Introducing HardwareBased Intelligence and Reconfigurability on Industrial IoT Edge Nodes Apostolos P. Fournaris, Christos Alexakos, Christos Anagnostopoulos, Christos Koulamas, and Athanasios Kalogeras ATHENA Research and Innovation Center draining a node from battery, memory, or p ­rocessing power. Editor’s note: Operations related to statistical Industrial automation is an important application area for the manipulation of collected senInternet-of-Things (IoT) technologies. In this article, Fournaris et al. discuss IoT system architectures for computing at the edge of the sor data on the time domain Internet and the use of FPGAs as edge devices. (like Pearson’s Convolution, Kur—Marilyn Wolf, Georgia Institute of Technology tosis) as well as industrial process anomalies and embedded The need for advanced intelligence in the machine-learning algorithms may introduce a very industrial environment may lead to moving intelli- high toll on an edge node’s expected behavior (e.g., gence from the cloud down to the local network level long life cycle, fast responsiveness, etc.). Also, addior directly into embedded devices. Thus, fog or edge tional edge node features, like security strengthencomputing emerges as a serious research challenge ing, introduce additional performance and hardware on edge nodes. This highlights the need for nodes that resource overhead that cannot always be handled by are capable of handling complex functions beyond the IIoT edge device without compromises in other simple sensing and actuation loops. Functionality device nonfunctional requirements. for local and embedded processing of collected In this article, we present a system architecture data to extract possible industrial decisions with- that can handle the overall manufacturing chain of out the need for cloud communication is becoming an Industry 4.0 plant and divide it into different laya reality in current and future edge-based Industrial ers, focusing on the level of intelligence that each Internet-of-Things (IIoT) deployments. However, this layer can handle. For this system, we propose a high edge node intelligence may have a considera- mechanism that can efficiently migrate IIoT comble footprint on a device’s hardware resources, thus putational functionality and edge intelligence to the system end node devices. Since such devices Digital Object Identifier 10.1109/MDAT.2019.2908547 do not always have the resources to support comDate of publication: 1 April 2019; date of current version: plex operations, we propose the executions of these operations through hardware means inside 22 July 2019. July/August 2019 Copublished by the IEEE CEDA, IEEE CASS, IEEE SSCS, and TTTC 2168-2356/19©2019 IEEE 15 Authorized licensed use limited to: University of Liverpool. Downloaded on January 18,2022 at 06:17:52 UTC from IEEE Xplore. Restrictions apply. SI: Circuits and Systems for VLSI Internet-of-Things Devices the node SoC. To bypass the lack of flexibility that ASICs have, we propose the use of FPGA technology on the end-node device. Recent advancements in FPGA technology enable us to use SoC core designs that merge powerful Advanced RISC Machine (ARM)-embedded processors with FPGA fabric. As an example of our proposed approach, we introduce a prototype concept of an edge node for our system that has hardware support for a statistical feature extraction mechanism and for detecting outlier behavior through novelty vector extraction using the extracted features and a training set of known good values. The nearest neighbor machine-learning algorithm is used for novelty extraction. Using this concept, we can support condition monitoring operations for industrial assets on the edge level. This approach was prototyped on an actual embedded system setting (Smartfusion2 SoC board combined with the CC2538-based Openmote board for wireless communication) and the two procedures (i.e., feature and novelty extraction) were realized in pure software and the proposed hardware or software codesigned way (using the Smartfusion2 SoC FPGA fabric). The experimental results confirmed that the proposed approach had considerable savings in consumed energy and led to very small time delays compared to traditional purely software solutions. Related works During the past years, the dominance of cloud computing was mainly driven by two main factors [1]. First, the exploitation of centralized resources guaranteed the reduction of administrative and operative costs. Second, companies preferred to consume services provided by large providers via the Internet than invest in acquiring a data center due to additional cost. However, the advent of the Internet of Things (IoT), with the continuous deployment of smart networking devices that generate large volumes of data, in conjunction with the physical separation of cloud services