Elsevier

Computers in Industry

Volume 125, February 2021, 103389
Computers in Industry

Vertical data continuity with lean edge analytics for industry 4.0 production

https://doi.org/10.1016/j.compind.2020.103389Get rights and content

Highlights

  • Necessity of high-performing analysis to deal with the growing volumes of data in production

  • Combination of signal acquisition and simultaneous data evaluation necessary

  • Approach implements vertical data continuity by combining signal acquisition and simultaneous data evaluation

  • Enables the use of data for near real-time modelling of the production system

Abstract

Industry 4.0 is characterized by the digitization and networking of machines and systems in production. The amount of data in production is increasing, providing information about processes and thus enables the autonomous monitoring, control and optimization of value creation processes. However, there have been several open challenges and current research questions identified. In particular, new solutions need to be scalable and high-performing to deal with the growing volumes of data close to real-time. The work at hand tackles these research gaps by presenting an approach to realize vertical data continuity by combining signal acquisition and simultaneous data evaluation in a decentralized system without the use of time-consuming external cloud solutions. The approach has been evaluated in laboratory as well as in industrial settings.

Introduction

Industry 4.0 is characterized by the digitization and networking of machines and systems in production (Heinze, 2017). The amount of data in production is increasing, providing information about processes and thus enables the autonomous monitoring, control and optimization of value creation processes ([Hecker et al, 2017], Steven, 2018). Condition monitoring uses this data in production plants to obtain relevant information about the condition of plant components in almost real-time (Abele et al., 2015). Consequently, production plants can be monitored autonomously, plant faults can be detected early, maintenance requirements can be recognized in advance and measures can be planned according to individual needs of plant administration (Reinhart, 2017).

However, there have been several open challenges, current research questions and problems identified in research (Peres et al., 2018). There is still a need to further combine real-time streams of data from the shop-floor with historical data at both the resource and system levels, as well as closing the loop to autonomously act on the results of the predictive analytics. In the context of Industry 4.0 systems, new solutions should be flexible as well as scalable and high-performing in order to deal with growing volumes of data. Solutions should also be highly adaptable, being capable of changing even after deployment by learning from newly generated knowledge, adapting at both the analysis and action fronts. In detail, the identified challenges are: (i) Increase of unanalyzed data, (ii) Heterogeneous and partly manual data acquisition, partly erroneous data of different granularity, (iii) Insufficient use of collected data for the optimization of the own production system, (iv) Lack of real-time analysis and operational utilization of data, (v) Detection of anomalies afterwards by data mining in the cloud; and (vi) Hardly central data analysis due to the high amount of data at sampling rates ≥ 10 khz (Peres et al., 2018, O’Donovan et al., 2019).

The work at hand tackles these research gaps by presenting an approach to realize vertical data continuity from the measuring point to the utilization of production-related data in a modeling of the production system, e.g., in a Digital Twin of the production system. The paper especially addresses the possibilities of implementing decentralized, near-realtime signal preprocessing algorithms and machine learning methods on resource constrained computing systems at the machines edge, that deal with high sampled current signals, perform information aggregation and derive machine-based knowledge. This Lean Data Analytics approach avoids subsequent data mining, the unnecessary storage of large amounts of data and thus eliminates waste in terms of lean production (Küfner et al., 2018). For this purpose, an intelligent sensor called CogniSense is developed which records signals in high resolution, analyses them decentrally, thus reducing the existing data volume without relevant loss of information and exclusively passes on findings to a superordinate optimisation instance. Machine learning methods (ML) are used to recognize patterns in current profiles, learn normal conditions and report deviations - ideal prerequisites for the implementation of predictive maintenance (PM). The analysis is performed decentrally and in near real-time on the hardware, without the use of a superordinate cloud and the associated latency and bandwidth restrictions. According to the ISO/IEC 2382, the term real-time describes the “pertaining to the processing of data by a computer in connection with another process outside the computer according to time requirements imposed by the outside process” (ISO, 2015). This does not result in a clear value for the time interval of the term real-time and it has to be defined specifically for the application. For the presented approach, no immediate data transfer is necessary, so that the term near real-time is defined as a time interval of less than one second. The sensor is also able to safely monitor various electrical loads and can therefore be used universally. Furthermore, it is possible to detect operating states and to record them over time. This data is provided for real-time optimization of production planning and control (PPC). The Digital Twin, a virtual counterpart of the physical system, that can be used for simulations based on real-time acquired and synchronized data, focusing on production planning and control (Cimino et al., 2020, [Kritzinger et al, 2018], [Uhlemann et al, 2017]), is generated purely from production data with ML. Therefore it is able to learn correlations in the production system itself without time-consuming discrete modelling and to make predictions based on current and historical production data. Thus, the overall system advances contemporary approaches in terms of data quality and modelling as well as the timeliness of data and models.

The main contribution of the approach lies in the combination of signal acquisition and simultaneous data evaluation without the use of external cloud solutions in one system, resulting in near real-time analysis capability. Embedded systems are used for this purpose. These are computing machines that are largely hidden and integrated into a technical context (Gessler, 2014). This enables the use of the data for near real-time modelling of the production system and allows optimisation and adaptation of the PPC in very short control cycles. This is achieved by combining an ultra-modern microcontroller (MCU) with a single-board computer (e.g., Raspberry Pi 4, referred to as RPI), which has been miniaturized to the maximum possible extent according to the current state of the art. This composition allows signal acquisition with very high resolution and sampling rate while providing sufficient memory and computing power for complex pattern recognition and machine learning methods. Furthermore, this combination splits the computing effort into a preliminary and a main analysis. The Digital Twin of the production system is automatically generated from the data thus provided and enables the relationships between influencing and target variables to be quantified and made usable in daily operation.

