Statistical Process Monitoring for IoT-Enabled Cybermanufacturing: Opportunities and Challenges

Proceedings of the 20th World The International Federation of Congress Automatic Control The International Federation of Congress Automatic Control Proceedings of the 20th World Toulouse, France, July 9-14, 2017 Proceedings of the 20th World Congress Proceedings of the 20th World Congress Toulouse, France, July 9-14, 2017 Available online at www.sciencedirect.com The International Federation of Control The International International Federation Federation of of Automatic Automatic Control The Automatic Control Toulouse, France, July 9-14, 2017 Toulouse, France, July 9-14, 2017 Toulouse, France, July 9-14, 2017

ScienceDirect

Statistical Monitoring for IoT-Enabled Cybermanufacturing: PapersOnLine 50-1 14946–14951 Statistical Process Process IFAC Monitoring for(2017) IoT-Enabled Cybermanufacturing: Opportunities and Challenges Statistical Process Monitoring for IoT-Enabled Cybermanufacturing: Statistical for Cybermanufacturing: Opportunities and Challenges Statistical Process Process Monitoring Monitoring for IoT-Enabled IoT-Enabled Cybermanufacturing: Opportunities and Challenges Opportunities and Challenges , Opportunities and Challenges **. Jin Wang* Q. Peter He* Q. Peter He*,**. Jin Wang*