and end devices has, in many cases, increased the average network latency and jitter to unacceptable levels [2]. As a result, a new trend emerged that dictates the migration of cloud functionalities toward the network edges and the company premises [3]. Many attempts in the literature propose architectures that implement the aforementioned paradigm and introduce new terms like mobile cloud computing (MCC), cloudlet [4] fog computing, edge computing [5], which 16 is now called multiaccess edge computing, and mist computing [6]. All of these share their common effort to bring the computational resources nearer to the end user and differ in the targeted use cases. It must be noted that the edge computing acts complementary to the cloud and not competitively. Living in the dawn of a new cyber world, we realize that proximity matters, and its importance can be summarized into four reasons [1]: highly responsively services due to low latency and jitter, scalability via edge analytics, improved security and privacy, and enforcement of cloud redundancy policies. Industry 4.0 and the Industrial Internet Consortium (IIC)i reference architectures strongly support the inclusion of smart edge devices in an IIoT infrastructure. This means that edge devices should be able to handle intelligent functions beyond sensing and actuating. Also, edge devices must be able to collect data from multiple sensors in a real-time fashion. As indicated in [7] and [8], using an edge node micro control unit (MCU) for multiple sensor data acquisitions (acting as a multisensor core controller) may be a difficult task since MCUs are not focused on offering parallelism. Fully migrating all node functionality on FPGA technology (with no MCUs) tended to be problematic, since FPGAs do not have analog-to-digital and digital-to-analog converters (ADC and DAC) or floating point units. For this reason, the use of FPGA or complex programmable logic device (CPLD) in combination with MCUs (in the form of SoCs) has been proposed in order to migrate parallel processing for multisensory control to dedicated hardware components in the FPGA or CPLD fabric [7], [8]. This concept was expanded in [9], where a solid FPGA SoC-based solution is proposed. In that work, the main functionality of the embedded edge node is left to an ARM processor inside the FPGA SoC, while the device communication (using IEEE 802.15.4) is handled by a dedicated IP core inside the SoC’s FPGA fabric. Although the design in [9] is interesting, several aspects of the overall concept were not actually implemented, while the edge node intelligence is constrained on efficient communication management and realization. Similarly, in [7], the node intelligence is only focused on optimizing, through FPGA hardware components, the parallelism of a multiple sensor data acquisition core controller. Our view is that FPGA technology i https://www.iiconsortium.org/ IEEE Design&Test Authorized licensed use limited to: University of Liverpool. Downloaded on January 18,2022 at 06:17:52 UTC from IEEE Xplore. Restrictions apply. can be used to considerably increase the edge node intelligence, beyond the previously presented works. Overall architecture supporting edge detection The aim of the proposed approach is the introduction of hardware-assisted, yet reconfigurable, intelligence in integrated system IIoT edge nodes that will implement the Industry 4.0 aspects in a manufacturing environment. Such a system provides an overall solution for IIoT in manufacturing, by addressing the challenges at the various level of the automation pyramid. It supports an infrastructure for collecting data from the IIoT sensors while it permits the processing of this data using different algorithms with different complexity in different layers. The proposed overall system architecture is based on the IIoT Volume G1: Reference Architecture (IIRA) introduced by IIC in 2017 while defining the structural components of a system that follows the principles of Industry 4.0. IIRA separates the applications and systems into three tiers. First is the edge tier that collects data from the edge nodes and comprises IIoT sensors or ­actuators as well as controllers (end nodes) of the manufacturing devices at the plant level. Second is the platform tier whose applications are responsible to preprocess, transform, and analyze the data from the edge tier in order to forward specific information to the next tier or to reference behaviors and events on the manufacturing execution environment. Furthermore, these tier systems can execute commands to actuators or control devices for making changes in the production chain. Third is the enterprise tier that implements domain-specific applications and enterprise decision support systems that issue control commands to the platform tier and edge tier. With