists of the following partners: Empolis Information Management GmbH (consortia leader), German Research Center for Artificial Intelligence (DFKI), Robert Bosch GmbH, Fraunhofer IOSB-INA, Trier University of Applied Sciences—Environmental Campus Birkenfeld, ArtiMinds Robotics GmbH, coboworx GmbH, and Salzburg Research.

Within EASY, this consortium aims to overcome technical barriers in industrial manufacturing by creating an open and standardized Edge-Cloud Continuum (ECC) that should enable the optimized execution of Analysis and Control Processes. This continuum is a distributed infrastructure that includes the described edge devices as well as a central cloud, and enables dynamic distribution of computations across all contained devices [35, 39]. AI methods are used to plan and execute both types of processes in the ECC. These methods can run both, centralized on the cloud and decentralized on the edge. The inclusion of the edge nodes allows analyzing the high-frequency data in near real-time, and also enables data protection for companies through non-sharing. This standardized, freely available ECC should lower the entry barriers for small- and medium-sized enterprises to these technologies. To demonstrate and evaluate the described aspects, various prototypes are developed.

In the remainder of this paper, an overview of the architecture of the EASY project and the considered processes is given in Sect. 2. In Sect. 3, the AI methods applied to these processes are presented. The use cases to demonstrate the project’s results are described in Sect. 4. Afterward, Sect. 5 provides a summary and an outlook for future research in the context of our work.

2 EASY Architecture and Processes

In this part, the overall EASY framework targeting the project’s goals is presented. First, the basic architecture is introduced in Sect. 2.1. Then, the processes within this are described in Sect. 2.2 and classified in the Business Process Management (BPM) [12] research state.

2.1 Architectural Overview

The ECC is a distributed environment composed of computing and network infrastructure [24, 35, 39]. Here, edge devices offer local data processing and storage as well as service execution, whereas the cloud provides the same functions on a server network with higher capabilities. These edge nodes can be run by individual members (i.e., industrial companies) while a separate entity provides a central cloud platform. In the ECC, data and computations can be flexibly moved between such edge devices and the cloud. This allows the transfer of data to the cloud to be minimized, as the compute resources available on the edge can be used to at least partially process the data. Therefore, the significant cost involved with large data transfers can be reduced. However, the cloud orchestrates the ECC and determines where analysis or control processes should be computed within the continuum. In addition, this architecture can address the issue of data privacy by allowing companies to avoid moving specific data to a cross-enterprise cloud solution. Therefore, certain assumptions such as the integrity of the cloud platform provider have to be made.

Fig. 1
figure 1

An overview of the EASY architecture

Full size image

We will develop such a continuum for the EASY project in the form shown in Fig. 1. The lowest level is formed by individual assets on the production floor, such as manufacturing robots or IoT control components [21, 45]. These assets are usually constrained in terms of computing power. However, they can still be used to perform small computational tasks. To enhance the compute power at these positions, the assets are connected to edge nodes, which form the second-lowest layer. At these nodes and along the continuum, services can be run to process, aggregate, or simply pass data to other entities. Specialized computing infrastructure at the edge, such as compute clusters or big data storage, could also be employed before moving to the cloud. In this architecture, the path between the edge and the cloud is paved with different layers of nodes providing increasing computational capabilities. This enables a dynamic and optimized execution of AI services, in terms of metrics, e.g., the project’s namesake energy efficiency, but also other Green AI and Sustainable Software Engineering criteria [49], such as resource and data efficiency [23]. To identify and optimize such metrics required for measuring these values in the ECC, a reference model and an exemplary measurement method [17] are used as basis. Within the ECC, the AI services are stored in a central repository based on the Gaia-X architecture [47]. By using these services on the edge nodes, the data only requires transmission to the nearest one with sufficient resources and, in most cases, not all the way to the cloud.

2.2 Processes in EASY

In EASY, we focus both on Analysis and Control Processes and address these dynamically in the ECC using AI techniques:

  1. 1.

    Analysis Processes: These processes are geared towards examination and analysis of data, e.g., based on IoT sensor streams [45]. Considering this, local analyses such as visual quality control [32] or error detection [15] are to be carried out. Regarding the ECC, the analysis will identify the use of resources such as computing capacity and energy consumption. In addition, aspects of the decentralized architecture that can increase data security are explored [10, 11].

  2. 2.

    Control Processes: These processes concern the management of the value creation processes, meaning the manufacturing facilities and their resources. This includes the dynamic allocation of these resources throughout the ECC, as well as the automated planning of production processes and their flexible and correct execution [6, 27]. In addition to the sustainability criteria, the processes should be robust and flexible so that in cases of deviations they can still be executed or adapted accordingly, even in cases of deviations [16] or changed metrics [6].

In the BPM research field, flexible analysis and production processes have already been addressed [29]. Both processes should focus on their resource efficiency, as criterion [9] already specified above for each type. To consider this and other sustainability metrics accordingly, semantic information about the processes must be available [25] and provided in a suitable semantic structure [30]. This must be an appropriate digital representation of the devices in the ECC or the production resources, e.g., like a semantic Asset Administration Shell [4, 41]. The approaches mentioned so far rarely use extensive AI methods. However, these offer great potential for optimizing adaptive process management and go beyond manually performed adaptations [27, 33]. Therefore, the EASY project targets this issue in the ECC and in physical smart factories. The AI methods used to manage these processes are presented next.

