Production Engineering https://doi.org/10.1007/s11740-018-0797-0

QUALITY ASSURANCE

A novel approach for data‑driven process and condition monitoring systems on the example of mill‑turn centers Dominik Kißkalt1   · Hans Fleischmann1 · Sven Kreitlein1 · Manuel Knott1 · Jörg Franke1 Received: 17 July 2017 / Accepted: 18 January 2018 © German Academic Society for Production Engineering (WGP) 2018

Abstract Implementing condition monitoring functionality in production machinery often proves to be a difficult task. Device- and process-specific algorithms must be created while inhomogeneous industrial communication networks hinder the aggregation of control signals and process variables. Further challenges arise from the advance of flexible cyber-physical systems (CPS) and the industrial internet of things (IIoT). They demand a service-oriented condition monitoring architecture, which seamlessly adapts to quickly changing production topologies. In this context, data-driven systems which are capable of unsupervised learning are promising approaches. The aim is the autonomous identification of significant process variables and patterns. This paper describes a machine learning approach for a condition and process monitoring system on the basis of pattern recognition within structure-borne noise of rotating cutting machinery. Process states are defined under application of non-negative matrix factorization (NMF). A production model is learned and deployed on the basis of Gaussian mixture models (GMM) and hidden Markov models (HMM) in a two stage process. Additionally a generic framework to ease the implementation of decentralized condition monitoring functionalities is given. A decentralized component, the monitoring module, constitutes a part of a holistic condition monitoring architecture managed by a central server. The approach is evaluated on the example of mill-turn centers. Keywords  Condition monitoring systems · Smart factory · Cyber-physical systems · Machine learning · Unsupervised learning · Machine tools

1 Introduction Condition monitoring (CM) is integral to the efficient operation of production systems. By recognizing the need for maintenance activities, production downtimes are minimized [1]. The purpose of corresponding condition monitoring systems (CMS) is to gather relevant data and offer insight on the current state of critical machine components and the likely progression of their condition. Furthermore, advances in data analysis and machine learning offer high potential for maintenance optimization, but they need to be tightly integrated in CM processes [2]. Contrary to CM, process monitoring systems are not observing the condition of specific components but the * Dominik Kißkalt [email protected] 1



Institute for Factory Automation and Production Systems, Friedrich-Alexander-Universität Erlangen-Nuremberg (FAU), Egerlandstraße 7‑9, 91058 Erlangen, Germany

manufacturing process itself by detecting anomalies within a manufacturing cycle. This concept allows conclusions to the product quality and has the potential to reduce time- and cost-intensive end-of-line testing. Heterogeneous production environments and processes often render the implementation of respective algorithms difficult. Algorithms must be ported to programming languages, which the used monitoring and automation devices support. Therefore the effort required to setup a CMS especially becomes challenging for highly flexible production systems. The decentralized organization of production entities and their digital representations within cyber-physical systems (CPS) enable quickly changing production topologies [3]. Thus, CMS must be able to handle varying system structures to process higher level system interdependencies. Frequently changing production plans cause the need of regular adaption of algorithms and models. This trend is likely to rise due to the increasing individualization of production systems. The conventional design method for CPS is based on the comprising task of manual modeling expert

13

Vol.:(0123456789)



According to Isermann [8] the detection of anomalous behavior in industrial systems can be either based on knowledge, signals or models. Knowledge-based approaches depend on measured variables by instruments, whereas states are defined and assigned by the accumulated experience of human operators. In contrast, signal-based anomaly detection implies limit and trend checking of single measurement variables. If complex systems with several interdependent variables are to be monitored, mathematical models need to be created in order to illustrate the associated complex real-world problem as entire as possible. Normally such models are designed manually, which implies several disadvantages: Firstly, knowledge about the system’s internal variables is required from start. Specific domain knowledge is usually difficult to obtain, as for example an experimental setup causes plant downtimes and the deployment of human and financial resources. Furthermore models are applicable only for the specific use case and therefore require updates as soon as the production environment changes. [9–12]. As an alternative data-driven approaches which are based on machine learning have been used in several industrial applications (e.g. [9, 10, 13, 14]). The application is at two phase process: In the training phase the

13

Process Models

Cycled Production Data Cycled Production Data

State Detection

...

and Process Model Training

...

