Real-Time Information Base as key enabler for Manufacturing Intelligence and “Industrie 4.0”

Full text

Real-Time Information Base as key enabler for
Manufacturing Intelligence and “Industrie 4.0”
Building the bridge between “real” Real-time Exploitation and Big Data via mathematically
grounded Information Fundamentals
Gerhard Luhn, PhD
SYSTEMA Gmbh
Dresden, Germany
gerhard.luhn@systemagmbh.de
Dirk Habich, PhD
Technical University Dresden
Dresden, Germany
dirk.habich@tu-dresden.de
Katrin Bartl
X-FAB Semiconductor Foundries
Erfurt, Germany
katrin.bartl@xfab.com
Johannes Postel, Travis Stevens, Martin Zinner, PhD
SYSTEMA Gmbh
Dresden, Germany
{johannes.postel, martin.zinner}@systemagmbh.de
travis.stevens@SYSTEMA-USAcorp.com
Abstract—This paper provides an overview and details
concerning a new and fundamentally based solution concept to
support “real” real-time Fab analysis, fast new-product ramp
and inherent knowledge discovery. Basic use cases and business
drivers will be highlighted, and some insight will be given into a
new and mathematically grounded system and methodology.
Influences of database technologies will be discussed, and the
research system of TU Dresden database group will be
introduced. This is followed by a rough comparison with prior
art systems, which highlights advantages of the new concept. This
approach substantiates some core concepts of “Industrie 4.0”.
Keywords—real-time information bases; Industrie 4.0; linear
system based information processing; decompositional – Cyber-
Physical – system model; solution for “shortest algorithm problem”
I. INTRODUCTION AND BACKGROUND
The vision of an integrated data collection and information
processing system goes back to the 1990’s, and SEMATECH
is still maintaining its CIM Data Model. But for different
reasons the initial, very ambitious CIM concept failed, and
was replaced by data collection and processing systems on
defined functional scopes. Manufacturing Execution Systems
(MES) addressed the functions and tasks between the ERP
level and the equipment automation level, including
functionalities like management of work in progress,
equipment utilization, dispatch/scheduling, manufacturing job
management and others. Besides these, additional, mostly
bottom-up grown applications and frameworks like APC,
FDC, complex integration of test equipment and others led
and are still leading to an ever growing degree of
diversification in the application landscape of modern
semiconductor production sites.
However, most of the applications are maintaining their own
data schemes and volumes. As a consequence, the overall
system neglects the availability of integrated and
comprehensive information, but provide instead numerous and
growing sources of data. Additionally, such data is of reduced
value, because only the correlation to other data creates the
required information in order to continue the production
cycles and transactions (example: all the data which represents
different characteristics of a production run, like cycle time of
the run, equipment setup parameters, quality control
parameters, other data used to identify the next process step,
next equipment and other boundary conditions (link times
etc.)). Other concepts like CEP (complex event processing) or
CPS (Cyber Physical Systems [1]) point toward an integrative
direction, while deriving meaningful information from atomic
events based on a decompositional system model. This is
indeed a starting point for the proposed approach.
This paper is structured as follows. The second chapter starts
with an example of the main focal point to achieve “real” real-
time exploitation and introduces the research design. The two
central use cases will be introduced and developed in the third
chapter. The fourth chapter shows empirical results, which
have been achieved together with Technical University
Dresden. The research system of TU Dresden database group
will be introduced. Finally, the fifth chapter introduces the
Real-Time Information Base as key enabler for Manufacturing
Intelligence.
[Finanziert aus Mitteln der Europäischen Union und des Freistaates
Sachsen, Projektantragsnr.: 100 120 140 / sponsored by European Union and
State of Saxony/Germany, project number.: 100 120 140]

II. RESEARCH DESIGN AND MATHEMATICAL FOUNDATIONS
In addition to the levels of automation and MES, and
complementary to ERP systems, a so called “Business
Intelligence” layer should serves as an integrated approach to
support business management and decisions within companies,
and is usually implemented with components of Data
Warehouse Systems (DWHs); e.g. for OLAP analysis and
further optimization and engineering of Fab / production
processes. The major drawback is that – caused by the
increasing number of heterogeneous, and often isolated,
bottom-up designed applications – the industrial processes are
lacking fundamental concepts to realize information (not data)
at the BI level [2]. This holds also true for the semiconductor
industry. Main tasks on the BI level are calculations of
performance indicators and other analytical tasks; but these
indicators are typically calculated, after the actual event occurs.