Devarshi Shah*.,, Nader Vahdat** Jin Wang* Q. Peter He* **. Devarshi Shah*. Nader Vahdat** Q. **. Jin Jin Wang* Wang* Q. Peter Peter He* He*,**. Devarshi Shah*. Nader Vahdat** Devarshi Shah*. Nader Vahdat** Devarshi Shah*. Nader AL Vahdat** *Department of Chemical Engineering, Auburn University, 36849USA (Q. P. H.: Tel: 334-844-7602; *Department of Chemical Engineering, Auburn University, AL 36849USA (Q. P. H.: Tel: 334-844-7602; e-mail: [email protected]; J. W.: [email protected]; D. S.: [email protected] ). *Department of Chemical Engineering, Auburn University, AL 36849USA (Q. P. H.: Tel: 334-844-7602; e-mail: [email protected]; J. W.: [email protected]; D. S.: [email protected] ). *Department of Engineering, Auburn University, AL (Q. P. 334-844-7602; ** Department of Chemical Engineering, Tuskegee University, AL 36806 (Q.Tel: P. H.: e-mail: *Department of Chemical Chemical Engineering, Auburn University, AL 36849USA 36849USA (Q.USA P. H.: H.: Tel: 334-844-7602; e-mail: [email protected]; J. W.: [email protected]; D. S.: [email protected] ). ** Department of Chemical Engineering, Tuskegee University, AL 36806 USA (Q. P. H.: e-mail: e-mail: J. D. ). [email protected]; N. V.: [email protected] )} e-mail: [email protected]; [email protected]; J. W.: W.: [email protected]; [email protected]; D. S.: S.: [email protected] [email protected] ). ** Department of Chemical Engineering, Tuskegee University, AL 36806 USA H.: e-mail: [email protected]; N. V.: [email protected] )}(Q. P. ** ** Department Department of of Chemical Chemical Engineering, Engineering, Tuskegee Tuskegee University, University, AL AL 36806 36806 USA USA (Q. (Q. P. P. H.: H.: e-mail: e-mail: [email protected]; N. V.: [email protected] )} [email protected]; N. N. V.: V.: [email protected] [email protected] )} )} [email protected]; Abstract: Initiated from services and consumer products industries, there is a growing interest in using Abstract: Initiated from services and consumer products industries, there is a growing interest in using Internet of Things (IoT) technologies in various industries. In particular, IoT-enabled cybermanufacturing Abstract: from services consumer products industries, there is growing interest in Internet ofInitiated Things (IoT) in various industries. In particular, IoT-enabled cybermanufacturing Abstract: Initiated from technologies services and and consumer products industries, there isareaaa usually growing interest in using using starts to draw increasing attention. Because IoT devices such as IoT sensors much cheaper and Abstract: Initiated from services and consumer products industries, there is growing interest in using Internet Things (IoT) technologies in various industries. In particular, IoT-enabled cybermanufacturing starts to of draw increasing attention. Because IoT devices such as IoT sensors are usually much cheaper and Internet of Things (IoT) technologies in various industries. In particular, IoT-enabled cybermanufacturing smaller the traditional sensors, in there is a industries. potential for instrumenting manufacturing systems with Internet than of Things (IoT) technologies various In particular, IoT-enabled cybermanufacturing starts to than draw the increasing attention. Because IoTa devices IoT sensors are usually much cheaperwith and smaller traditional sensors, there is potentialsuch for as instrumenting manufacturing systems starts increasing attention. Because such as are much massive number of sensors. The premise is IoT that devices the big data from IoT cheaper sensors and can starts to to draw draw increasing attention. Because IoT devices such subsequently as IoT IoT sensors sensorscollected are usually usually much cheaper and smaller than the traditional sensors, there aa the potential for instrumenting manufacturing massive number of sensors. The premise is is that big data subsequently collected from IoTsystems sensorswith can smaller than the traditional sensors, there is potential for instrumenting manufacturing systems with Therefore, data-driven statistical process monitoring (SPM) is be utilized advance manufacturing. smaller thantothe traditional sensors, there is a potential for instrumenting manufacturing systems with massive number of sensors. The premise is that the big data subsequently from IoT sensors Therefore, data-driven statistical collected process monitoring (SPM)can is be utilized to advance manufacturing. massive number of sensors. The premise is that the big data subsequently collected from IoT sensors can expectednumber to contribute significantly to theisadvancement of cybermanufacturing. In this work, state-ofmassive of sensors. The premise that the big data subsequently collected from IoT the sensors can data-driven statistical process monitoring is be utilized advance manufacturing. expected to to contribute significantly to theTherefore, advancement of cybermanufacturing. In this work, the(SPM) state-ofTherefore, data-driven statistical process monitoring (SPM) is be utilized to advance manufacturing. the-art in cybermanufacturing is reviewed; an IoT-enabled manufacturing testbed (MTT) Therefore, data-driven statistical technology process monitoring (SPM)was is be utilized to advance manufacturing. expected to contribute significantly to the advancement of cybermanufacturing. In this work, the state-ofthe-art in cybermanufacturing is reviewed; an IoT-enabled manufacturing technology testbed (MTT) was expected to contribute significantly to the advancement of cybermanufacturing. In this work, the state-ofbuilt to explore the potential of IoT sensors for manufacturing, as well as to understand the characteristics expected to contribute significantly to the advancement of cybermanufacturing. In this work, the state-ofthe-art cybermanufacturing reviewed; manufacturing (MTT) was built to in explore the potential ofis IoT sensors an forIoT-enabled manufacturing, as well as totechnology understandtestbed the characteristics the-art in is manufacturing technology testbed (MTT) was of data produced by the IoT sensors; finally, an theIoT-enabled potentials and challenges associated with big data analytics the-art in cybermanufacturing cybermanufacturing is reviewed; reviewed; an IoT-enabled manufacturing technology testbed (MTT) was built to explore the potential of IoT sensors for manufacturing, as well as to understand the characteristics of data produced by the IoT sensors; finally, the potentials and challenges associated with big data analytics built to to explore explore the potential potential of of IoT IoTsystems sensors is fordiscussed; manufacturing, aspropose well as as to to understand theanalysis characteristics presented by cybermanufacturing and weas statistics pattern (SPA) built the sensors for manufacturing, well understand the characteristics of data produced by the IoT sensors;systems finally, the potentials and and we challenges with big data analytics presented by cybermanufacturing is discussed; proposeassociated statistics pattern analysis (SPA) of produced by IoT finally, as a promising SPM tool forsensors; cybermanufacturing. of data data produced by the the IoT sensors; finally, the the potentials potentials and and challenges challenges associated associated with with big big data data analytics analytics presented by cybermanufacturing systems is discussed; and we propose statistics pattern analysis (SPA) as a promising SPM tool for cybermanufacturing. presented by cybermanufacturing systems is discussed; and we propose statistics pattern analysis (SPA) presented by cybermanufacturingInternet systemsof is discussed; andstatistical we propose statistics pattern analysis (SPA) Keywords: Cybermanufacturing, Things, sensors, process monitoring, faultreserved. detection, as a promising SPM tool for cybermanufacturing. © 2017, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights as a promising SPM tool for cybermanufacturing. Keywords: Cybermanufacturing, Internet of Things, sensors, statistical process monitoring, fault detection, as a promising SPM tool for cybermanufacturing. fault diagnosis, statistics pattern analysis Keywords: Cybermanufacturing, Internet fault diagnosis, statistics pattern analysis Keywords: Cybermanufacturing, Internet of of Things, Things, sensors, sensors, statistical statistical process process monitoring, monitoring, fault fault detection, detection, Keywords: Cybermanufacturing, Internet of Things, sensors, statistical process monitoring, fault detection, fault diagnosis, statistics pattern analysis fault diagnosis, statistics pattern analysis fault diagnosis, statistics pattern analysis 2015, 2012). One potential enabler for these advanced/smart 1. INTRODUCTION 2015, 2012). One potential enabler for these advanced/smart or cyber- manufacturing is industrial IoT. Industrial IoT 1. INTRODUCTION 2015, 2012). One potential for these or cybermanufacturing isenabler industrial IoT. advanced/smart Industrial IoT 2015, potential for advanced/smart 1. INTRODUCTION devices2012). for One manufacturing include 2015, 2012). One potential enabler enabler for these thesesensors/actuators, advanced/smart The concept of the Internet of Things (IoT) is not new as it was 1. or cyber- for manufacturing is industrial industrial IoT. Industrial IoT IoT 1. INTRODUCTION INTRODUCTION devices manufacturing include IoT. sensors/actuators, The concept of the Internet of Things (IoT) is not new as it was or cybermanufacturing is computers with wirelessis networks, etc., Industrial which have cyber- manufacturing industrial IoT. Industrial IoT first coined in 1999 in the MIT Auto-ID Center, which has or for include sensors/actuators, The concept concept of the the Internet ofMIT Things (IoT) is is not new new as it it was was computers withmanufacturing wireless networks, etc., which have first coined in 1999 in theof Auto-ID Center, which has devices devices for manufacturing include sensors/actuators, The of Internet Things (IoT) not as contributedfor significantly to different aspects sensors/actuators, of manufacturing manufacturing include since gained of tremendous and importance. The concept the Internetmomentum of Things (IoT) is not new asRecent it was devices with wireless networks, etc., which have first coined 1999 in the MIT Auto-ID Center, which has computers contributed significantly to different aspects of manufacturing since gainedin tremendous momentum and importance. Recent with networks, etc., first 1999 in Auto-ID Center, which such as automation and tracking. However, therewhich is one have area advances in in radio, mobile, and cloud technologies computers with wireless wireless networks, etc., which have first coined coined in 1999network, in the the MIT MIT Auto-ID Center, which has has computers contributed significantly to different aspects of manufacturing since gained tremendous momentum and importance. Recent such as automation and tracking. However, there is one area advances in radio, network, mobile, cloud technologies contributed significantly to different aspects of manufacturing since gained tremendous momentum and importance. Recent that has beensignificantly largely overlooked so faraspects – because industrial IoT to different of manufacturing have supported the development of the generation IoT contributed since gained tremendous momentum and first importance. Recent as automation tracking. there is one area advances in radio, mobile, and cloud technologies that has been largely and overlooked soHowever, far – because industrial IoT have supported the network, development of the first generation IoT such as automation and tracking. there is area advances in radio, mobile, and technologies devices as sensors usuallyHowever, much cheaper smaller such as such automation and are tracking. However, there and is one one area services and (Tarkoma and Ailisto, 2013). advances in products radio, network, network, mobile, and cloud cloud technologies such that has been largely overlooked so far – because industrial IoT have supported the development of the first generation IoT devices such as sensors are usually much cheaper and smaller services and products (Tarkoma and Ailisto, 2013). been largely far have thanhas sensors, so is a industrial potential IoT of that hasthe beentraditional largely overlooked overlooked sothere far – – because because industrial IoT have supported supported the the