Smart Manufacturing infrastructures have to be based on network technolo-

gies which enable a secure (encryption, authentication, robustness, safety), vertical and horizontal cross-domain and cross-layer communication between stationary and mobile objects (as virtual objects, sensors, actors, devices, things or systems). Network technologies have to comply with specific requirements related to e.g. real-time, safety, security, data amounts, wired or wireless, passive or active, etc. [23] . Lower field-levels require time frame abilities of seconds or milliseconds for response, reliability, resolution, and repair (e.g. control, or real- time statistics of the process), whereas higher levels only require time frames of weeks or months (e.g. for production planning or accounting) [23] .

Potential suitable technologies are e.g. ZigBee, (Industrial) Ethernet, 6LowPan, ZigBee, Wi-Fi or Bluetooth.

Communication protocols in smart manufacturing should enable a robust, loosely coupled, time-synchronized, secured, and semantically based communication.

It is hardly possible in smart manufacturing infrastructures, where thousands of communicating objects are reachable from everywhere, to exchange information traditionally loop based (e.g. on field level (ISA-95 levels 1 - 2) via a bus in a defined frequency to enable a deterministic information exchange). That generates a need for new approaches to reduce the traffic in networks whilst holding on to all requirements. Such a new approach is e.g. the paradigm “service-ori- ented architectures (SOA)”. SOA describes an architectural pattern in which functionalities and features of systems, components, applications, etc., are provided as a service in a network where all services can be found and accessed using publish and subscribe mechanisms in form of complex event processing as an alternative for control loops and to reduce traffic in the network. Services are “logical representations of a repeatable business activity that has a specified outcome (e.g. check customer credit, provide weather data, consolidate drilling reports), is self-contained, may be composed of other services and is a ‘black box’ to consumers of the service [24] ”.

Potential suitable technologies are e.g. implemented DPWS specifications or OPC UA.

Architectures, in general, are describing combinations of components/modules and their interaction, and should provide a unified structure and wording for used terms. An architecture should include a logical, a development, a process and a validation view, and should provide scenarios for a validation as proposed by Philippe Kruchten in his 4+1 architectural view model [25] . A smart manufacturing architecture should also provide a unified structure and wording covering mandatory aspects in smart manufacturing as product, system or order life cycles, value streams, information flows, or hierarchical layers. Such architectures are currently under development.

Example projects and architectures that contribute to this area are e.g. RAMI 4.0 [26] , Arrowhead [27] , CyPros [28] or Fi-Ware [29] .

(Physical) reachable objects inside a smart manufacturing network (e.g. digitalised and virtualised field level devices, systems, material, integrated humans, virtual concepts (e.g. of products in the design phase), etc.), have to fulfil a range of requirements. Objects should communicate using a unified communication protocol, at least at the application level, and should be based on a unified semantic to enable a mutual identifiability and understanding. The object itself should provide its own features as a service (e.g. state information or functionalities), and should be able to provide its own description next to extended information as manuals, specifications or wear information. All these have to be kept next to further requirements related to security, safety or quality of service [26] [30] .

Finally, various applications that use services of deployed objects to realise e.g. control systems, systems of systems through service orchestration, or―as focused in this work―Big Data analysis applications can be implemented.

The definition of Big Data is still under discussion and many suggestions were made. Gartner made a proposal in 2011 where he suggested to categorise Big Data through 3Vs (Volume of Data, Variety of Data and Velocity of Data) [31] . This categorization (/definition) is widely accepted. Newer definitions have also added a fourth “V” standing for “veracity” which describes the quality of captured data that can vary and affect the accurate analysis [32] .

Big Data in Smart Manufacturing systems are big amounts of (continuously generated) data produced by machines, ambient sensors (temperature, vibration, humidity, etc.), controllers, (manufacturing) systems, etc. [33] available in a great variety as e.g. in form of signal/information streams, log files, master data, manual entered operator data, etc.

It is also conceivable for Big Data analysis, to include further data sources from the enterprise level systems, supply chains, marketing and sales, Product Lifecycle Management (PLM) systems, social media, website browsing pattern or from business forecasts [34] .

As described in Section 3.1, a smart manufacturing infrastructure should enable data access to nearly all data, available in its environment. Data should be pre-processed and aggregated, and send to receivers/subscribers (e.g. a Big Data analysis system) in an event-based way (see paragraph about “Saving Data” below) e.g. to reduce the amount of traffic in a network. A Big Data analysis system should also save these data to enable a historical analysis for detecting long-term patterns as e.g. related to wear and tear of devices.