That is, all this data on BI level lacks real-time cadence, and
for this reason is not used for further Real-time analytics and
control. Typical reports on BI-level are Fab Cycle time reports,
production flow reports, production plan fulfillment reports,
equipment utilization reports and others.
Fig. 1 classical DWH / BI Architecture; major drawbacks
Additionally, the growing complexity of applications
decrease the value of the information which is represented
within the data structures of those applications. This growing
complexity hinders the ability to easily introduce new
functionalities (for example a functionality to additionally
manage energy efficiency of production equipment and
corresponding process steps). The research design of the
proposed solution is based on the concept to overcome such
blocking points through the introduction of a new
methodology. This methodology has already been developed
by SYSTEMA, supported by TU Dresden and initially
evaluated by XFAB.
The current approach is motivated by the relationships
between the logical descriptions of objects of the real-world
system and the real-time information system. For example, a
specific movement of a part of a machine may correspond to
the engineering concept "process start". Another movement or
event may correspond to an engineering concept called "cycle
time". Even another set of events may correspond to the
concept "production costs". Such concepts may be used in
different kinds of systems (MES, ERP, etc.). But all these
different kinds of systems could theoretically share the same
foundational ontology, or decompositional base system. Such a
foundational ontology is given by the “state tracking model”
and the “decomposition model”, as described by Wand and
Weber [3]. But unfortunately in practice – due to the effect of
encapsulation diversification – the different applications do
not share similar description models or ontologies. Figure 2
shows this problem. Any Fab event might be embedded into
many dimensionalities, such as Time (including different
periodic schemes, such as hours, shifts, months etc.), Product
(including different product hierarchies), and Production
Process (including hierarchies and also recursive structures).
The clouds are representing different application domains.
Fig. 2 the dimensional embedding of Fab events; Prior art
(bottom): fundamental isomorphism of Fab events to linear
spaces is unknown (with regard to aggregation of information);
reason for complex and error-prone data processing
The problem domain is to be described as the availability of
information (in a most general logical, qualitative and
quantitative sense) in order to monitor, supervise, and qualify
any kind of industrial / business / financial process, to steer,
control, drive and optimize such processes. Assuming different
objects or processes (like business processes, financial
processes, engineering processes), which are characterized
through specific and well defined figures. Typically, such
figures are given as performance indicators, engineering
measurements (for example: physical measurements (within
semiconductor industry termed inline-data), functional
measurements (within semiconductor industry termed test-
data), derived measures (example from the semiconductor
industry: yield)), or logical associations / attributions in the
general and abstract case. This includes for example
engineering measurements, aggregation of values required for

performance indicators, logical values with regard to specific
definitions, for example the logical state of an equipment
(“unscheduled down”), of a lot, a (sub-)product (“on hold”); in
general: of all kinds of material parts (product parts, equipment
parts etc.) and / or processes and sub-processes (business
processes, physical processes in production, equipment, etc.).
SYSTEMA has conducted analysis of the mathematical
structure describing how information is derived from these
scenarios. This study has concluded that any corresponding
system model in all the different domains and applications
must incorporate the structure of the decompositional system
model, as defined above. That is, because of the compositional
characteristics, any parameter or data component, which
describes the behavior of subsystems on the lowest level of
granularity, can be grouped and aggregated with corresponding
parameters using historical records, and within the
mathematical concept of linear information spaces. The
decompositional system model preserves the linearity of the
overall model, and defines the corresponding linear relations of
the historical records. It can be conceptualized as a
nondeterministic system (a Cyber-Physical System [4]).