2 Model‑based condition monitoring

algorithm learns about the system’s signal features, its states, and their relationships. Whereas in the monitoring phase the acquired informations incorporated in behavior models are applied to detect anomalies within the recorded data streams. The proposed CMS does not require preloaded expert knowledge and is able to quickly adapt to changes within the production environment and products. Therefore it offers high flexibility and low need for service at the same time. Figure 1 shows the sequence of the model learning process. After capturing a sufficient amount of data, process models are built and different states are detected by the system. In the next step, related states are accumulated in order to reduce complexity. By defining possible transitions between the remaining states the eventual production model is created. The behavior of automated production systems and CPS can be described with hybrid systems, which comprise the modeling of discrete and continuous behavior [15]. In this paper the input scope is limited to continuous process variables under waiver of discrete control signals. Therefore it is possible to dispense with machine control intervention, which is frequently an obstacle because of its complexity and inaccessibility. Furthermore the flexibility of the CMS increases as it is system-independent and can be applied on other cases more easily by taking advantage of machine learning techniques. The fast adaption of production systems to new requirements like changing product variants (“plugand-produce”) is a major topic in the research area of cyberphysical production systems [16].

...

knowledge and lacks the required flexibility. Hence there is a rising trend of machine learning approaches which are capable of automated model identification [4]. The CMS described in this paper is applied on a use case of supervising mill-turn centers in a large scale production scenario. Worn ball screw drives (BSD) in linear axes represent a frequent reason for machine failures which cause large downtimes under application of a reactive maintenance strategy. Previous predictive maintenance approaches were based on the evaluation of internal control data, such as the BSD’s backlash trend or the motor torque curve. These were not able to fulfill the the purpose of dependent BSD failure prediction because of the huge variety of interference effects. The condition monitoring approach described in this paper is based on the recording of BSD vibrations similar to the research of Lee et al. [5] and Schopp [6]. The paper is organized as follows: In Sect. 2 the principles of model-based anomaly detection are given. Section 3 describes a concept of automated model identification. In Sect. 4 the CMS is classified according to the Reference Architecture Model for Industry 4.0 (RAMI 4.0) and the software architecture of its components is described [7]. Section 5 validates the presented concept on a use case. Finally, in Sect. 6 the results are concluded and an outlook to further work is given.

Production Engineering

... State Reduction and Product Model Training

Production Data Recording

Process Variables Recording and Provisioning

Evaluation −

Mill-Turn Center

... 1 n

Product Model

Fig. 1  Methodological representation of the model learning and execution phase of the monitoring system

Production Engineering

Due to the waiving of discrete machine control data, the CMS needs to be capable of identifying different states of a manufacturing cycle by recognizing reoccurring patterns within the recorded data. A model identification concept for the present case which origins from the field of speech recognition, is described in the next section.

3 Production model identification and deployment To ease the understanding of the implemented machine learning approach, first the used mathematical principles are explained. Afterwards the utilized methodology will be stated.

3.1 Principles Accurate and precise CM of hybrid systems presupposes knowledge about the current active state. Without extracting this information from the discrete control signals, assumptions based on the recognition of previously learned signal patterns can fulfill this requirement. Therefore, different models for feature extraction and state estimation were established in this work. 3.1.1 Non‑negative matrix factorization Non-negative matrix factorization (NMF) is a statistical procedure that finds usage in the field of pattern recognition. It is used for data aggregation and features detection. Using NMF, a given non-negative matrix X is decomposed into two smaller matrices W and H (Eq. 1) under application of a gradient descent algorithm [17]. W represents a set of features and H stands for the corresponding activation coefficients. Target of the modeling process is to minimize the approximation error of Eq. (1) by optimizing the squared Frobenius norm (Eq. 2) [17, 18].