reference to the IIRA architecture, the proposed overall system architecture consists of various functional components distributed in the three tiers, as depicted in Figure 1. In our approach, the edge tier consists of IIoT sensors or actuators, either simple or smart, and controllers like programmable logic controllers (PLCs) for the operation of the actual industrial devices or machines in the plant. This tier hosts the proposed smart edge IIoT nodes, which can both collect data Figure 1. Overall architecture. July/August 2019 17 Authorized licensed use limited to: University of Liverpool. Downloaded on January 18,2022 at 06:17:52 UTC from IEEE Xplore. Restrictions apply. SI: Circuits and Systems for VLSI Internet-of-Things Devices and process it for timely event detection using hardware-assisted decision-making algorithm implementations. With the proposed edge nodes, we can execute at the end node level, fast calculations of statistical and machine-learning models on the collected raw data, inferencing accurately for the detection of an anomaly. Also, the nodes can execute predefined commands realizing the system response on the detected event. The response can include either automatic actions, such as a device shut down, where the interconnection with the device controllers is used, or signals through human interaction devices that inform the human employees about how to react in the case of a detected event or just passing the event to the platform tier in order to process it and react in a wider segment of production process. Apart from the intelligence that the edge tier has obtained, its nodes can preprocess data and send only specific and valuable information to the other two tiers, minimizing the size of the transmitted data, and eventually the time of the transmission. For example, instead of a raw time series data segment that has been sampled with some kilohertz sampling frequency, the edge node is able to do calculations and send values, such as mean and peak values, statistical moments, spectral components, etc., and far more intelligent information like malfunctioning events and machine anomalies. Furthermore, the edge tier could consist of classical IIoT sensors that monitor various variables and send values to the systems of the platform tier. Also, human interaction devices, such as buttons, that are used in case of alarm can propagate an event at this layer. In some cases, wherever IIoT sensors are not possible to interconnect immediately with the edge IIoT platform, due to the incompatibility of communication or data exchange protocols, gateways are used. These gateways are responsible for establishing the connection between sensors and systems at the next layer. The platform tier contains two main systems: 1) the IIoT platform and 2) the manufacturing control mediator (MCM). The IIoT platform collects data from different sensors, processes them, and infers event detection and diagnosis. The difference here from the edge tier is that the event detection utilizes the combined processing of data from different edge nodes. The IIoT platform is hosted in servers, physical or virtual, installed in the factory premises or on the cloud. In any case, IIoT platforms have more available computational power than the edge 18 nodes, permitting the execution of more complex algorithms for event detection and diagnosis that make use of more data. The MCM is a centralized integrated system interoperating with both the IIoT infrastructure systems and industrial systems in the plant. The main responsibility of this system is the execution of a command that consists of a flow of actions that must be executed in the production chain when an event is detected. In each case, the MCM must orchestrate the necessary changes in the process workflow of the production systems in order to make applicable required event-related changes in the production. Alexakos et al. [10] used a multiagent system decentralizing manufacturing execution system to play the role of MCM. The Visetak multiagent system (MAS), as the system is called, is based on the logic that each order coming from the enterprise resource system (ERP) layer is processed by an order agent which, by communicating with other agents, binds the appropriate factory resources (production devices and machines) that are controlled by their own resource agent. The scope is that, when a change is made in the plant, i.e., a machine stops, or its configuration changes, the MAS will reschedule the production chain with alternative flows and factory resources that are available. As agents define a distributed system architecture and do not require any predefined information about other agents, such an approach introduces fault tolerance features [11]. The enterprise tier hosts the classical enterprise information systems (ERP, CRM, etc.) in combination with modern systems for data analytics, mainly for big data that are finally collected by the IIoT and industrial systems infrastructure. The role of enterprise tier systems is to provide tools to the business analysts for analyzing data received from the factory, in order for the latter to infer abnormalities that affect the whole pro­duction line. This inference will lead to the changes in decision making at the enterprise level but will also define or update rules that are used for decision making in the other two tiers. Apart from the manual definition of rules for decision making, as already mentioned, the other two tiers (i.e., the edge nodes and IIoT platform) can provide resources for the execution of complex decision making or prediction algorithms. Thus, at this layer, using machine-learning or statistical methods on big data, new predictive models can be generated that can be forwarded to IIoT IEEE Design&Test Authorized licensed use limited to: University of Liverpool. Downloaded on January 18,2022 at 06:17:52 UTC from IEEE Xplore. Restrictions apply. platforms or the smart edge nodes to be used for data processing and inference. Hardware-based reconfigurable smart edge node for IIoT Focusing on the edge tier of Figure 1 system architecture, this article proposes an edge node architecture that utilizes hardware-based intelligence. Thus, an edge node can evaluate its data locally and provide appropriate responses without the need to always transmit data to the IIoT cloud, creating local intelligence loops without cloud or fog intervention. In several cases, where computer intensive or memory demanding intelligent operations are performed at the edge level, it would be impossible to retain high responsiveness when relying solely on software implementations. The same reason prohibits the inclusion of advanced features in the IIoT end nodes, such as strong security and privacy services. Hardware-assisted edge nodes, as proposed, can become a viable solution to the previously mentioned issues since commercial SoCs that include an MCU along with an FPGA fabric have been launched in the market. The edge node proposed architecture is structured around a synchronous SoC solution that includes an MCU processor and an FPGA fabric. The MCU has an advanced microcontroller bus architecture (AMBA) bus (APB, AHB, or AXI) that can communicate with all MCU peripherals [RAM–ROM memory, I/O interfaces (CAN, I2C, SPI)], which can also be used for the interconnection of dedicated IP Core peripherals, implemented in the FPGA fabric for specific tasks, thus boosting the edge node intelligence. We assume that the main software functionality, like operating systems (OSs) and communication, is handled by the edge node MCU. However, the edge node software stack includes specific drivers that can be used by an IIoT developer to utilize the FPGA fabric dedicated peripherals. In terms of hardware, each peripheral must be able to communicate with the interconnection bus, thus it includes a bus slave interface. The internal hardware structure of the FPGA peripherals is related to the peripheral’s actual f­unctionality. Determining what kind of peripherals need to be implemented in FPGA technology instead of software is highly relevant to the actual tasks that the edge node needs to perform. Analyzing the functional and nonfunctional requirements of the edge node behavior and extracting needed, achievable, performance indicators is the first step that we follow to match the previous goal. Components and functions needed for handling the edge node intelligence are described and modeled in software and are evaluated against the performance indicators. Those components that cannot match the indicators need to be hardware accelerated are described in hardware description language (HDL) and are implemented as IP cores in the FPGA fabric of the SoC. Through its SoC FPGA fabric, apart from hardware performance, the node gains the ability to reconfigure its hardware functionality over time according to the IIoT infrastructure needs. Thus, IIoT designers can update the edge tier functionality at the end node hardware level without changing the actual nodes themselves. The overall process of deploying both hardware and software on an edge node, along with the overall architecture of the node itself, is described in Figure 2. Initially, the desired edge tier functionality Figure 2. Design and deployment of reconfigurable hardware-assisted edge node intelligence on FPGA SoC-based architectures. July/August 2019 19 Authorized licensed use limited to: University of Liverpool. Downloaded on January 18,2022 at 06:17:52 UTC from IEEE Xplore. Restrictions apply. SI: Circuits and Systems for VLSI Internet-of-Things Devices that we want to deploy is defined in the enterprise