Fig. 3 Schematic representation of the new methodology:
any Fab event (atomic data portion) gets immediately
transformed into incremental Information Fundaments
A fundamental premise of the newly developed
methodology (Fig. 3), is that all incoming data elements are
immediately captured and transformed to components of
information. For example, out of the event “production lot
process step terminated at [date : time]”, will result in ‘real-
time’ calculation of the information component, which
represents the proportionate partial data value for the desired
characterization of the production process (kinds of
performance indicator, or any other desired aggregation). That
is, any such further calculation is taking place and occupying
linear vector spaces in a strictly mathematical sense; any
further processing takes place within the linear information
framework. This approach is practicable, since the
corresponding information is relatively small and is being
processed - within a newly designed ETL - as soon as it is
available. Such minimized computational effort is not possible
within common solutions even for approaches utilizing small-
scale aggregation strategies (i.e. data to be aggregated is split
into small batches, which fit in memory).
The linearity of these information spaces offers very
important properties because any data component can be
processed independently enabling the desired Real-Time
capability of the overall system. For this reason, any further
information on such data components - all performance
indicators, KPIs and the like are calculated based on the
desired atomic data components – and are calculated in Real
Time. The decompositional base system model is consistently
defining linear information spaces, without loss information,
and preservation of the ability to integrate information across
the whole business and/or industrial process, including
financial processes. The concept of hierarchical system
decomposition includes the capability of chaining and
hierarchically nesting such base systems, while preserving the
linearity of the informational spaces. To summarize, the
proposed solution and methodology offers following
advantages:
(i) the availability of continuously updated values for
performance indicators, or as-desired data aggregation
functionality at any point in time;
(ii) minimized usage of resources/energy through a new
system model;
(iii) highest performance to calculate any data aggregation
function;
(iv) support for complex and high performance ad-hoc queries;
(v) support for knowledge discovery analytic databases and
more generally the creation of new information.
Knowledge discovery in databases becomes possible due to
the high performance and flexibility of the present
methodology.
III. ARCHITECTURAL DESIGN AND BASIC USE CASES
The classical ETL process holds following structure:
1) Data Loading (from MES and other sources)
2) Data Transformation and Aggregation
3) Data Supply (of the Aggregated Data) into the OLAP
Structures

In summary, the same information is accessed multiple
times during the conventional ETL and aggregation steps. In
particular, these systems first transform and store operational
data and then re-access the data during the aggregation
processes, and finally calculate performance indicators [5].
Each of these steps causes corresponding CPU-load, I/O-load
and communication effort. For this reason prior art systems
hinder Real-Time aggregation because data extraction and
further transformation and aggregation processes are holding a
sequentialized, non-parallel structure, which conflicts with the
goal of parallel-processed Real-Time Data Warehousing [6].
The basic conflict arises from the problem that in prior art
systems the aggregation procedures which are required to
calculate the KPIs are started in batch mode. This barrier still
exists even if the refresh and updating cycles are minimized or
redesigned to run incrementally. Data Warehouse updates
cannot be executed anymore during off-peak hours. Some
authors argue that update anomalies may arise when base
tables are affected by updates that have not been captured by
the incremental loading process [7]. The next incremental load
could, but must not necessarily resolve such anomalies. Similar
anomalies may arise for deletes and inserts. As a consequence
ETL jobs have to be modified and extended in order to prevent
such anomalies and to approach at least a Real-Time
characteristics of the system.
In contrast to this the present method and system is based
on an analysis of the overall and fundamental informational
structure. As a consequence – because of the logical
independency of each business / production process element –
any data which characterizes such business / production
processes can already be created and extracted out of the
development of any single business process element during
current execution time (and will be stored in fundamental data-
components, i.e. fundamental atomic datasets). This leads to
following new system architecture; note that the BI-level is
now integrated into the communication infrastructure.