This paper is structured as follows: In Section 2 we discuss related research. Section 3 introduces the design of the embedded system, i.e., details about the signal acquisition and analysis hardware as well as the concepts for the developed microcontroller application. In Section 4, we describe our concepts and architecure for Lean Data Analytics and the implemented machine learning concepts. Section 5 presents our full-fledged evaluation in laboratory as well as in industrial settings. The paper is finally concluded in Section 6

Section snippets

Related work

In this section, we give an overview of related research. Many efforts have been put into research of the different aspects of analysis and modelling of manufacturing in Industry 4.0. In Lee et al. (2014a) the authors overview the recent advances regarding Cyber Physical Production Systems (CPPS) identifying self-predictiveness and self-awareness as key characteristics to gain insight into digital production. It has been identified that several sources of information remain untapped.

Further

Signal acquisition and analysis hardware

The hardware concept of the sensor system comprises the parts shown in Fig. 1 with the main components (a) RPI, (b) CogniSense electronics with the MCU (ATSAMV71) and the other peripherals (low-pass filter, galvanic isolation, etc.) and (c) individual current sensors with BNC connection (e.g., Fluke i30s).

In order to be able to evaluate and operate the system in an industrial environment, a housing for top-hat rail mounting (switch cabinet) was designed and manufactured in an in-house 3D

Machine learning methods

Due to the heterogeneity of case studies in industry it became clear that it is necessary to cover the widest possible range of machine learning methods. The best results could be achieved by a few specific model types. Therefore two different types of machine learning methods were distinguished for the implementation: (i) Models trained and evaluated within the CogniSense system, and (ii) Externally trained models that are installed on the CogniSense system.

Two of the most established software

Laboratory evaluation

The complete system was successfully combined and tested in the Fraunhofer technical center in Bayreuth. A test facility for wear simulation developed for this purpose was retrofitted with the CogniSense system so that specific simulated patterns can be measured and detected. These can be recognized in the current profiles and the ML methods for deriving key figures can be trained with it. On the basis of this procedure, it was also possible to evaluate the CogniSense system in a laboratory

Conclusion and future work

In this paper the development and evaluation of an ML-based decentralized sensor system, called CogniSense ist described, which allows an automated analysis and monitoring of production plants and thus provides a near-realtime data basis for the realization of the Digital Twin. The evaluation of the prototypical system was based on the analysis of current profiles with regard to different wear conditions as well as on system failures and their classification. The extensive testing of the system

Authors’ contributions

Küfner, Thomas: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Project administration, Writing – Review & Editing

Schöenig, Stefan: Validation, Writing – Original Draft, Writing – Review & Editing, Supervision, Visualization, Project administration, Funding acquisition

Jasinski, Richard: Validation, Software, Resources

Ermer, Andreas: Project administration, Funding acquisition

Declaration of Competing Interest

The authors report no declarations of interest.

References (32)

  • E. Abele et al.

    Effiziente fabrik 4.0

    ZWF Zeitschrift für wirtschaftlichen Fabrikbetrieb

    (2015)
  • A. Canito et al.

    An architecture for proactive maintenance in the machinery industry

    International Symposium on Ambient Intelligence

    (2017)
  • A. Chan et al.

    Using bayesian networks to improve fault diagnosis during manufacturing tests of mobile telephone infrastructure

    J. Oper. Res. Soc.

    (2008)
  • A.K. Choudhary et al.

    Data mining in manufacturing: a review based on the kind of knowledge

    J. Intell. Manuf.

    (2009)
  • C. Cimino, E. Negri, L. Fumagalli, Review of digital twin applications in manufacturing, Comput. Ind....
  • R. Gessler

    Entwicklung eingebetteter Systeme. Vergleich von Entwicklungsprozessen für FPGA- und Mikroprozessor-Systeme Entwurf auf Systemebene

    (2014)
  • Cited by (15)

    • A sustainable deep learning framework for fault detection in 6G Industry 4.0 heterogeneous data environments

      2022, Computer Communications
      Citation Excerpt :

      They argued that the existing solutions for Industry 4.0 missed the incorporation of artificial intelligence and visualization techniques for better predictive maintenance operations. Kufner et al. [17] dealt with a large amount of data in production, and suggested a vertical-based solution while merging sensors data in a distributed cloud environment. Kaupp et al. [18] identified various kinds of contextual faults in a log of data collected from sensors in a smart factory system.

    • Towards edge computing in intelligent manufacturing: Past, present and future

      2022, Journal of Manufacturing Systems
      Citation Excerpt :

      This is an inevitable part of any manufacturing system and also requires real-time or near real-time analysis. The value of time interval for the term real-time is not clear as according to the ISO/IEC 2382, the term real-time describes the “the processing of data by a computer in connection with another process outside the computer according to time requirements imposed by the outside process" [88]. This term, real-time may vary from application to application.

    • Monitoring on a shoestring: Low cost solutions for digital manufacturing

      2021, Annual Reviews in Control
      Citation Excerpt :

      The data streams from the sensors are analysed at the sensor node, another peripheral node, a switch, or the gateway. Typically, the devices in the data collection network are constrained in terms of computing resources, which necessitates the use of various computation offloading strategies or lean methodologies for performing analytics within the edge (Küfner, Schönig, Jasinski, and Ermer (2021)). Nowadays, specialized hardware8 are being developed, which are capable of standalone edge analytics, even for high-speed, high-volume data streams.

    View all citing articles on Scopus
    View full text