tier. At this level, the hardware accelerated intelligence that needs to be deployed is decided. Then, the IIoT designer uses an edge node device architecture model and its connection with deployed sensors or actuators to describe the needed intelligence. The description is a mixture of block design, software code, and HDL development, which is determined by the SoC vendor toolset. The needed functionality described modeling is then simulated using SoC provider toolset, following SoC FPGA design and verification approaches, to validate the description compliance with the functional and nonfunctional edge node requirements. Successful verification leads to the actual hardware and software code implementation for edge node deployment. This set consists of application updates, new software drivers for the deployed hardware peripheral in the edge node FPGA fabric as well as the actual FPGA configuration file. This information is transmitted to the platform tier in the IIoT Platform that forward it to the appropriate edge tier device. When the deployment set reaches the edge node, it is installed to the SoC fabric. The actual installation is highly dependent on the SoC manufacturer provided technology. Currently, each SoC FPGA manufacturer provides his/her toolset for the installation procedure and supports different installation mechanisms that may need manual user intervention. Adaptive manufacturing and condition-based maintenance use case To further refine and describe the proposed generic methodology, using the proposed system architecture in the previous section, we focus on the design of a smart edge node capable of detecting malfunctions of an associated industrial asset (i.e., a production machine), offering basic condition monitoring services for preventive and predictive maintenance operational scenarios. Typically, an IIoT edge node sensor mechanism is installed on the industrial machine to collect data and transmit them to a fog gateway or the cloud for processing. Then, a cloud service will determine, from the collected data, some training data set (known good values), the health status of the machine, and detect or predict possible malfunction events. End node data are collected sparsely within a day (once per day) and can be used only for predicting long-term machine malfunctions, since, at this data acquisition rate, no short-term conclusions can be made. Part of the previously described func- 20 tionality can be transferred to the edge node if this node is capable of executing statistical analysis or machine-learning algorithms, thus extracting novelty features locally and evaluating them. This approach, showcased in [12], suffers from several implementation issues when software is used for event detection. RAM must be reserved on the edge node for collecting a considerable amount of data. Feature extraction and event detection execution have a significant execution latency while the detection window is small due to the reduced size of available memory for collecting sensor data. The overall software computation process remains power consumption-hungry, and thus, can only be executed a few times a week in order not to drain the edge node battery (in battery-powered systems). Therefore, the condition monitoring time-step can only support long-term maintenance prediction windows and cannot be exploited for other relevant and safety-related scenarios, such as to trigger a device response to remedy in real time other acute machine problems. In light of these issues, in this article, we apply the proposed architecture approach and transfer some of the edge node smart functionality from software to hardware. More specifically, we implement a dedicated hardware peripheral for extracting features from collected sensory data and extract novelty or outlier behavior using stored well-known good values (training set). These hardware components implement the same method utilized in [12], which processes the collected data in the time domain. Furthermore, an additional hardware component is used for data integrity code generation and validation using the hash-based message authentication code (HMAC) algorithm. As a prototype design of the mentioned class of edge nodes, we are using a two-processor board system structure, separating the application from the com­munication processor, coupled with several exchangeable sensor interface boards. The sampling, signal processing, and detection components are executed in an AVNET SmartFusion 2 kickstart development kit board featuring the Microsemi Smartfusion 2 SoC consisting of an ARM Cortex M3 processor and an FPGA fabric. The IoT networking stack, namely, the IEEE-802.15.4 wireless interface and the Internet Protocol version 6 (IPv6), Routing Protocol for Low-Power and Lossy Networks (RPL), and Constrained Application