Fig. 4 Real-Time Information Base in order to enable Real-
Time Manufacturing Intelligence (RT-MI) within a target IT
Framework – horizontal and vertical Integration (Industrie 4.0)
It should also be noted – from the standpoint of data
management as the biggest challenge – that all atomic data is
immediately updated and stored within the RT-MI level. That
is for the first time immediate drill-down functionality will be
provided for any kind of information. And this information will
be provided within the Fab-wide perspective of this Real-Time
system. In summary, core concepts of “Industrie 4.0” are
substantiated [1].
The first use case concentrates on the replacement of
classical Reporting services by the core components of a Real-
Time Information System. All reports are replaced by the new
system, including continuous updates in Real-Time of any
reporting figures during the reporting period (e.g. key
performance indicators). A Real-Time Transformation Process
replaces the classical ETL process and creates in a first step
Basic Atomic Datasets. The basic atomic datasets contain all
the information necessary for reporting and data mining, as
they refer to the lowest level of granularity required for
effective decision making and knowledge discovery in
databases. In contrast to basic atomic datasets, the
Fundamental Atomic Datasets contain summarized
information from a well-defined subset of the basic atomic
datasets which are regarded as an entity (transaction) from the
relevant processing / reporting point of view (including ad-hoc
analysis and data mining / knowledge discovery in databases).
Real-Time Aggregated Datasets contain continuously
aggregated, predefined performance indicators and evaluations.
The second use case focuses on the capability for a new
dimension of ad-hoc analytics and Knowledge discovery with
regard to a new Real-Time Information Base enabling also
cross-functional analytics for engineering tasks. State-of-the-art
implementations cover only heuristically and historically
grown information structures without any sensitivity to
algorithmic complexity and informational efficiency of the
solution. Fig. 4 shows the main categories of the new data
structures (Basic Atomic Data, Fundamental Atomic Data,
Real-Time Aggregated Data).
IV. EMPIRICAL RESULTS AND FURTHER ANALYSIS OF DATABASE
TECHNOLOGIES
A. Empirical Results
Fig. 5 Comparison to prior art: typical batch processes are
replaced by continuous aggregation of information, building up
the RT-Information Base

Fig. 5 shows an aggregation run over 10.000 data sets. The
classical, batch-oriented solution (red color) requires 900 sec
runtime. The Real-Time aggregated solution requires 8 sec.
runtime (runtime reduction of 100x). The reduced runtimes
hold for larger datasets. In further analysis we have compared a
row oriented database (MS SQL) versus a column oriented
solution (IQ Server). We applied a complex test set including
long-term load tests and full fab simulation as input data and in
parallel complex data aggregations on the MI-level. There is no
“best system” for the tested use cases (from low data volumes
to terabytes). The row oriented system showed stable
performance for all test cases. The column oriented solution
uses much less storage volume (due to the internal organization
of the data). But this solution failed for:
• Tests including continuous updates on the MI-level
(which is required for Real-Time behavior)
• Multi-user operation (typically, the reporting systems
are not used by hundreds of users)
B. Research Database System of TU Dresden
But these existing solutions are outperformed by the
research system of TU Dresden. Even if we consider that the
biggest effect comes from the new logical model. But still we
need to deal with huge amounts of data. In order to efficiently
support our envisioned solution we also have to enhance data
management systems on various levels, e.g. on an architectural
level. The ever-growing demand for more computing power
forces hardware vendors to put an increasing number of
multiprocessors into a single server system which usually
exhibits a non-uniform memory access (NUMA). In-memory
database systems running on NUMA platforms face several
issues such as the increased latency and the decreased
bandwidth when accessing remote main memory.
To cope with these NUMA-related issues, NUMA-
awareness has to be considered as a major design principle for
the fundamental architecture of a database system.