For analysing Big Data in smart manufacturing, it is necessary to have:

・ enough data to make useful analysis.

・ complete data to avoid miss-interpretations [35] .

・ correct data which should describe real processes.

・ available meta-data and data descriptions to simplify the analysis process.

What this means in detail is often dependent on the concrete use case.

Produced Big Data in smart manufacturing systems can be analysed from (real- time) data streams (short term pattern) up to full historical data buckets. However, it is important to think about reducing the data without losing meaningful data. Thousands of sensors cannot send data each millisecond to a data analysis solution through an smart manufacturing infrastructure. Reducing traffic can be done through different combinable approaches:

・ send event-based data (send data only if needed).

・ send data aggregated (averages, sums, etc.).

・ send bigger data packages in form of log files in which (aggregated) data are stored locally.

Data sent to a data analysis system has to be saved in an appropriate database as described in section 3.4.1.

The process of making Big Data analysis is often similar to the Cross-Industry Standard Process for Data Mining (CRISP-DM) as shown in Figure 7.

In any case Different experts are needed before a Big Data analysis can start. At least two experts are needed to make the Big Data analysis itself: a data scientist who understands the analysis methodologies and who is able to use the tools; and an IT expert who is able to deploy, configure/administrate databases and computer clusters, needed for the analysis of Big Data [35] .

In addition, two other experts are commonly needed: One expert who has clear expectations of Big Data analysis outcomes, and another expert who has knowledge about the (input) data for the analysis [35] .

In manufacturing this could mean, that a production site expert, who has clear expectations of big data analysis outcomes is needed (as e.g. an optimisation of a KPI); and a manufacturing expert, who knows the production process and the available data needed for the analysis [35] .

In general it is very important, and a big challenge, that all involved experts get a

clear common understanding of aimed Big Data analysis expectations/goals [35] .

Coming back to the analysis process itself, as already introduced, a typical analysis process often follows the CRISP-DM standard [40] [41] , which is divided into six phases [42] , as indicated in Figure 7:

Each phase includes different tasks:

1^st Phase―Understanding the Business: includes the determination of business objectives, situation assessment, and the determination of Big Data analysis expectations/goals.

2^nd Phase―Data Understanding: includes the collection of initial data, exploration and description of data, and the verification of data quality.

3^rd Phase―Data Preparation: includes the selection, cleaning, construction, integration and formatting of data.

4^th Phase―Modelling: includes the selection of a modelling technique, the generation of a test design, and the building and assess of a model.

5^th Phase―Evaluation: includes the evaluation of Big Data analysis results, knowledge discovery and planning of the next steps.

6^th Phase―Deployment: includes the exploitation of lessons learned, as for example the use of new knowledge e.g. for predictions, further observations, decision support or automation [35] .

Big Data analysis is a heuristic process. Analysing Big Data compared to related topics as classical Data Mining, requires the consideration of great amounts of data that are not processable through classical data mining approaches. Big Data analysis needs therefore approaches to access and process these amounts of data very fast, and needs therefore scalable, linear and parallelizable algorithms, as well as storage approaches. Simple Big Data analysis includes some methods that fulfil these requirements for fast processing of Big Data.

Interim results coming from simple Big Data analysis are often basis for further data mining to reduce the complexity of Data, preferably without losing valuable information. Simple Big Data analysis results are coming among others from descriptive statistical methods as calculations of averages, shapes of distribution histograms (normal, exponential, Poisson), but can be also sums, minimums and maximums values.

Such simple big data analysis results can be final results, but also interim results for further processing in data mining, artificial intelligence or machine learning approaches.

Statistics, graph analysis, artificial intelligence and machine learning approaches are suitable for data mining in smart manufacturing. A small overview will be provided in the following for each sub-topic indicated in Figure 6.

Statistical methods in Big Data analysis are used e.g. for the identification of clusters in data, the identification of associations, or the analysis of regressions. Meaningful data as the identification of clusters and associations in data, potentially helps to understand how statistical obtained values correlates, and how they affect each other.

Through such information can be derived happenings, behaviours, and relationships of/between smart manufacturing system components, processes and supply chains. Such knowledge is useful to improve virtual models of them, used for simulation and prediction, as basis for several use cases as the improvement of KPIs (performance, efficiency, etc.), or to make regression analysis to enable the determination of system control for real-time adaptions based on the results.