Protocol (CoAP) components, are implemented on a Contiki OS-based CC2538 IEEE Design&Test Authorized licensed use limited to: University of Liverpool. Downloaded on January 18,2022 at 06:17:52 UTC from IEEE Xplore. Restrictions apply. Figure 3. The embedded edge node prototype architecture. Openmote board that is connected through an Xbee shield to the AVNET system and communicates over using universal asynchronous receiver–transmitter (UART) (Figure 3). Using the dedicated peripherals on the FPGA fabric shown in Figure 3 architecture, we can collect sensor data on the dedicated storage area of the FPGA and feed them directly to the feature extraction (FE) unit that calculates seven different Features for windows of 1024 samples. The FE peripheral employs a direct memory access (DMA)-like mechanism for sample collection (16bit sample storage on dedicated FIFO memory) and using a software-based triggering, processes samples to extract rms, kurtosis, skewness, peak, crest factor, shape factor, and impulse factor features. The FPGA fabric also has a novelty extraction (NE) peripheral using a stored training dataset, and the extracted feature values are calculated using the nearest neighbor machine-learning algorithm (implementing matrix multiplication), outlier, and abnormal behavior for a sampled industrial device [12]. Figure’s 3 edge node functionality was prototyped using VHDL code (FPGA fabric) and C code software (ARM core) on the Smartfusion2 SoC. In Table 1, experimental results of the above-mentioned implementations are provided and comparisons are made when using only software (SW) and software–hardware (SW/HW) codesign for the FE window and NE window procedure. The ARM Cortex M3 datasheet specifies a power consumption of 31 μW/MHz, which results in 4.96-mW power consumption at ARM’s 160-MHz chip. The FPGA fabric without the FE/NE peripheral (pure SW) consumes 121.681 mW. The FE/NE peripheral inclusion adds 32.292 mW to Fabric’s power consumption. In Table 1, we can observe that the time delay for FE and NE on HW/SW solu- Table 1. Experimental results. July/August 2019 21 Authorized licensed use limited to: University of Liverpool. Downloaded on January 18,2022 at 06:17:52 UTC from IEEE Xplore. Restrictions apply. SI: Circuits and Systems for VLSI Internet-of-Things Devices tion is considerably decreased compared to pure SW solution. Also, energy consumption is considerably reduced using HW acceleration. The NE HW implementation has excessively low values due to the utilization of FPGA fabric’s dedicated multiplier units, which reduce time delay and energy consumption. Smartfusion 2 SoC has a dedicated hardware HMAC peripheral that we can use with no additional cost on HW resources and thus does not appear in Table 1. By using dedicated FPGA peripherals and considering the low power consumption specifications of the SmartFusion 2 SoC, we can achieve considerable energy consumption reduction compared to a pure software use case implementation. This enables us to increase the number of data acquisition and processing operations within a day (compared to the software solution of [12]) and monitor an industrial machine on the short term to achieve timely respond without IIoT cloud involvement. Long-term malfunction prediction using advanced data analytics performed at the enterprise tier using cloud mechanism is still supported in the above use case. Edge node devices must transmit processed data, such as novelty vectors and malfunction events, to the cloud infrastructure but since preprocessing of events is done locally cloud communication happens sparsely. Edge intelligence can be also enhanced by guaranteeing the integrity of data reaching the cloud by utilizing the SmartFusion 2 SoC dedicated HMAC peripheral. The proposed architectural interventions exploit all available reconfigurability mechanisms to improve the overall system intelligence and enhance its adaptability to the rapidly changing conditions in the modern manufacturing environment. The latest SoC approaches where reconfigurable logic is combined with hardcoded MCUs introduces a lot of capabilities on introducing adaptable hardware assisted intelligence on an IIoT end node and the IIoT system. In this article, we proposed such an approach that permits the processing of data using different algorithms with different complexity at different layers. Through a specific test case and its provided experimental results, we demonstrated that introducing intelligence on the IIoT edge using our approach is feasible and has significant benefits in condition-based maintenance by reducing detection speed and energy consumption. 