Additionally, we observe a worsening of the scalability of
latches and atomic instructions. This is a result of the cache
coherence maintenance overhead induced by the NUMA
system. These issues are already measurable on wide-spread
server systems consisting of four or eight multiprocessors. The
demand for more parallel hardware forces vendors to put even
more multiprocessors into a single server system (e.g., Oracle
SPARC M6 with up to 96 multiprocessors or SGI UV2000 that
are sold as HANA-Boxes). We also expect emerging
technologies like 3D DRAM/CPU stacking to let NUMA
characteristics appear in a single multiprocessor, where each
core has its local low latency and high bandwidth main
memory. To allow database systems to scale-up on current and
future platforms, NUMA- awareness has to be considered as a
major design principle for the fundamental architecture of a
database system. Recent research revealed that a data-oriented
architecture (DORA) enables disk-based database systems to
scale-up on multicore systems in the context of transactional
workloads. DORA uses a thread-to-data instead of the
conventional thread-to-transaction assignment and thus
dramatically reduces contention on lock tables as well as latch
contention on data objects. Since such architecture relies on
logical partitioning, load balancing is necessary to adapt the
partitioning to workload changes. More recent works extended
the data-oriented approach by a physically partitioned disk
page buffer pool and a NUMA-aware as well as workload-
aware data placement algorithm, which tries to minimize inter-
socket communication.
In our work, we have developed a database system called
ERIS, a NUMA-aware all-in- memory storage engine for tera-
scale analytical workloads [8]. ERIS is also based on a data-
oriented architecture and thus splits data objects into partitions,
which are assigned exclusively to individual workers. Workers
are pinned on a designated core of the platform and execute
data commands (i.e., scan, lookup, or insert/upsert) on their
partitions. In contrast to existing approaches, ERIS aims at
tera-scale analytical workloads that are executed purely in
main memory and fundamentally differ from transactional
workloads. First, transaction-oriented systems try to cluster
partitions that belong to the same transaction class to minimize
inter-socket communication as well as interference between
trans- actions running in parallel. However, analytical queries
require a high degree of parallelism to execute with low la-
tency and thus data is best spread out as much as possible.
Second, analytical queries often generate huge amounts of
intermediate results and inevitably generate inter-socket
communication on NUMA systems. Hence – in opposition to
transactional workloads – the effective handling of inter-
mediate results, whose size grows with the size of the base
data, and the routing between workers are mission critical
components of our NUMA-aware storage engine. Third,
analytical workloads are usually read-only and thus the storage
subsystem should not implement comprehensive locking and
latching mechanisms, which could force a large scan operation
to block. The more feasible use of a non-blocking multi-
version approach or the use of staging tables eliminates the
need to optimize for minimal lock table contention. Finally,
analytical workloads execute scans or lookups on large data
objects. Hence, a storage engine that relies on logical
partitioning needs a topology-aware load balancing that is able
to quickly balance huge data objects with minimal impact on
the overall query throughput.
V. REAL-TIME INFORMATION BASE AS KEY ENABLER FOR
MANUFACTURING INTELLIGENCE
A prototypical implementation of this Real-Time
Information System has been evaluated and tested at X-FAB
Erfurt. For the first time, operational, strategic and tactical data
become available and add direct value for engineering and
optimization tasks along the supply chain. A first production
system optimization has been done. X-FAB manages a specific
database, representing a so called “process step / equipment
time model”. This database manages the production time of all
process steps, including their mapping to equipment, and
further including dynamic dependencies of equipment loads. It

serves as an interface between production planning and
production control.
X-FAB has now implemented a functionality covering
dynamic analysis and visualization of remaining equipment
runtime, as based on the “equipment time model”. In simple
words, initiated on any “equipment / process step – start event”
formerly invisible remaining equipment runtimes are becoming
visible and exploitable from within a Fab-wide perspective for
every process run. This will directly be used as input for further
dispatching and manufacturing flow control. As a first
outcome, reduction in cycle time becomes measurable. X-FAB
anticipates a potential improvement of around 15% with regard
to production efficiency. It needs to be noted that the
integration of the module “process step / equipment time
model” has been done most advantageously within the context
of the new overall system architecture (one could argue, that
this functionality could be a typical MES component; but the
team has consciously eliminated the boundaries given through
MES architectures, such as: missing connectivity to other, non-
MES related information components like fab performance
indicators – inspired through the “Industrie 4.0” vision).