development development of of the the first first generation generation IoT IoT that devices such as sensors are usually much cheaper and smaller services and products (Tarkoma and Ailisto, 2013). than the traditional sensors, there is a potential of Initiated from services and consumer products industries, there devices such as sensors are usually much cheaper and smaller services and products (Tarkoma and Ailisto, 2013). instrumenting systems with massive number of sensors. The devices such as sensors are usually much cheaper and smaller services and products (Tarkoma and Ailisto, 2013). Initiated from services and consumer products industries, there than the traditional sensors, there is a potential of systems with massive number of potential sensors. The is a growing interest in using IoT technologies in various instrumenting than the traditional sensors, there is a big data from these IoT sensors be used of to thecollected traditional sensors, there iscanathen potential of Initiated from services consumer industries, there than is a growing interest and in using IoT products technologies in various Initiated from services and products industries, instrumenting systems with massive number of sensors. The big data collected from with thesemassive IoT sensors can of then be usedThe to industries. Many countries have invested significantly onthere IoT instrumenting Initiated from services and consumer consumer products industries, there systems number sensors. advance manufacturing. Currently this type of industrial IoT instrumenting systems with massive number of sensors. The is a growing interest in using IoT technologies in various industries. Many countries have invested significantly on IoT is aa growing interest using various data collected from these IoT sensors can be used to advance manufacturing. Currently this type ofthen industrial IoT initiatives based on thein that technologies IoT can be anin is growing interest inpremise using IoT IoT technologies ineffective various big big data collected from these IoT sensors can then be used application has notfrom drawnthese much attention academic big data collected IoT sensorsfrom can either then be used to to industries. Many countries have invested on IoT initiatives based on the premise that IoT significantly can be an effective industries. Many countries have invested significantly on IoT advance manufacturing. Currently this type of industrial IoT has not drawnCurrently much attention fromofeither academic way to improve traditional physicalsignificantly and information industries. Many countries have invested on IoT application advance manufacturing. this type industrial IoT researchers or industrial practitioners. One possible reason is advance manufacturing. Currently this type of industrial IoT initiatives based on the premise that IoT can be an effective way to improve traditional physical and information initiatives based based on the the premise premise thathave IoT acan can be an an effective effective not drawn much attention from either academic researchers has or industrial practitioners. One possible reason is technology infrastructure, and will significant impact application initiatives on that IoT be has drawn much either academic that the benefits of such havefrom not been application has not not drawnapplications much attention attention from eitherrecognized academic way to improve traditional physical and information technology infrastructure, and will have a significant impact application way to improve traditional physical and information or industrial practitioners. Onenotpossible reason is that the benefits of such applications have been recognized on productivity and innovation. way to improve traditional physical and information researchers or possible reason or tested. Therefore, in practitioners. this work, weOne built an IoT-enabled researchers or industrial industrial practitioners. One possible reason is is technology infrastructure, and will have aa significant impact researchers on productivity and innovation. technology the benefits of such applications have not been recognized or tested. Therefore, in this work, we built an IoT-enabled technology infrastructure, infrastructure, and and will will have have a significant significant impact impact that the of not recognized manufacturing testbed have (MTT) to explore the that the benefits benefits technology of such such applications applications have not been been recognized on productivity Despite the factand thatinnovation. the industrial IoT is still in its infancy, that on productivity and innovation. tested. Therefore, in this work, we built an IoT-enabled manufacturing technology testbed (MTT) to explore the on productivity Despite the factand thatinnovation. the industrial IoT is still in its infancy, or tested. work, built IoT-enabled potential industrial in IoTthis sensors, as an to gain a better or tested.ofTherefore, Therefore, in this work, aswe wewell built an IoT-enabled many applications are being developed and deployed in or manufacturing technology testbed (MTT) to explore the potential of industrial IoT sensors, as well as to gain a better Despite the fact that the industrial IoT is still in its infancy, many applications are being developed and deployed in manufacturing technology testbed (MTT) to explore the Despite the fact that the industrial IoT is still in its infancy, understanding on the characteristics of the big data produced manufacturing technology testbed (MTT) to explore the various industries including healthcare inventory and supply Despite the fact that the industrial IoT is still in its infancy, potential of industrial IoT sensors, as well as to gain a better many applications are being developed and deployed in understanding on the characteristics ofwell the as bigtodata produced various industries including healthcare inventory and supply potential of industrial IoT sensors, as gain a better many applications are being developed and deployed in by the IoTofsensors. industrial IoT sensors, as well as to gain a better chain management,are transportation, workplace and home many applications being developed and deployed in potential on the characteristics of the big data produced various industries including healthcare inventory and supply by the IoT sensors. chain management, transportation, workplace home understanding various including healthcare inventory and understanding on on the the characteristics characteristics of of the the big big data data produced produced support, security, and surveillance, etc. (Xu et al., 2014). various industries industries including healthcare inventory and supply supply understanding by the IoT sensors. With future cybermanufacturing equipped with massive IoT chain management, transportation, and home support, security, and surveillance, etc.workplace (Xu et al., 2014). by the IoT sensors. chain management, transportation, workplace and home thefuture IoT sensors. With cybermanufacturing equipped with massive IoT chain management, transportation, workplace and home by it is expected that the data produced by these support, security, and surveillance, etc. (Xu et al., 2014). In manufacturing, advanced/smart manufacturing and sensors, support, security, etc. et future cybermanufacturing with massive IoT sensors, it is expected that theequipped data produced by these support, security, and and surveillance, surveillance, etc. (Xu (Xu et al., al., 2014). 2014). and With In manufacturing, advanced/smart manufacturing future with IoT manufacturing systems will growequipped exponentially into so call With future cybermanufacturing cybermanufacturing equipped with massive massive IoT cybermanufacturing are drawing increasing attention as well. With sensors, it is expected that the data produced by manufacturing systems will grow exponentially into sothese call In manufacturing, manufacturing and cybermanufacturing areadvanced/smart drawing increasing attention as well. sensors, it is expected that the data produced by these In manufacturing, advanced/smart manufacturing and data”it (Qin, 2014). Four (4V’s)produced are oftenbyused to sensors, is expected that V’s the data these The essence of these trends is the application of increasingly In manufacturing, advanced/smart manufacturing and “big systems into so call “big data” (Qin, 2014).will Fourgrow V’s exponentially (4V’s) are often used to cybermanufacturing drawing attention as well. manufacturing The essence of theseare trends is theincreasing application of increasingly manufacturing grow into call cybermanufacturing are increasing as thesystems essence will of big dataexponentially (Zikopoulos et al.,so manufacturing systems will grow exponentially into so2012; call powerful and low-cost computation and networked cybermanufacturing are drawing drawing increasing attention attention as well. well. characterize “big data” (Qin, 2014). Four are often to The essence of these trends is the application of increasingly the essence of big V’s data (4V’s) (Zikopoulos et al.,used 2012; powerful and low-cost computation and networked characterize “big data” (Qin, 2014). Four V’s (4V’s) are often used to The essence of these trends is the application of increasingly and Eaton, Volume (theare size/scale of the “big data” (Qin, 2014).2011): Four V’s (4V’s) often used to information-based technologies in application manufacturing enterprises. Zikopoulos The essence of these trends is the of increasingly characterize the essence of big data (Zikopoulos et al., 2012; powerful and low-cost computation and networked Zikopoulos and Eaton, 2011): Volume (the size/scale of the information-based technologies in manufacturing enterprises. characterize the essence of big data (Zikopoulos et al., 2012; powerful and low-cost computation and networked Varietythe (the form/format the(Zikopoulos data), Velocity (the rate characterize essence of big of data et al., 2012; There is aand general consensus that factories plants data), powerful low-cost computation and and networked Zikopoulos and Eaton, 2011): Volume (the size/scale of rate the information-based technologies manufacturing Variety (the form/format of the data), Velocity (the There is a general consensusin that factories enterprises. and plants data), Zikopoulos and Volume (the the information-based technologies in enterprises. the data being 2011): produced), Veracity of Zikopoulos and Eaton, Eaton, 2011): Volume and (the size/scale size/scale of (the the connected to the Internet are more efficient, productive and of information-based technologies in manufacturing manufacturing enterprises. data), Variety (the form/format of the data), Velocity (the rate of the data being produced), and Veracity (the There is a general consensus that factories and plants connected to the Internet are more efficient, productive and data), Variety (the form/format of the data), Velocity (the rate There is a general consensus that factories and plants uncertainty/reliability of the data). Such big data will present Variety (the form/format of the data), Velocity (the rate smarter isthan non-connected counterparts et al., data), There a their general consensus that factories(Davis and plants produced), Veracity (the uncertainty/reliability data). Suchand big data will present connected to their the Internet are more efficient, productive smarter than non-connected counterparts (Davis et and al., of of the the data data being beingof the produced), and Veracity (the connected the data being produced), and Veracity (the connected to to the the Internet Internet are are more more efficient, efficient, productive productive and and of uncertainty/reliability of the data). Such big data will present smarter than their non-connected counterparts (Davis et al., uncertainty/reliability of the data). Such big data will present smarter than their non-connected counterparts (Davis et al., uncertainty/reliability of the data). Such big data will present smarter than their non-connected counterparts (Davis et al.,15511 Copyright © 2017 IFAC Copyright © 2017, 2017 IFAC 15511 2405-8963 © IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Copyright © 2017 IFAC 15511 Peer review under responsibility of International Federation of Automatic Control. Copyright © © 2017 2017 IFAC IFAC 15511 Copyright 15511 10.1016/j.ifacol.2017.08.2546