X ≈ WH

(W, H ≥ 0)

1 1∑ arg minW,H ||X − WH||2 = (X − WHij )2 2 2 i,j ij

(1)

The model is defined by its mean, weight and the covariance matrices. It can be written as a linear superposition of Gaussians (Eq. 3) [19].

p(𝐱) =

K ∑

Its parameters are determined by the implementation of the expectation maximization algorithm. Extending this concept into the multidimensional space, distributions are represented by a combination of multiple single dimensional GMMs. In the context of this work GMMs are used to model the probability distributions of the NMF activation coefficients within the processing states. 3.1.3 Hidden Markov models The basic theory of Markov chains describes systems by a set of states and a matrix defining transition probabilities in between these states. Thereby the likelihood of transitioning to a specific state is only dependent on the current state, whereas the traveled path stays insignificant. In a HMM not the states themselves but only the system’s emissions are observable. The probability of a specific emission to occur depends solely on the current state the system is in [20]. The concept of HMMs can be extended to display a more comprehensive reflection of the production system. The hierarchical hidden Markov model (HHMM) illustrates several state levels, whereas each level is considered to be a self-contained HMM in which only one state is active at the same time [21]. Figure 2 shows an exemplary HHMM for the conduct of a production plant. Plant Level

On

Product Level

3.1.2 Gaussian mixture models Arbitrary distributions can be represented by combinations of Gaussian distributions called Gaussian mixture models.

P1

Off P2

Pn

Process Level

(2)

The feature identification and recognition capabilities of NMF help the system to describe the shape of the single manufacturing steps and to determine the current position in the production sequence.

(3)

𝜋k  (𝐱|𝜇k , 𝚺k )

k=1

PS1

AP1

E1

PS1

AP1

TC

TC

AP2

PS2

AP2

En

PS2

PSn

PSn

Legend

Inac ve State

Ac ve State

Transi on P Product E Error

PS Processing Step AP Axis Posi oning TC Tool Change

Fig. 2  Hierarchical hidden Markov model for plant behavior modeling

13

Production Engineering

power spectral density [ B



Fig. 3  Signature of the recorded vibration signal, Top: time domain, Bottom: power spectral density

In a condition monitoring context HMM are suitable for describing hybrid production plants with an unknown control signal vector [22]. Since the control signal is unknown, the plant’s status or rather the current manufacturing stage is hidden. By observing specific machine emissions like structure-borne noise the system’s current active state can be estimated through application of the HMM.

3.2 Methodology The data scope for the system is entirely limited to the structure-borne sound of the BSD. Fig. 3 shows the recorded vibration of a single production cycle and the corresponding power spectral density (PSD) over time.1 In the following chapters, aside from methodical representations, the signal will be entirely handled in the frequency domain. During production the workpieces are processed by the mill-turn centers in multiple sequentially arranged manufacturing operations which are separated by repeating tool exchange actions. This behavior is also observable in the frequency domain representation of the vibration signal (see Fig. 3 bottom). The dark marks below 50 Hz represent BSD acceleration on the monitored axis, whereas the vertical lines indicate BSD movement of adjoining axes transmitting mainly noise over the connecting structures. The reoccurring pattern of narrowly and uniformly spaced axis movements represent the tool exchange operations. In between tool exchanges the processing steps with their characteristic signal signatures are localized. Due to the fact that the single process states and boundaries can be determined by eye, we establish and later prove

1   In compliance with trade secret protection, information about the timing and the complete production sequence can not be displayed.

13

the claim that this can also be obtained by applying machine learning approaches. Figure 4 shows the applied methodology for state detection. The recorded vibration signal is transferred into the frequency domain via Short-Time Fourier Transformation (STFT).2 Under application of NMF the signal is decomposed into its components representing the system’s emissions or retrospective the observations. Next, the state sequence with the highest likelihood is estimated by parsing the observations to the Viterbi-algorithm. Model-based condition monitoring of production processes presupposes models which can depict the procedural sequence and shape of the process variables in a sufficiently accurate form. Conditioned by the omission of control signals the states stay hidden. For fulfilling the above stated requirement a target set of states has to be defined for estimating the HMM parameters. Therefore the following logical subdivision of mill-turn centers processing is utilized as a definition. Single states represent logically indivisible machine operations like axis-movement, milling of a plane or simply machine stasis during tool exchanges. Goal of the learning phase is to find the characteristic features for the described atomic machine operations and splitting the time series into a set of target states. 3.2.1 Learning phase After the implementation of the CMS into a production system the learning phase is conducted. Thereby data is recorded and processed synchronously to production. The recorded sets have to be error-free and represent the nominal data. The recording time span has to include multiple

2

 Note that the window’s width has to be wider than the shortest detectable signal pattern.

frequency

Power Spectral Density

Production Engineering

time Xt+1

Xt+n

Features W

Decoding Ht

Ht+1

Ht+n

Wn

W1

H1 Hn

activation

Coefficients [H]