Fig. 6 Automatic Batch creation on upstream Process Steps
Those two examples may be seen as representatives of the
first use case, as outlined in paragraph III (replacement of
classical reports, and usage of real-time KPI metrics). The next
example will point towards the second use case (enabling of
knowledge discovery and ad-hoc creation of new chains of
value). Both use cases are based on the Real-Time Information
Base, as outlined in Fig. 4.
The next figure indicates an ad-hoc analysis and further
data drill-down capability with regard to production bottleneck
scenarios.
Fig. 7 Usage of the Real-Time Information Base for
Knowledge Discovery and Ad-hoc creation of new chains of
value; the upper part (heatmap) is showing a production line
(ordered by its Process Steps), and the evolution of certain
KPI’s (in this example: WIP) over Time; this permits to
visualize the appearance of dynamic bottlenecks
But even more important, new planned MEMS products
will add tremendous complexity to the Fab. Formerly linear
organized production routes are getting transformed towards
Multi Device Integration. This implies potentially increasing
variations of orders of process steps. This complexity needs to
be represented within new kinds of Fab and Production
indicators, and at the same time provide a categorically new
concept of dynamic and Real-Time Fab Exploitation, including
new manufacturing prognosis capabilities. This is required in
order to minimize learning cycles and enable smooth
production ramps of the new MEMS products. As a
consequence, the implementation of the new solution will be
seen as a key enabler in order to achieve the mission of the
Fab. The next figure shows an example of the usage of Real-
Time Fab KPI’s in order to optimize the production

throughput. In this example WIP KPI’s are calculated in Real-
Time, and WIP thresholds (which are dynamically adjusted
with regard to actual equipment setup) are used to
automatically initiate batch creation and prioritization on
upstream process steps.
In summary, the Real-Time Information System is seen as a
base fundament in order to achieve X-FAB’s vision of Real-
Time Manufacturing Intelligence and Industrie 4.0.
ACKNOWLEDGMENT
The authors wish to acknowledge the close cooperation of
the entire project team with TU Dresden Database group – led
by Prof. Wolfgang Lehner – and with XFAB IT Automation
department – led by Volker Schmalfuss. Without their
motivation and support the reach for a more fundamental
solution concept would not have taken place, and the detailed
empirical results would not have been achieved.
REFERENCES
[1] H. Kagermann, W. Wahlster, J. Helbig, “Recommendations for
implementing the strategic initiative INDUSTRIE 4.0” Final report of
the Industrie 4.0 Working Group, Frankfurt/Main, April 2013
[2] H. Lasi, “Industrial Intelligence – a business intelligence-based
approach to enhance manufacturing engineering in industrial companie,”
8th CIRP Conference on Intelligent Computation in Manufacturing
Engineering, proceedings, Procedia CIRP 12 (2013), pp. 384-389
[3] Y. Wand, and R. Weber, "Toward a theory of the deep structure of
information systems." Information Systems Journal. Volume 5, Issue 3,
pages, 213-223, July, 1995
[4] E. A. Lee and S. A. Seshia, “Introduction to Embedded Systems - A
Cyber-Physical Systems Approach”, LeeSeshia.org, 2011.
[5] P. Ponniah, "Data Warehousing Fundamentals: A Comprehensive Guide
for It Professionals," 2010, John Wiley & Sons
[6] M. Thiele,W. Lehner, D. Habich,"Data-Warehousing 3.0 - Die Rolle
von Data-Warehouse-Systemen auf Basis von In-Memory Technologie.
[the role of Data-Warehouse-Systems based on In-Memory
Technology]" IMDM, 2011: 57-68
[7] T. Jörg, S. Dessloch: "Near Real-Time Data Warehousing Using State-
of-the-Art ETL Tools." BIRTE, 2009: 100-117
[8] T. Kiefer, T. Kissinger, B. Schlegel, D. Habich, D. Molka, W. Lehner,
“ERIS live: a NUMA-aware in-memory storage engine for tera-scale
multiprocessor systems,” proceedings, SIGMOD Conference 2014: pp.
689-692