Proceedings of the 20th IFAC World Congress Toulouse, France, July 9-14, 2017 Q. Peter He et al. / IFAC PapersOnLine 50-1 (2017) 14946–14951

new challenges and opportunities to data-driving applications for manufacturing industries, including fault detection and diagnosis, process optimization and control, predictive maintenance, etc. In this work, based on the findings obtained from a manufacturing technology testbed (MTT) system, we focus on the challenges that cybermanufacturing presents to statistical process monitoring (SPM), with applications in fault detection and diagnosis. The rest of the paper is organized as follows: Section 2 presents the design and configuration of the MTT system and the characteristics of the data obtained from various experiments on the IoT-enabled MTT. Section 3 briefly reviews the state-of-the-art in SPM, and presents statistics pattern analysis (SPA) as a candidate for the SPM of cybermanufacturing. Section 4 discusses the potentials and challenge of the SPM for cybermanufacturing and how SPA could meet the potentials and address the challenges. Section 5 draws conclusions.

14947

temperature sensors were then fixed to the rods at different heights. Fig. 1 shows the setup of the IoT sensors within the CSTR made of transparent polycarbonate material. For the data management system, the final architecture follows the publish/subscribe model enabled by the MQTT protocol, a lightweight messaging transport protocol that allows for a small code footprint while utilizing minimal bandwidth. More information about data acquisition and storage can be found in (Shah et al., 2017).