Coefficient Calculation

Xt

Hidden Markov Model

time

State 1

State 2

State 3

State n

Fig. 4  Methodology of state detection via NMF-decoding and GHMM

production cycles in order to model the probabilistic behaviors of the production unit in sufficiently accurate manner.3 After splitting the training set into multiple production cycle sets, the model learning phase is triggered. For each production cycle a low dimensional NMF is conducted. Subsequently, the activation coefficients are assigned into two clusters by applying spectral clustering. Based on the characteristic behavior of NC machines the time series is split into single machine operations labeled with each standstill or activation. The transition time between these two super states defines the single state boundaries. After the identification of single machine operations, another higher dimensional NMF is conducted on the production cycle sets with sparsity restrictions. The sparsity restrictions ensure the necessary distinctness between the learned features in order to ensure that the recorded signal can not be reconstructed in arbitrary combinations of learned features. Thus, robust recognition of the active process state is ensured. By splitting the activation coefficients according to the prior detected state transitions multiple sets of H representing the single machine operations are formed and consistently stored in the the form of GMMs. In the next step, reduction of the possible state space and thus limitation of necessary computational capacities is accomplished by merging the prior defined states. This comes with the trade-off in reducing the models’ accuracy. In order to maintain a meaningful accuracy the merge order is conducted according to the Jensen–Shannon divergence between all states. Outgoing from the pair of states with the lowest distance, the states are iteratively merged until

3   Especially hydraulic driven machine operations like the operating of the tool holder reveal often a non deterministic behavior.

the divergence of all remaining states among themselves exceeds the given threshold. Updating the divergence information of the merged states to the remaining states is done by individually selecting the greater option of the both states to merge. Mapping of the initial to the merged states is subsequently used to define the transition probability matrix of the corresponding HMM. 3.2.2 Monitoring phase After training the process and product models the monitoring phase can be triggered. The continuous monitoring process consists of three successive processing steps: 1. Applying STFT and multiple filtering steps on the incoming vibration signal. The signal in the frequency domain is sliced into short overlapping intervals. By defining the window length the trade-off between sensing short timed events and necessary timing sensitivity has to be addressed. 2. The PSD intervals are decoded into the current activation coefficients by using the learned features for the active manufacturing mode of the production unit. 3. The resulting activation coefficients are compared against the active state of the product model. If the divergence exceeds a given level, the coefficients are compared to states which can be transitioned from the active state according to the product model. If no matching state can be found, an anomaly will be assumed and reported.

4 Software architecture After defining the concepts for online data analysis, the design of a corresponding hard- and software architecture can be conducted. In an industrial internet of things (IIoT) service-oriented information processing becomes particularly significant. The execution of complex calculations is often provided as a service by server platforms, as decentralized embedded systems are not economical or lack the required processing power [7]. Central server platforms allow for the efficient utilization of hardware resources and avoid dimensioning decentralized hardware for occasional peak loads. Serviceoriented architectures (SOA) help to manage system complexity by extensive modularization. In previous work, an architecture for CMS is introduced, which is designed to meet the demands of sophisticated CPS. It comprises four modules [23]. To clarify their respective scopes, the modules are classified according to RAMI 4.0 (see Fig. 5) [24].

13



Production Engineering

Condition and Process Monitoring Machine Learning

Layers Business Functional Information

OPC UA Administration Shell

Communication Integration Asset

Maintenance

Fig. 5  Alignment of the proposed CMS in the RAMI 4.0 reference model

In this case, the responsibility of a monitoring module (MM) is to collect the necessary data, process the retrieved information and offer its result in a semantically accessible way. Multiple MM can form a distributed CMS. The distributed CMS is managed by a central Cloud Module. It organizes the system structure and configures the MM. Besides storing long-term data, it also executes resource-intensive calculations. The resulting models are transferred back to the MM. The MM serves the purpose of collecting data, executing the CM algorithms and providing the results. Its objective is to offer performant data processing, while being easily configurable. Maintenance personnel is integrated into the CM process via HMI Modules. The HMI Module visualizes data and manages knowledge on the basis of web browser technologies. Its primary connection point is the Cloud Module, but it can also connect to single MM in order to retrieve online data. An important feature of the HMI Module is the possibility to train the algorithms responsible for the CM process by giving feedback about the generated conclusions. The conceptual software architecture and the schematic deployment of the learned models is displayed in Fig. 6. By using ISA95-compliant information models a consistent terminology is provided. This contains the information representation from the shop floor where the vibration data is gathered way up to the downstream to the anomaly detecting and model learning system modules. Thereby an efficient integration of actors in the total system can be ensured. The implementation of the information models is compatible on RAMI 4.0 and uses the Open Platform Communications Unified Architecture (OPC UA) communication standard. By using its rich service oriented architecture the complex data can be modeled and provided while fulfilling