Statistical methods are commonly used in manufacturing for scheduling and diagnosis, but are also regularly used on the Enterprise Resource Planning (ERP) level (s. ISA-95) to establish mathematical correlations between their profits, expenditures and produced goods.

Another approach to get information out of Big Data is graph analysis. In graph analysis data are represented in a graph [44] . Many things can be represented in a graph, as social media networks, power grids, knowledge (in form of ontologies), street maps, etc. In the smart manufacturing domain a graph could e.g. represent IoT interactions, energy flows, processes or supply chains [45] .

The analysis of paths, connectivity, community, or centrality of/between objects inside a graph, are parts of graph analysis [46] . Path analysis is used to proof reachability between two objects in a graph, to find a path with no repeating edge, to find cycles, or to find a shortest path between nodes. Connectivity analysis is used to provide answers of how strong an object is connected with other objects, how robust a graph is, or could provide answers of the similarity of graphs. Community analysis is used to identify objects that are often interacting (communities/clusters) and its evolution. Centrality analysis is used to get information about the importance of nodes (e.g. in case that a node is connected to many other nodes, which could be an indication of a high importance, or a single point of failure). Based on the specific use case in manufacturing (as already said, e.g. representation of IoT interactions, energy flows, processes or supply chains), such analyses can provide meaningful information about the analysed system.

In addition, the Dijkstra algorithm, a popular example in graph analysis, is very useful for finding shortest paths in smart manufacturing systems. A graph could represent a flexible transportations system, where it is necessary to find the shortest path between start and destination. To do so, iterative algorithms, such as Dijkstra’s, perform tasks on an input, over and over again, which leads to expensive and unnecessary reloading of data [47] . In cases such as these, bulk synchronous parallel (BSP) [48] processing of the data constitutes a good alternative. In systems that follow this model, computation is done through successive super steps, where computation occurs simultaneously in each of the participating nodes. Once this is done, a final and global synchronization step is needed.

A potential use case for graph analysis in smart manufacturing could be the optimization of the energy efficiency, through the representation of energy grids or processes as a graph. One approach could be the optimization of processes through the identification of the lowest/highest energy consumption over time (in nodes and edges of the graph). Through this information, and in combination with a parametrized Dijkstra algorithm, the most energy efficient route or control setting for a future similar task can be chosen. Another approach could the optimization of the control. A manufacturing process, represented as a graph, could support the identification and prediction of components, which are not used in a specific period within a manufacturing system. Such information would be used to shut down components, or to turn them into a low consumption/standby mode during this time.

However, graph analysis is a very computing intensive process and mostly limited to hardware performance capabilities, given that classical approaches store the entirety of the graph in memory. To improve this circumstance, it is currently being studied to load smaller partitions into memory and to handle communication at this level. An overall comparison on the several existing Graph models is given in [49] .

Artificial Intelligence (AI) concerns computational heuristic methods. An AI method is considered as such, if it is capable of learning something by its own means. This research field is often materialised in the form of Neural Networks (NN) and Decision Trees (DT) (two of the most intensively researched fields within AI). This field has a variety of applications in an equally wide spectrum of other research and enterprise areas. Concerning manufacturing, both NN and DT are widely used as diagnosis approaches, often as part of other self-learning paradigms. Behind classical manual diagnosis, individual diagnosis solutions are often used for manufacturing systems. Researchers are working on new approaches based on, for instance, automatic (predictive) failure detection through fault tree analysis (FTA) [50] , or self-learning approaches [51] . These approaches handle also systems with a higher complexity (evolvable and emergent systems with changing physical and logical conditions and behaviours). FTA is a methodology for fault detection based on tree composed and linked logical Boolean functions. States which could result in a defined combination to an error, have to be identified through the FTA, and will result on of the top tree level in an error state (e.g. when an unwanted combination of sub-states appears) [52] .