2. IoT-ENABLED MANUFACTURING TECHNOLOGY TESTBED (MTT) 2.1 MTT Design and Setup

Fig. 1. IoT sensors in the CSTR system

Simulation is a powerful, flexible tool often utilized by control engineers to understand complex dynamic systems and to test out new algorithms. However, the fidelity of the simulated system is limited by the understanding on the system, (i.e., the model that describes the system). Currently, industrial IoT is still in its infancy and there is insufficient understanding on the property, capacity and performance of IoT sensors to enable accurate simulation. Therefore, in this work, instead of relying on simulation to generate the big data delivered by IoT devices, we developed an IoT-enabled MTT to understand the properties and characteristics of IoT sensors, as well as to identify the challenges and opportunities presented by IoTenabled manufacturing systems. Currently available IoT devices on the market are mainly for daily use consumer products such as cell phones and home security systems, and limited options are available for industrial applications. Based on the availability, functionality, cost and potential industrial applications, we decided to use temperature sensors to develop the IoT-enabled MTT system, which is a continuous stirred tank reactor (CSTR) equipped with 28 IoT sensors (water proof DS18520 IoT temperature sensors), plus corresponding data acquisition, transmission and storage systems. As discussed earlier, these IoT sensors are small and easy to embed. Therefore, they offer the opportunities to instrument systems with mass number of sensors. With 28 sensors, the IoT-enabled MTT allows us to measure the temperature distribution within the reactor directly, without assuming ideal mixing. In the final design of the IoT-enabled CSTR system, sensors are placed in three different levels: top, middle and bottom levels in the tank; in each level, sensors are distributed uniformly across the cross-sectional area of the tank. The selected design not only allows for the capture of non-ideal mixing within the tank, but also allows easy scale up. The corresponding sensor housing unit was designed and fabricated in house. The base of the sensor housing unit (i.e., the tank lid) was fabricated using 3D printing, and 12 rods were inserted into the base and fixed with epoxy. IoT

2.2 Designed Experiments and Findings In order to test the functionality of the IoT-enabled MTT, as well as to gain better understanding of the behavior of the IoT sensors, steady state behavior and step response of the MTT were tested and analyzed. Additional experiments have been designed and are currently being conducted to investigate the mixing pattern on heat (mass) transfer within the CSTR. Steady-state behavior In order to study the behavior of IoT sensors, we first collected data over a period of time from the CSTR that is stabilized at room temperature. Fig. 2 shows the data collected from 14 sensors, with results from seven sensors on each subplot. In Fig. 2, the bold red line represents the reference temperature obtained using a mercury thermometer; while the other thin color lines represent measurements obtained from different IoT sensors.

(a)

(b)

Fig. 2. Steady-state behavior of IoT temperature sensors Fig. 2 shows that the IoT sensor responses are quite different from the traditional mercury thermometers in the following aspects. First, many IoT sensors exhibit noisy behaviour at steady state. The spikes maybe because the analog reading happens to be in the vicinity of the middle of two grids. Second, the sensor readings fall on fixed grids. In most of the cases, the IoT readings change in the multiples of 0.0625 oC; a couple of IoT sensors changes in the multiples of 0.5 oC, as the

15512

Proceedings of the 20th IFAC World Congress 14948 Q. Peter He et al. / IFAC PapersOnLine 50-1 (2017) 14946–14951 Toulouse, France, July 9-14, 2017

orange line shown in Rpi-3 in Fig. 2. This is most likely caused by the precision of the built-in analog-to-digital converter. Although it is expected that the grid will become finer when more digits are used in the future IoT sensors for higher precision, such grids will continue to exist. Third, the sensors all show different levels of persistent bias over time, which is likely caused by the sensing element used within the IoT devices. Traditional sensors are usually equipped with certain calibration mechanisms, which can correct any drifts or shifts that might occur to the sensing element before each experiment. However, there is no calibration option for IoT sensors. As a result, the IoT sensors would have biases which may change over time. These characteristics should be considered when developing new data analytics approaches. Step response In these experiments, the temperature in the reactor was initially at room temperature (between 20 oC to 22 oC) ; after 5~10 minutes of sampling, the water temperature was suddenly changed to 38 oC (hot step) or 4.5 oC (cold step). Such step changes were achieved through moving the sensors and their housing unit together into a duplicated identical reactor with the same amount of hot or cold water. Data collection continued even during the switch. After the switch, another 20~30 minutes of samples were collected. Fig. 3 (a) and (b) show the step responses obtained through seven IoT sensors during hot and cold step changes, respectively. Fig. 3 (c) shows the zoom-in of the transient response during the hot step change (i.e., Fig. 3 (a)); while Fig. 3 (d) shows the zoomin of a later stage after the cold step change (i.e., Fig. 3 (b)). The gradual decrease or increase of the temperature after a step change was due to the heat transfer between the reactor and the ambient environment.

(a)

(b)

(c)

(d)

is similar to traditional sensors; (2) for gradual temperature changes, IoT sensors show “stiction” behaviour. In other words, the temperature change, either increase or decrease, has to be larger than a certain threshold before the sensor readings change. Such behavior is similar to a sticky valve. The grid of the sensor reading is similar to their steady state behavior, in the multiple of 0.0625 oC or 0.5 oC. Sampling interval Besides different dynamic responses between IoT and traditional temperature sensors, another major difference is the sampling frequency or sampling interval. In our current configuration, the sampling frequency is mainly dictated by the IoT sensors. Whenever an IoT sensor sends a reading to the data collector, it will be collected and stored in the database. Therefore, samples are not collected at a fixed sampling interval, instead, they are collected at various time intervals. This can be seen from the zoom-in plot in Fig. 3(c). As a result, over the same period of time, different sensors will provide different number of readings. The sampling interval distribution for sensor #2 and sensor #8 are given in Fig. 4. Over the same sampling period (5960 seconds), sensor #2 and #8 collected 7022 and 6982 samples, with a sampling frequency of 1.178 Hz and 1.171 Hz, respectively. The main reason for such distributions is due to the time needed by the hardware (IoT sensor) to generate a response, and the specification of the digital sensors usually provides information on how fast the sampling frequency could be. However, the software does play a significant role as well. For example, in our initial configuration, only one data collector (a function) was implemented to collect data from all the sensors, which resulted in decreased sampling frequency when more sensors are plugged into the network. After we switched to multi-agent setup where each data collector is responsible to collect data from one specific sensor, we were able to maintain the sampling frequency, no matter how many sensors were plugged into the network.