13

the requirements in terms of scalability and multi-platform support. The designed information model which represents the mill-turn centers is demonstrated in Fig. 7.

5 Evaluation For the evaluation scenario a critical interval of the production cycle was selected. Due to comprehensibility reasons and intellectual property concerns some processing steps are displayed in a condensed form. Figure 8 illustrates the feature and anomaly detection abilities of the CMS applied to a 12-s data production sample, containing the milling of two different hexagons separated by a tool change operation. Thereby the NMF was conducted with each three components for the features and the activation coefficients with no sparsity restrictions. The results show a clear recognition and differentiation between three production operations: ­C1 describes the activity of the moving BSD, ­C2 detects the milling operations of the first hexagon and ­C3 those of the second one. By applying slice-wise spectral clustering of the NMF coefficients, state transitions can be detected which is displayed by changing background color. For evaluating the anomaly detection capability the prevalent error pattern of machine chattering was considered. In our scenario the disruptive vibrations between cutting tool occur at approximately 700 Hz. The subplot in row three shows the calculated residuum between the measured and the reconstructed signal of the learned NMF model. If the residuum exceeds a specified confidence interval, the occurrence of a production fault is assumed.4 One can see the transgressions of the residuum over the threshold within the processing of the second hexagon where the chattering error occurred. If an anomaly was detected, the erroneous section is parsed to another NMF model for decoding into its activations coefficients. This particular model contains all features learned on the basis of previous detected and classified anomalies. In the course of this, a similar procedure to the previous one is applied. If none of the features can be recognized, the occurrence of a presently unknown anomaly is assumed. Hereafter the data containing the anomaly is sent to the Cloud Module triggering the updating of the known anomaly model and the following redistribution to all Monitoring Modules. By forwarding the relevant information, the user can label unknown anomalies (e.g. tool breakage, collision with workpiece, etc.). Based on the mapping between fault labels and

4

  In this case the confidence interval is four times the standard deviation of the nominal detected residuum.

Production Engineering R

OPC UA Client

OPC UA Server

Model Training

Model Management

Anomaly Mapping

OPC UA Client

R Monitoring ModulesOPC UA Server

Human Agent

Storage

read / write access

request / respone communication channel bidiractional Communication channel

Monitoring ModulesOPC UA Server Internal data bus

Anomaly Detection

Anomaly Models

State Estimation

Process Models 1

Signal Processing

OPC UA Client

Anomaly Detection

Anomaly Models

State Estimation

Process Models n

Monitoring Module n

Controller

Monitoring Engine

Monitoring Engine

R

Agent

R

Controller

Monitoring Module 1

Legend QMS | HMI Module

OPC UA Client

Cloud Module

MMS | HMI Module

R

Production Line 1 Local model deployment

Signal Processing

OPC UA Client

R

R

R

Production Line n

R

OPC UA Server

OPC UA Server

OPC UA Server

OPC UA Server

Provisioning Module

Provisioning Module

Provisioning Module

Provisioning Module

Mill-Turn Center 1

Mill-Turn Center n

Mill-Turn Center 1

Mill-Turn Center n

Production Line 1

Cloud Module

Production Line n

Fig. 6  Left: Software architecture of the condition and process monitoring system divided into four modules according to the Fundamental Modeling Concepts standard; Right: schematic presentation of production specific model deployment

ISA-95 Base Information Model

ISA95ClassType EquipmentClassType

PhysicalAssetClassType

ISA95ObjectType

PhysicalAssetType

EquipmentType

ISA-95 Common Object Model MachineClassType

MillingMachine ClassType HasISA95 A ribute

Manufacturer:: CompanyType

ModelNumber:: BaseDataType HasISA95 ClassProp erty

BallScrewDrive ClassType

MeasurementDevice ClassType VibrationSensorClass Type HasISA95 A ribute

VibrationX::