Self-learning approaches are constantly observing systems in order to learn the normal behaviour of a system, and to detect anomalies and failures, based on the knowledge extracted. Deterministic finite automaton (DFA) [53] is one popular research field in this direction where, algorithms observe systems to build deterministic finite automats, which represents a correct behaviour of the system. In case where the system gets into a state which is not covered by the model (generated in the learning phase), a pre-defined reaction will be triggered, e.g., sending an alarm or start a troubleshooting routine [50] . This method also allows to evaluate the progression of a failure. Another research field that has been around for several years and has grown in importance with the current distributed and learning concerns is that of bio-inspired algorithms. These algorithms, e.g. the Artificial Immune Systems ( [54] [55] [56] [57] [58] ), emulate the behaviour of several animals, and even humans, in order to provide machine learning capabilities.

These algorithms depend on their knowledge base. This means that they depend on existing knowledge (data) to obtain new information using their algorithms. Often, the result can be untrustworthy given that the knowledge base is insufficient to obtain conclusive results. For example, to understand what is the best opportunity to put a station on a minimum consumption standby, it is not enough to assess its behaviour twice, but rather for a whole month. By using Big Data (interim results) as source for their algorithms, the obtained results may prove themselves as much more trustworthy. Hence, AI can be used in Big Data analysis to perform, for instance, dynamically AI supported scheduling of the stations in a shop floor.

Machine Learning (ML) is currently described as computational methods that use experience to improve performance, or make accurate predictions [59] . Experience is often nothing more than a knowledge-base of a computational method/algorithm. This knowledge-base allows these algorithms to reason on whether the obtained results are better or worse than the ones previously known. If these results are better, the system’s mechanism can be changed, or not, according to the use cases.

Machine Learning represents a set of algorithms that are being widely used in a vast range of applications and in several use cases as energy analysis [60] [61] . Some of the most important algorithms are C4.5, k-means clustering, support vector machines, Apriori, expectation-maximization, k-nearest neighbours class, Naïve Bayes and CART [62] , used for instance to build decision trees, classification on matrices, identification of frequent item sets, finding of association rules or clustering [63] .

From a Big Data perspective, specially taking into account the smart manufacturing domain, ML is viewed as a mean to obtain better performance results, as in obtaining lower energy consumptions, or obtaining a product faster. On the other hand, ML is also very important to, firstly, understand the processes e.g. which cause more tear and wear to the machines, and secondly, to determine the best alternative to those processes. This introduces another whole importance to ML, since it enables machines in a smart manufacturing environment to diagnose themselves faster and more efficiently [64] .

With the growing amount of available data The former RDBMS centralised and concurrent, became inappropriate to handle huge chunks of Big Data with the growing amount of available data with changed structures [67] .

Hence, to cope with this change of data towards data with Big Data characteristics, a new type of database emerged which is called NoSQL. This new kind of NoSQL databases [68] are suited for distributed systems, parallel access, and fast processing of poly-structed huge amounts of data. Contrary to RDBMS, NoSQL databases are scalable and provide better performance rates on Big Data, whilst allowing the users to benefit from a database where the structure of data is irrelevant. Some widely used technologies are MongoDB [69] , Cassandra [70] and HBase [71] [72] (Figure 9).

The majority of data processing tools, frameworks or algorithms use the Google “Map Reduce” algorithm which is a programming model for parallel processing of big data sets [73] . These algorithms consist of two distinct tasks: Map and Reduce. Put simply, the Map task finds different keys (tokens) in the data, whilst the reduce task groups the different entries of data according to the existent

keys. This algorithm was implemented in the free Apache Hadoop framework, which is one of the older solutions, but up to now a base/core element of many Big Data technologies. Hadoop includes the Hadoop Distributed File System (HDFS) and the Map Reduce Algorithm. Optionally are provided several additional extensions, e.g. for the management of large data sets in a Hadoop-Cluster [74] , Data Warehouse functionalities [75] , real-time functionalities [76] , etc.

Currently several open source data processing tools that provide comprehensive sets of operations to extract knowledge from existing data, are available (see Figure 10). Table 3 presents some of the most widely used free (at least for academic use) or open-source data processing tools. A comparison of these engines can be found in [77] .

Table 4 presents some of the most widely used free (at least for academic use) or open-source Data Processing Engines.

Many of the existent frameworks for ML in Big Data are associated with data processing engines [88] , but, the majority of ML frameworks (see Figure 11) can

be used with several data processing engines, even though some are limited due to language limitations. Table 5 presents some of the most widely used free (at least for academic use) or open-source Machine Learning frameworks.

Another technology which should be also mentioned is MLbase [92] , which is one of the very first attempts to develop a framework that is, at the same time, both a Data Processing Engine and a Machine Learning Tool.