(a)

(b)

Fig. 3. Step responses of the IoT sensors

Fig. 4. Sampling frequencies vary among IoT sensors. The time interval is in second.

From Fig. 3 we can observe the following: (1) different IoT sensors have slightly different time constants and delays. The time constants estimated from their step response ranges from ~2.9 to ~5.3 second, with a mean of ~4.1 second. This behavior

Discussion Although additional experiments are on-going, these initial tests have revealed several major differences between IoT and traditional sensors. In particular, the inherent bias (which can

15513

Proceedings of the 20th IFAC World Congress Toulouse, France, July 9-14, 2017 Q. Peter He et al. / IFAC PapersOnLine 50-1 (2017) 14946–14951

be quite large depending on the device) and varying sampling frequency seem to stand out the most, and require additional consideration/treatment in order to use the data collected by the IoT sensors for process monitoring. For the sensor biases, one may argue that taking average of the readings across different sensors will likely remove the bias, as long as the number of sensors is large enough. However, taking average essentially throws away the additional information provided by the large number of sensors on the spatial distribution of the process properties, such as heat transfer or mass transfer. For varying sampling frequency, the easiest way to handle it would be down sampling, and assume zero-order hold between two adjacent readings. Again, this approach would lose the benefit of fast sampling provided by the IoT devices. In the case of processes with fast dynamics, it is desirable to get updated information as frequently as possible. Obviously, practical and effective approaches are needed to address these new characteristics of data produced from IoT sensors. In the next section, we present a brief review of statistical process monitoring, discuss the challenges that the current solutions are facing, and introduce a statistical pattern based framework as an initial attempt to address the current and new challenges. 3. STATISTICAL PROCESSING MONITORING FOR MANUFACTURING The massive amount of data that will be produced by the future cybermanufacturing systems offer the potential of significantly improved manufacturing, as they will provide information that is not available in current manufacturing systems. However, how to extract useful information from the "big data” presents significant challenges to process monitoring. Here we briefly review the history of data-driven process monitoring, i.e., statistical process monitoring (SPM). SPM has been widely employed in manufacturing since the first generation SPM (i.e., statistical process control, or SPC) was pioneered by Walter Shewhart at the Bell Laboratories in the 1920s (Oakland, 2007). One of the most noticeable examples of SPC’s contribution to manufacturing is the great success of Japanese manufacturing in the 1970s (Oakland, 2007). In the 1980s, it was recognized that process variables (e.g., process temperature and pressure) and their correlations to product quality, which were not utilized by SPC, could provide extra benefit to process monitoring (MacGregor and Kourti, 1995). The utilization of both process and quality variables led to the birth of the 2nd generation SPM or multivariate SPM (MSPM). Principal component analysis (PCA), partial least squares (PLS), and their variants in many ways form the basis of MSPM. MSPM was a significant leap forward in terms of fault detection performance as compared to SPC, especially for large and complex chemical processes where process variables are highly correlated due to the physical and chemical laws and principles that govern the process such as mass and energy balances, thermodynamics and reaction kinetics. As a result, MSPM methods have become the industrial standard methods for process monitoring – they have been widely implemented in many process industries and numerous successful stories have been reported (Qin, 2012, 2003).

14949

However, the success of MSPM has been limited by its underlying assumption of a static or steady-state linear process and a multivariate Gaussian distributed state variables. Driven by the frequently changing market demands, fewer and fewer processes are operating continuously around a fixed operation point, where the assumptions of the linear steady-state process and Gaussian distributed data are satisfied. Therefore, new methods are needed to effectively monitor dynamic nonlinear process with non-Gaussian distributed data. We expect that the new approaches, termed as the 3rd generation methods, should offer a more direct and generally applicable way to handle process dynamics, nonlinearity and nonnormality. To this end, nonlinearity and/or process data non-Gaussianity, dynamic variations of PCA/PLS, such as dynamic PCA/PLS (DPCA/DPLS), dynamic-inner PCA/PLS (DiPCA/DiPLS), nonlinear variations of PCA/PLS, such as kernel PCA/PLS (KPCA/KPLS), and other statistical methods, such as independent component analysis (ICA), have been proposed (Dong and Qin, 2015; Galicia et al., 2012; Kano et al., 2003; Ku et al., 1995; Lee et al., 2004a, 2004b, 2004c). From a different viewpoint, the authors proposed a new SPM framework termed statistics pattern analysis (SPA) for continuous and batch process monitoring as shown in Fig. 5 (a) and (b)(He and Wang, 2011; Wang and He, 2010). In the SPA framework, instead of monitoring the process variables, the statistics patterns, which consists of various process statistics, are monitored to perform fault detection. In this way, various statistics, including higher-order statistics, can be used to capture the process characteristics that cannot be captured by PCA such as nonlinearity and non-Gaussianity. With additional information other than variance-covariance structure extracted from the process data, the SPA method is able to detect faults that are difficult or cannot be detected by the traditional SPM methods. It has been shown that for continuous process monitoring, SPA can deliver much enhanced performance for fault detection and diagnosis, particularity for processes with strong nonlinearity (Wang and He, 2010). In addition, for batch process monitoring, SPA not only can eliminate the data preprocessing steps needed by the 2nd generation MSPM methods, but also can deliver significantly improved monitoring performance (He and Wang, 2011). Because SPA relies on various statistics of process variables to characterize the state of a process, we believe it has the potential to address the new challenges presented by cybermanufacturing systems. This will be discussed further in the following section.