22 Acknowledgments This work was supported in part by the Project “I3T—Innovative Application of Industrial Internet of Things (IIoT) in Smart Environments” under Grant MIS 5002434 through the Action for the Strategic Development on the Research and Technological Sector, in part by the Operational Programme “Competitiveness, Entrepreneurship, and Innova­ tion” under Grant NSRF 2014-2020, and in part by the Greece and the European Union through the European Regional Development Fund. References [1] M. Satyanarayanan, “The emergence of edge computing,” Computer (Long. Beach. Calif.), vol. 50, no. 1, pp. 30–39, Jan. 2017. [2] R. Roman, J. Lopez, and M. Mambo, “Mobile edge computing, Fog et al.: A survey and analysis of security threats and challenges,” Future Gener. Comput. Syst., vol. 78, pp. 680–698, Jan. 2018. [3] M. Chiang and T. Zhang, “Fog and IoT: An overview of research opportunities,” IEEE Internet Things J., vol. 3, no. 6, pp. 854–864, Dec. 2016. [4] Y. Wang, I.-R. Chen, and D.-C. Wang, “A survey of mobile cloud computing applications: Perspectives and challenges,” Wireless Pers. Commun., vol. 80, no. 4, pp. 1607–1623, Feb. 2015. [5] H. H. Pang and K.-L. Tan, “Authenticating query results in edge computing,” in Proc. 20th Int. Conf. Data Eng., pp. 560–571. [6] J. S. Preden et al., “The benefits of self-awareness and attention in Fog and Mist computing,” Computer (Long. Beach. Calif.), vol. 48, no. 7, pp. 37–45, Jul. 2015. [7] Q. Chi et al., “A reconfigurable smart sensor interface for industrial WSN in IoT environment,” IEEE Trans. Ind. Inform., vol. 10, no. 2, pp. 1417–1425, May 2014. [8] S. Bao et al., “FPGA-based reconfigurable data acquisition system for industrial sensors,” IEEE Trans. Ind. Inform., vol. 13, no. 4, pp. 1503–1512, Aug. 2017. [9] T. Gomes et al., “Towards an FPGA-based edge device for the Internet of Things,” in Proc. 2015 IEEE 20th Conf. Emerg. Technol. Factory Autom., pp. 1–4. [10] C. Alexakos et al., “IoT integration for adaptive manufacturing,” in Proc. 21st IEEE Int. Symp. RealTime Comput., 2018. [11] M. Y. Afanasev et al., “Modular industrial equipment in cyber-physical production system: Architecture and integration,” in Proc. IEEE 2017 21st Conf. Open Innov. Assoc., pp. 1–9. IEEE Design&Test Authorized licensed use limited to: University of Liverpool. Downloaded on January 18,2022 at 06:17:52 UTC from IEEE Xplore. Restrictions apply. [12] C. Alexakos et al., “Production process adaptation to IoT triggered manufacturing resource failure events,” in Proc. 2017 22nd IEEE Int. Conf. Emerg. Technol. works, and serious gaming. Anagnostopoulos has a Diploma in electrical and computer engineering from the University of Patras, Greece (2004). Factory Autom., pp. 1–8. Apostolos P. Fournaris is a Principal Researcher with the Industrial Systems Institute (ISI), ATHENA Research and Innovation Center, Patras, Greece. His research interests include hardware and software codesign, hardware attack resistance, WSN security, and trusted systems. He is a member of the IACR, the IEEE, the IEEE Computer Society, and the IEEE Circuits and Systems Society. Christos Alexakos is a Research Associate with the Industrial Systems Institute, ATHENA Research and Innovation Center, Patras, Greece. His research interests include knowledge management, information integration, and systems architecture in enterprise interoperability, manufacturing automation, Industrial Internet of Things, and cloud manufacturing. He is a program committee member of several international conferences and workshops. Christos Anagnostopoulos is a Collaborating Researcher with the Industrial Systems Institute, ATHENA Research and Innovation Center, Patras, Greece (since 2007). His current research interests include manufacturing systems, industrial net- July/August 2019 Christos Koulamas is a Research Director with the Industrial Systems Institute (ISI), ATHENA Research and Innovation Center, Patras, Greece. His research interests include real-time distributed embedded systems, industrial networks, wireless sensor networks and their applications in industrial and building automation, transportation, and energy sectors. He is a Senior Member of the IEEE and a member of the Technical Chamber of Greece. Athanasios Kalogeras is a Research Director with the Industrial Systems Institute, ATHENA Research and Innovation Center, Patras, Greece. His research interests include cyber-physical systems, Industrial Internet of Things, flexible manufacturing systems, collaborative manufacturing, industrial integration, and interoperability. He has a PhD in electrical and computer engineering from the University of Patras, Greece. He is a Senior Member of the IEEE. Direct questions and comments about this article to Apostolos P. Fournaris, Industrial Systems Institute, ATHENA Research and Innovation Center, Patra, Greece; fournaris@isi.gr. 23 Authorized licensed use limited to: University of Liverpool. Downloaded on January 18,2022 at 06:17:52 UTC from IEEE Xplore. Restrictions apply.