Suboptimal or erroneous con guration data in rail automation systems may cause serious safety and cost issues. This research work aims to address continuous optimization of such data by (i) connecting con guration and operation data by integration of heterogeneous data sources, and by (ii) application of prescriptive analytics methods to propose decision options on how to correct and optimize con guration data.
In industrial and infrastructural systems, like rail automation, product con guration is the activity of engineering and customizing a product to meet the needs of a particular customer. The product in question may consist of mechanical parts, services, and software - each with various parameters and properties that re ect the variability. The result of a con guration process are the con guration data, i.e., a digital model of the system specifying all details for installation, operation and maintenance of the facility. In the case of a rail automation system, con guration data contain, e.g., the bill of material, detailed con guration plans of all hardware parts, the station and track topology, the screen layout of the operator terminals, parametrization of the control software, etc.
While con guration data is speci ed at engineering time, operation data is continuously generated by trains and the interlocking and control systems at operation time. The amount of operation data generated each day is huge, because it contains all positions and speeds of all trains, as well as logs of all telegrams exchanged by the di erent subsystems.
The combination of these two data models - con guration data and operation
data - allows for a feedback loop which has the potential to detect hardware
defects or errors in the con guration data. Anomalies and unexpected behavior in
operation data can be detected by statistical methods like principal components
analysis or discriminant analysis. Error causes can only be detected by locating
the corresponding con guration objects in the con guration data. This promising
setting enables to build a prescriptive analytics framework [
Another interesting application enabled by the availability of con guration and operation models is predictive asset management, where prediction models for the obsolescence of the various hardware parts are computed from di erent heterogeneous data sources. Con guration and operation data are complemented here by sales and order forecast models and contextual data, like weather data.
Based on these challenges, the following research questions will be tackled in the proposed dissertation thesis: { Which data integration processes are suitable for preparing con guration and operation data for prescriptive analytics applications? { Which sequences of statistical methods are appropriate for predictive asset management and anomaly detection in con guration and operation data? { How can new or modi ed con guration rules/constraints be derived by prescriptive analytics methods in order to optimize con guration data? 2
Beyond the application of mere statistical analysis methods, data analytics
methods require a federated architecture of descriptive, predictive, and prescriptive
analytics in combination with data models and a data warehouse [
Data analytics in rail automation gains more and more interest [
Following the statistical results of descriptive and predictive analytics, we
believe in applying prescriptive analytics methods to propose optimal decision
options for optimizations in product con guration. Basically, con guration can
be de ned as a "special case of design activity, where the artefact being con
gured is assembled from instances of a xed set of well-de ned component types
which can be composed conforming to a set of constraints" [
Data analytics is a highly active research eld which has been driven mainly in business applications in the last decade. Recently, more and more industrial applications are adapting these methods, e.g., for sensor analysis for predictive maintenance. We explore the links between data analytics technologies and product con guration. The integration of con guration data in the process of analyzing operation data enables an easy localization and therefore a better understanding of the source of an anomaly.
In this work we contribute to solutions for two crucial problems in the domain of rail automation at the company Siemens AG O sterreich: (i) predictive asset management and (ii) prescriptive analytics for correcting/optimizing railway engineering data. In the rst scenario, forecasts of the form "How many assets of type A will be needed within the next N years?" are computed. Reliable forecasts are very important for guaranteeing the availability of all necessary modules in the future and allow for a solid version and lifetime management of module variants. The second scenario - the continuous correction and optimization of con guration data - is an important building block of guaranteeing safe and high-performance train operation.
The framework developed within the proposed dissertation goes beyond the rail automation use case. The methods developed are of general nature and may be adapted and applied to other industrial elds such as industry automation, power plants, or energy management. This is novel and highly promising, especially in the context of Smart Production and Industry 4.0.
According to the research questions, we design a framework of methods with the
following contributions.
1. Heterogeneous Data Model Integration. We need to integrate various
data sources of di erent formats, like Excel and XML. As di erent business
units use di erent tools and formats to maintain data, integration of data
is challenging and prone to errors. Existing data quality methods fall short
of a generalized approach that covers such a variety of data types in the
domain of rail automation. Our contribution extends the notion of signi ers
for a robust and at the same time typo-tolerant identi cation of objects of
di erent sources [
The project related to this research has been started in April 2016. The work follows the contributions described in Section 4 assuming that they build on each other.
The rst stage, Heterogeneous Data Model Integration, is nished. The
result of the this contribution is an approach \Using Signi ers for Data Integration
in Rail Automation" which was presented at the 6th International Conference
on Data Science, Technology and Applications [
The di erent data analytics tasks for predictive asset management and anomaly detection in operation data are de ned and documented in a user requirements speci cation. Next, we will study the applicability of di erent multivariate methods to our analytics tasks. The selection and application of statistical methods is a highly sensitive task since the results serve as basis for further prescriptive analytics methods to optimize rules and constraints in product con guration.