(a)

(b) Fig. 5 Statistics pattern analysis (SPA) based SPM framework

15514

Proceedings of the 20th IFAC World Congress 14950 Q. Peter He et al. / IFAC PapersOnLine 50-1 (2017) 14946–14951 Toulouse, France, July 9-14, 2017

for (a) continuous processes and (b) batch processes 4. ADDRESSING CYBERMANUFACTURING DATA CHARACTERISTICS Big data analytics is arguably a major focus in cybermanufacturing, and could become a key basis of competitiveness, productivity growth, and innovation (Manyika et al., 2011; O’Donovan et al., 2015; Qin, 2014). In this section, we discuss some challenges process monitoring could face in addressing the 4 V’s of big data generated in cybermanufacturing: volume, variety, velocity and veracity. We also discuss the potential of SPA in addressing them. For volume: Cybermanufacturing will generate a massive amount of data, both due to increased sensoring (more variables) and increased sampling (more observations). In addition, process nonlinearity and process data nonnormality will become more dominant or even become norm instead of exception for cybermanufacturing. Generally speaking, more observations do not pose a problem to existing MSPM methods or SPA. In fact, more observations are beneficial for SPA because they allow better estimation of statistics. However, significantly increased variables in cybermanufacturing is more difficult to handle than just large number of observations, particularly when different variables are sampled at different frequencies. Since SPA uses a window approach to compute different statistics, it can naturally handle the different sampling frequency by using all samples available within the window to compute the statistics. One concern associated with window approach is that detection delay. However, we have shown that the detection delay associated with SPA is similar or even shorter than that of PCA (Wang and He, 2010). This is mainly due to three reasons: first, although calculating statistics would average out the change presented by the last data point in the window, the standard deviation (i.e., control threshold) is also reduced by √𝑛𝑛𝑛𝑛 where n is the number of samples used to calculate the statistics, which allows tighter control limit for faster fault detection. Second, for second and higher order statistics that may play a key role in monitoring nonlinear processes, square and cubic of the process variable are involved, which would significantly amplify the change carried by the newest data point in a window. Finally, since the statistics follow a Gaussian distribution to a much better degree than process variables due to the CLT, the control limit is much tighter than that for process variables in 2nd generation MSPM methods, particularly for nonlinear processes. Note that the inherent bias among different IoT sensors that measure the same process property, such as the temperature sensors in the MTT system, does not pose significant challenge to SPA, as such effect is limited to variable mean only, without affecting other statistics. In addition, if the bias does not vary over time, it will not significantly affect data-driven SPM as the bias will be taken into account in the model that is built based on the historical data that contain the bias. Finally, effective variable (statistics) selection will play a key role to improve SPA’s performance, and will help address the above discussed challenge of large number of variables to certain extent, but more likely drastically new approaches are needed to fundamentally address these challenges. We envision that

some alternative process/data representations will emerge that utilize the complete set of variables rather than the filtered or pre-selected variables. For variety: Cybermanufacturing could generate different forms of data, monitor different parts of the system, measure different phases of the process, which can be sampled at very different frequencies. Current MSPM approaches (e.g., PCA) usually make use of one type of data at a time because the data have to form a matrix or matrices. In this aspect, SPA has the potential to integrate different data types relatively easily, because the statistics of different data types (such as machine data and metrology data) are much easier to integrate than the original data. Note that different statistics from different data forms or different parts of the process are simply augmented together for process monitoring; there is no restriction on the number of the selected statistics. If PCA is applied to monitor the statistics (i.e., SP), the same sampling frequency of statistics are required. This can be easily satisfied by using all samples available for a given variable within a fixed window to calculate the desired statistics. In addition, we have shown that the information contained in the raw data would be altered more or less during data pre-processing, which may negatively impact the monitoring performance (He and Wang, 2011). Therefore, minimizing required data pre-processing would be desirable for cybermanufacturing systems. For velocity: In cybermanufacturing, it is expected that streaming or online models will often be used for process monitoring (e.g., diagnostics and prognostics) and control; To address large volume of streaming data for real-time statistical analysis and online monitoring, SPM methods that do not require data pre-processing such as SPA, or methods that require minimum and automated data preprocessing, will have advantage in handling steaming data. For veracity: As shown in this paper, for cybermanufacturing, it is expected that the veracity of the data (e.g., data quality or cleanness such as missing data, outliers, noises, delays and data asynchronism) will be significant. While the traditional MSPM methods emphasize the cleanness of the data to prevent potential misleading conclusions, it is expected that the future MSPM methods should consider data errors or messiness as unavoidable, and are robust to the imperfections in the data (Qin, 2014). Because missing data, outliers, or data uncertainty has much less impact on various statistics than variable themselves, SPA offers advantages in this aspect. For example, the mean of a variable will not be affected significantly by the noise or infrequent missing data or outliers. Despite the advantages that SPA possesses, it also has certain limitations. For example, how to optimally select the statistics to effectively monitor a process is still an open question. For cybermanufacturing, with the increase of measured variables, available statistics that can be used for process monitoring increases exponentially (e.g., considering cross-variable statistics among different variables). If all available statistics are included for process monitoring, it not only significantly increase the computation load unnecessarily, but also could deteriorate the monitoring performance. 6. CONCLUSIONS In manufacturing, advanced/smart manufacturing and cybermanufacturing are drawing increasing attention. One