3 AI Methods in EASY

In the EASY project, various established AI methods address processes in the ECC and are investigated for optimizing resource usage. In the following, we present the techniques of AI Planning (Sect. 3.1), Case-Based Reasoning (Sect. 3.2) and Distributed Learning (Sect. 3.3) in this context.

3.1 AI Planning

AI Planning aims at solving a state transformation problem, where the goal consists of finding a sequence of steps to transform a discrete world model from an initial state to a desired goal state [14, 19]. This technique is already applied in the BPM area to increase automation and support [33]. In EASY, we investigate this technique to facilitate flexible Analysis and Control Processes dynamically in the ECC. The relevant analysis process mainly consists of service orchestrations regarding computation resources in the ECC. Here, the goal state is achieving an optimal distribution of computation and data across the individual instances, e.g., for data aggregation. As the result, this is federated among the participated edge nodes and the cloud. For the control processes, the manufacturing environment is considered, where the goal state is a desired product with specific characteristics. The planning problem thereby consists of finding a sequence of executable manufacturing actions which lead to the desired products. Both, the generated analysis and the control processes, can be executed automatically. A drawback of solving complex planning problems is the high computational complexity needed, especially when used on edge devices. To mitigate this, we will combine AI Planning with other AI techniques [18].

3.2 Case-Based Reasoning

Case-Based Reasoning (CBR) is an AI method for experience-based problem-solving [1, 7]. Problems and their corresponding solutions are stored as cases that form the basis of addressing new problems. Using similarity as a criterion, suitable cases are identified, and their solution is adapted. In the context of the EASY project, both analysis and production processes are created using CBR. In the manufacturing domain, CBR can be used to perform analysis processes such as predictive maintenance or identifying data quality issues based on IoT time series data [31, 42, 44]. The advantage of using CBR in this context is that in comparison to other AI methods only a few error cases are required. To optimize production processes, a case is represented as a workflow in the sub-field of Process-Oriented Case-Based Reasoning (POCBR) [36]. In the context of planning of processes, POCBR is used to reduce computational complexity and increase flexibility by reusing already solved problems [28]. Thus, existing plans are retrieved and, if necessitated, adapted for the new requirements by AI Planning. In EASY, we use this for flexible planning and execution of the processes. These CBR applications for analysis and control are to take place on the edge as well as in the cloud dynamically.

3.3 Distributed Learning

Distributed Learning is a Machine Learning (ML) [37] approach performed across multiple computing resources [40]. Traditional ML and Deep Learning [8] systems rely on large amounts of centrally stored data, so that most (locally stored) data cannot be used due to computational complexity. The classical approach in ML or DL would be to send the (small) data sets of the single machines to a central server. There, a model is trained based on the data. In practice though, data often cannot be transported over a network due to privacy or bandwidth issues [26, 48]. To address this in EASY, we use Federated Learning (FL) [34] as a distributed ML approach to keep the data local at each machine as an edge-node and learn for local analysis processes there. Instead of explicit data, the local model weights are shared in the ECC. Besides the advantage that no real data is sent, all nodes share their knowledge within the continuum. The respective edge devices can adapt the globally merged analysis model for their data without directly influencing the performance of the other local models. In addition to traditional manufacturing data analysis using FL, robotic learning and optimization is also contained in EASY based on this technique. For this purpose, existing previous work [3] is built upon to first pre-train predictive robot foundation models on large data sets in the cloud and then achieve model-based parameter optimization on the edge devices [2]. Additional to the described communication reduction in these applications, energy consumption should be reduced by the distributed FL methodology.

4 Use Cases and Demonstration

To showcase the approaches developed in the EASY project, six demonstrators are built based on different use cases, briefly described in the following. The participating and associated industrial partners aim to ensure that the demonstrators represent real-world use cases. In addition to qualitative aspects, quantitative parameters based on the metrics to be developed, such as energy or resource efficiency, are examined for each demonstrator.

Together with SmartFactoryKL and SmartFactoryOWL as associated partners, DFKI and Fraunhofer are creating two similar setups at the respective, cooperating locations. These are used to demonstrate the concept of shared production across multiple sites. This use case demonstrates the flexible and resilient control processes as well as analysis processes for quality control. To apply the POCBR approach, the CBR framework ProCAKE [5, 43] is used and extended for the application in the ECC.

Bosch realizes a demonstrator which is especially suited to invest analysis processes using FL. The demonstrator consists of several standardized, interconnected edge devices whose behavior is monitored and controlled using a graphical interface. In this setup, the data of several milling machines connected to edge devices are processed and models are learned federated on this basis.

FL approaches in the ECC are also being explored by ArtiMinds in a demonstrator that focuses on robot learning and optimization. Among others, force-control and vision-based handling are investigated. The implementation of the FL methods will also be integrated into the industrial robot data platform ArtiMinds LAR.Footnote 2

In another demonstrator, Coboworx and Salzburg Research, present analysis processes by monitoring the condition of an industrial robot in a distributed palletizing application. The reliable communication within the ECC is visualized and measured regarding anomaly detection, to prevent possible economic downtimes.

Furthermore, the Environmental Campus explores analysis processes with focus on the resource, data, and energy efficiency of distributed learning and applies these in demonstrators for analyzing multi-modal sensor data.

5 Summary and Outlook

In this paper, we present the idea of the EASY project, which aims at realizing a dynamic ECC for industrial manufacturing. Our focus is on the application of AI methods to BPM processes, namely to the presented Analysis and Control Processes. In this context, we describe various research areas for future work, addressed within the EASY project. We will demonstrate the project results in an industrial context and evaluate them regarding their energy and resource efficiency.