Manufacturer:: CompanyType

VibrationY::

ModelNumber:: BaseDataType

VibrationZ:: PhysicalAssetClassPropertyType

Third Party Equipment and Physical Asset Model

HasISA95 Property

Channel I

Channel II

FolderType

FolderType

X-Axis

XX-Axis

HasISA95 Property

HasISA95 Property

VibrationSensor

VibrationSensor

Y-Axis

YY-Axis

HasISA95 Property

HasISA95 Property

VibrationSensor

VibrationSensor

Z-Axis

ZZ-Axis

VibrationSensor

VibrationSensor

HasISA95 Property

Mill-Turn Center

SensorType DefinedBy EquipmentClass

PhysicalAssetClassPropertyType

PhysicalAssetClassPropertyType

LinearFeedAxisType

HasISA95 Property

HasISA95 Property

HasISA95 A ribute

HasISA95 Property

Manufacturer::

Type = CompanyType Value = ...

ModelNumber::

Type = BaseDataType Value = ...

(Attributes and properties of sensor instances are not shown for reasons of clarity)

Fig. 7  Information model of the mill-turn centers compliant to the ISA95 standard

13



Production Engineering

Fig. 8  Top: PSD of the structure-borne noise with erroneous machine behavior at 700 Hz from 6.5 to 10.5 s , Middle: Normalized NMF coefficients and detected states, Bottom: Residuum between recorded and reconstructed NMF model and detected states

the respective anomalies, already occurred anomalies can be recognized and classified into the corresponding faults. Thus, explicit knowledge about existing faults can be made persistent and usable synchronously to the manufacturing process. Thereby quality management systems or maintenance personnel directly can be informed about the occurrence of relevant faults.

6 Conclusion and future work In this paper a process and condition monitoring system based on pattern recognition and machine learning was developed and evaluated on an industrial use case. The implemented CMS is capable to learn models representing signal characteristics and manufacturing sequences without human input and synchronous to the manufacturing process in order to detect process anomalies in a self-sufficient matter. Thereby the preconditions are held minimal to ensure a maximum adaptability on altering production plans. Because of its adaptive data processing architecture, the developed overall concept is well suited for the realization of highly flexible, distributed CMS with machine learning capabilities. Through its generic design and the ability to handle arbitrarily structured data objects it provides a

13

solution neutral framework for the implementation of condition monitoring functionality. Service-orientation and modularization help to manage the complexity of CPS. In the future a more comprehensive system evaluation regarding system robustness and prediction accuracy is planned. Especially detection rates and necessary accuracy regarding different production faults like machine collisions and tool breakage have to be defined. Further the integration of the CMS into different network topologies have to be evaluated in an industrial environment. The automated work-flow and feedback in order to map detected anomalies to the correspondent failures has to be implemented and evaluated. Acknowledgements  We would like to thank the Robert Bosch GmbH for providing access to its production facilities.

References 1. Marwala T (2012) Condition monitoring using computational intelligence methods: applications in mechanical and electrical systems. Springer, New York 2. Schuh G, Stich V, Reuter C, Blum M, Brambring F, Hempel T, Reschke J, Schiemann D (2017) Cyber physical production control. In: Jeschke S, Brecher C, Song H, Rawat DB (eds) Industrial internet of things. Springer, Cham, pp 519–539