15515

Proceedings of the 20th IFAC World Congress Toulouse, France, July 9-14, 2017 Q. Peter He et al. / IFAC PapersOnLine 50-1 (2017) 14946–14951

potential enabler for these advanced/smart or cybermanufacturing is industrial IoT, with the premise that the big data subsequently collected from IoT sensors can be utilized to advance manufacturing. in this work, we built an IoTenabled manufacturing technology testbed (MTT) to explore the potential of industrial IoT sensors, as well as to gain a better understanding on the characteristics of the big data produced by the IoT sensors. The IoT-enabled MTT system is a continuous stirred tank reactor (CSTR) equipped with 28 IoT sensors, plus corresponding data acquisition, transmission and storage systems. Steady state behavior and step response of the MTT were tested and analyzed to examine the functionality of the IoT-enabled MTT, as well as to gain better understanding of the behavior of the IoT sensors. The IoT sensors exhibit the following characteristics: 1) many IoT sensors exhibit noisy or spiky behaviour at steady state; 2) IoT sensor readings fall at fixed grids; 3) most IoT sensors show different levels of persistent bias; 4) samples are collected at variable time intervals. We propose SPA based framework for cybermanufacturing process monitoring. Because SPA relies on various statistics of process variables to characterize the state of a process, we believe it has the potential to address the challenges exhibited in the data collected from IoT sensors. Specifically, for volume, more observations are beneficial for SPA because they allow better estimation of statistics. For variety, SPA has the advantage that different statistics from different data forms or different parts of the process can be simply augmented together for process monitoring. For velocity, because SPA does not require data preprocessing, it is suitable for online monitoring. For veracity, because missing data, outliers, or data uncertainty has much less impact on various statistics than variable themselves, SPA offers advantages in this aspect. SPA does have its limitations. For example, how to optimally select the statistics to effectively monitor a process is still an open question. ACKNOWLEDGEMENT Financial supports from National Science Foundation, NSFCBET #1547124 (He), and NSF-CBET #1547163 (Wang and Shah) are greatly appreciated. REFERENCES Davis, J., Edgar, T., Graybill, R., Korambath, P., Schott, B., Swink, D., Wang, J., Wetzel, J., 2015. Smart Manufacturing. Annu. Rev. Chem. Biomol. Eng. 6, 141–160. Davis, J., Edgar, T., Porter, J., Bernaden, J., Sarli, M., 2012. Smart manufacturing, manufacturing intelligence and demand-dynamic performance. Comput. & Chem. Eng. 47, 145–156. Dong, Y., Qin, S.J., 2015. Dynamic-Inner Partial Least Squares for Dynamic Data Modeling. IFACPapersOnLine 48, 117–122. Galicia, H.J., He, Q.P., Wang, J., 2012. Comparison of the performance of a reduced-order dynamic PLS soft sensor with different updating schemes for digester control. Control. Eng. Pract. 20, 747–760. He, Q.P., Wang, J., 2011. Statistics Pattern Analysis - A New Process Monitoring Framework and Its

14951

Application to Semiconductor Batch Processes. AIChE J. 57, 107–121. Kano, M., Tanaka, S., Hasebe, S., Hashimoto, I., Ohno, H., 2003. Monitoring independent components for fault detection. AIChE J. 49, 969–976. Ku, W., Storer, R.H., Georgakis, C., 1995. Disturbance detection and isolation by dynamic principal component analysis. Chemom. Intell. Lab. Syst. 30, 179–196. Lee, J.-M., Yoo, C., Choi, S.W., Vanrolleghem, P.A., Lee, I.-B., 2004a. Nonlinear process monitoring using kernel principal component analysis. Chem. Eng. Sci. 59, 223–234. Lee, J.-M., Yoo, C., Lee, I.-B., 2004b. Statistical process monitoring with independent component analysis. J. Process. Control. 14, 467–485. Lee, J.-M., Yoo, C., Lee, I.-B., 2004c. Statistical monitoring of dynamic processes based on dynamic independent component analysis. Chem. Eng. Sci. 59, 2995–3006. MacGregor, J.F., Kourti, T., 1995. Statistical process control of multivariate processes. Control. Eng. Pract. 3, 403–414. Manyika, J., Chui, M., Brown, B., Bughin, J., Dobbs, R., Roxburgh, C., Byers, A.H., Institute, M.G., 2011. Big data: The next frontier for innovation, competition, and productivity. Oakland, J.S., 2007. Statistical process control. Routledge. O’Donovan, P., Leahy, K., Bruton, K., O’Sullivan, D.T., 2015. Big data in manufacturing: a systematic mapping study. J. Big Data 2, 1–22. Qin, S.J., 2003. Statistical process monitoring: basics and beyond. J. Chemom. 17, 480–502. Qin, S.J., 2012. Survey on data-driven industrial process monitoring and diagnosis. Annu. Rev. Control. 36, 220–234. Qin, S.J., 2014. Process data analytics in the era of big data. AIChE J. 60, 3092–3100. Shah, D., Hancock, A., Skjellum, A., Wang, J., He, Q.P., 2017. Challenges and opportunities for IoT-enabled cybermanufacturing: what we learned from an iotenabled manufacturing technology testbed, in: Proceedings of Foundations of Computer Aided Process Operations / Chemical Process Control. Tarkoma, S., Ailisto, H., 2013. The internet of things program: the finnish perspective. IEEE Commun. Mag. 51, 10–11. Wang, J., He, Q.P., 2010. Multivariate process monitoring based on statistics pattern analysis. Ind. Eng. Chem. Res. 49, 7858–7869. Xu, L., He, W., Li, S., 2014. Internet of things in industries: A survey. IEEE Trans. Ind. Informatics 10, 2233–2243. Zikopoulos, P., Eaton, C., 2011. Understanding big data: Analytics for enterprise class hadoop and streaming data. McGraw-Hill Osborne Media. Zikopoulos, P., Parasuraman, K., Deutsch, T., Giles, J., Corrigan, D., 2012. Harness the Power of Big Data The IBM Big Data Platform. McGraw Hill Professional.

15516