Production Engineering 3. Oks SJ, Fritzsche A, Möslein KM (2017) An application map for industrial cyber-physical systems. In: Jeschke S, Brecher C, Song H, Rawat DB (eds) Industrial internet of things. Springer, Cham, pp 21–46 4. Niggemann O, Lohweg V (2015) On the diagnosis of cyber-physical production systems—state-of-the-art and research agenda. In: Twenty-ninth conference on artificial intelligence (AAAI-15) pp 4119–4126 5. Lee WG, Lee JW, Hong MS, Nam SH, Jeon YH, Lee MG (2015) Failure diagnosis system for a ball-screw by using vibration signals. Shock Vib. https:​ //doi.org/10.1155/2015/435870​ (Article ID 435870) 6. Schopp M (2009) Sensorbasierte Zustandsdiagnose und -prognose von Kugelgewindetrieben. Dissertation, Karlsruher Institut für Technologie, Institut für Produktionstechnik wbk 7. Fleischmann H, Kohl J, Franke J, Reidt A, Duchon M, Krcmar H (2016) Improving maintenance processes with distributed monitoring systems. In: IEEE 14th international conference on industrial informatics (INDIN), pp 377–382. https​://doi.org/10.1109/ INDIN​.2016.78191​89 8. Isermann R (2011) Fault-diagnosis applications—model-based condition monitoring: actuators, drives, machinery, plants, sensors, and fault-tolerant systems. Springer, Berlin 9. Vodenčarević A, Büning HK, Niggemann O, Maier A (2011) Identifying behavior models for process plants. In: IEEE 16th conference on emerging technologies & factory automation (ETFA), pp 1–8. https​://doi.org/10.1109/ETFA.2011.60590​80 10. Vodenčarević A, Büning HK, Niggemann O, Maier A (2011) Using behavior models for anomaly detection in hybrid systems. XXIII International Symposium on Information, Communication and Automation Technologies 1–8. https​://doi.org/10.1109/ ICAT.2011.61020​93 11. Henning S, Niggemann O, Otto J, Schriegel S (2014) A descriptive engineering approach for cyber-physical systems. In: Proceedings of the 2014 IEEE emerging technology and factory automation (ETFA) pp 1–4. https:​ //doi.org/10.1109/ETFA.2014.700528​ 6 12. Fleischmann H, Spreng S, Kohl J, Kißkalt D, Franke J (2016) Distributed condition monitoring systems in electric drives manufacturing. In: 6th international electric drives production conference (EDPC), pp 52–57. https​://doi.org/10.1109/EDPC.2016.78513​14 13. Maier A, Niggemann O, Koester M, Gatica CP (2013) Automated generation of timing models in distributed production plants. In:

14.

15. 16.

17. 18. 19. 20. 21. 22.

23.

24.

Proceedings international conference on industrial technology (ICIT), pp 1086–1091. https​://doi.org/10.1109/ICIT.2013.65058​ 23 Windmann S, Niggemann O (2015) Efficient fault detection for industrial automation processes with observable process variables. In: Proceedings international conference on industrial informatics (INDIN), pp 121–126. https:​ //doi.org/10.1109/INDIN.​ 2015.72817​ 21 Branicky MS (2005) Introduction to hybrid systems. In: HristuVarsakelis D, Levine WS (eds) Handbook of networked and embedded control systems. Birkhäuser, Boston, pp 91–116 Dionisio Rocha A, Peres R, Barata J (2015) An agent based monitoring architecture for plug and produce based manufacturing systems. In: Proceedings 13th IEEE international conference on industrial informatics (INDIN), pp 1318–1323. https​://doi. org/10.1109/INDIN​.2015.72819​26 Hoyer PO (2004) Non-negative matrix factorization with sparseness constraints. J Mach Learn Res 5:1457–1469 Berry MW, Browne M, Langville AN, Pauca VP, Plemmons RJ (2007) Algorithms and applications for approximate nonnegative matrix factorization. Comput Stat Data Anal 52(1):155–173 Bishop CM (2006) Pattern recognition and machine learning. Springer, New York Rabiner L, Luang B (1986) An introduction to hidden Markov models. IEEE ASSP Mag 3(1):4–16. https​://doi.org/10.1109/ MASSP​.1986.11653​42 Fine S, Singer Y, Tishby N (1998) The hierarchical hidden Markov model: analysis and applications. Mach Learn 32:41–62. https​:// doi.org/10.1023/A:10074​69218​079 Windmann S, Jungbluth F, Niggemann O (2015) A HMM-based fault detection method for piecewise stationary industrial processes. In: Emerging technologies & factory automation (ETFA), pp 1–6. https​://doi.org/10.1109/ETFA.2015.73014​65 Fleischmann H, Kohl J, Franke J (2016) A reference architecture for the development of socio-cyber-physical condition monitoring systems. In: 11th system of systems engineering conference (SoSE), pp 1–6. https​://doi.org/10.1109/SYSOS​E.2016.75429​63 VDI/VDE Society Measurement and Automatic Control (2015) Status report: reference architecture model industrie 4.0 (RAMI 4.0)

13