=Paper=
{{Paper
|id=Vol-3397/phd1
|storemode=property
|title=Process Mining on Distributed Time-Series Data
|pdfUrl=https://ceur-ws.org/Vol-3397/phd1.pdf
|volume=Vol-3397
|authors=Frederik Fonger
|dblpUrl=https://dblp.org/rec/conf/emisa/Fonger23
}}
==Process Mining on Distributed Time-Series Data==
https://ceur-ws.org/Vol-3397/phd1.pdf
Process Mining on Distributed Time-Series Data (PhD
Proposal)
Frederik Fonger1
1
Kiel University, Group Process Analytics, Hermann-Rodewald-Str. 3, 24118 Kiel, Germany
Abstract
Process mining techniques are used for the discovery of process models from recorded events, to analyze
the conformance of a specification derived from recorded events and a process model, and for predictive
analytics. However, mostly the recorded events come from (business) data of IT systems and process
mining techniques have been developed to process structured data that is at a high (business) level of
abstraction. Plenty of scenarios exist with low-level data and where process mining could give valuable
insights when analyzing these kinds of data. The purpose of this PhD proposal is to design a process
mining pipeline for time-series data in distributed settings. The challenges for process mining in such
a scenario are that usually no ground truth exist to learn and optimize against nor techniques exist to
efficiently process the high volume of time-series data, which is typical in such scenarios. To address
these challenges, we suggest a process analytics pipeline relying on the generation of synthetic data and
data sampling.
Keywords
PhD Proposal: Process Mining, Time Series, Data Sampling
1. Introduction
Process mining techniques are used for the discovery of process models from recorded events,
to analyze the conformance of a specification derived from recorded events and a process model,
and for predictive analytics. However, mostly the recorded events come from (business) data
of IT systems and process mining techniques have been developed to process structured data
that is at a high (business) level of abstraction. In plenty of scenarios (e.g., IoT settings) and
disciplines (e.g., natural or life sciences) low-level data is produced. One example for the analysis
of sensor data is in oceanography. With the rising importance of climate change, seaweed has
been identified as a large natural carbon storage possibility. By analysing time series data, we
can better understand the growth process of seaweed.
Process mining could give valuable insights when being capable of analyzing these kinds of
data [1]. For instance, time-series data is recorded in sensors and the data is analyzed with the
purpose to identify trends and patterns over time and make forecasts. The analysis of time-series
data could benefit from process mining. Its combination would allow to discover a process
and to analyze causal effects through the simulation of the process. However, time-series data
13th International Workshop on Enterprise Modeling and Information Systems Architectures (EMISA), May 11–12, 2023,
Stockholm, Sweden
Envelope-Open ffo@informatik.uni-kiel.de (F. Fonger)
Orcid 0009-0000-8445-8104 (F. Fonger)
© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
Cloud
Edge
Sensor 3 ... Sensor 2 ...
Sensor 1 generating
Distributed Sensors time-series data
Figure 1: Setting with distributed sensors and central processing units like Edge and Cloud.
is unstructured data and at a low level of abstraction, while process mining techniques were
developed for structured data that is at a high (business) level of abstraction. Therefore, time
series data needs to be processed appropriately to extract an event log.
Additionally, in IoT scenarios time-series data is generated distributed and mostly in a large
volume. This calls for an approach to efficiently process distributed time-series data for process
mining. To address this, we present a process analytics pipeline relying on the generation of
synthetic data and data sampling. Sampling allows to reduce the data size while preserving
the information within the data. The challenge of data sampling is to identify a representative
sample of the original dataset while reducing the volume of data.
The PhD proposal is structured as follows. The next section describes the research problem.
Section 3 summarizes a solution and challenges that have to be addressed. Related works are
discussed in Section 4, while the paper concludes with a summary.
2. Research Problem
The aim of the PhD proposal is to discover process models from distributed time-series data.
To achieve this, a pipeline to efficiently process and analyze the data has to be designed.
As mentioned above, distributed time-series data is generated in high volume, mostly with
inappropriate data quality and in scenarios with low latency. Therefore, we suggest a pipeline
relying on the generation of synthetic data and data sampling, while satisfying the requirements
in distributed settings. The synthetic data is generated on a low level by simulating the individual
sensors. The primary objective is to generate data for the development that is similar in
characteristics like the format, time intervals and synthetic dependencies with noise. The
synthetic dependencies are only for validating the developing methods and do not have to
mirror the real data. Real data will be used on the developed methods at a later point in time.
Fig. 1 shows a scenario with distributed sensors that a centrally processed at e.g., the edge and
cloud.
Existing process mining algorithms are not capable to process time-series data from such
scenarios. Activity recognition in time series is still an unsolved problem in time series analysis
[1]. In this respect, the mapping of activities onto a time series is still a challenge. In this PhD
proposal we will in particular focus on the volume and data quality aspect of time-series data
Raw data with
Time Series Event Log Process Model
control flow
timestamp activity timestamp activity case
01.01. Water Temp. up 01.01. Air Temp. up 1
01.02. Assets
Salt down 05.01. Assets
W. Temp. up 1
01.01. Chl. up 15.01. A. Temp. down 2
01.01. Wind up 17.01. W. Temp. down 2
Figure 2: Process analytics pipeline for time-series data for process mining
for process mining. This means that we also have to focus on the data perspective in distributed
settings to deal with the data quality aspect. Particularly, we suggest to generate synthetic
time-series data in order to bridge the gap. For this, we developed a structured approach to
map time-series data on control-flow patterns that we annotated for our purpose. Based on the
simulation of the patterns it is possible to generate synthetic data in varying quality, which is
again a crucial step for accurate results from machine learning techniques [2].
Fig. 2 shows the process analytics pipeline for distributed time-series data on an abstract level,
which we plan to implement in the PhD project. First, we map time-series data on control-flow
patterns in order to increase data quality through synthetic data1 . Next, we plan to apply
sampling strategies and to abstract activities from the data and finally to use process mining.
When generating an event log, the large uncertainty between measuring points in some datasets
is also essential to be considered.
3. Activities of the PhD proposal
To provide a solution several challenges have to be addressed: first, data sampling is necessary
to reduce the data size. Then, activity recognition for time-series is essential, too. To address
this, we plan to generate synthetic data with our tool, insert different levels of noise and then to
design a technique for activity abstraction. Depending on the use case, there may be long time
intervals between measurement points in the time series. This leads to uncertainty between
measurement points, which must also be taken into account. Finally, we have to evaluate our
techniques of mapping time-series data on process patterns. Particularly, we have to evaluate
and to quantify the occurrence of the patterns in real time-series data.
Table 1 shows an example of the time series data used in the approach. The real and the
synthetic data have the same data structure. Depending on the dataset, there are one or more
entries for each timestamp. In this example, the air temperature and the salinity are measured
weekly.
The next steps to implement the pipeline are improving the generating of synthetic data,
testing different approaches to identify activities in time-series data and designing a method
for dealing with uncertainty within the time-series data. Furthermore, a visualisation for the
synthetic time-series data generation will be implemented. We plan to evaluate different data
quality levels (i.e., less or more noise) to provide a solution for activity recognition. Also, we
1
This step is already completed. We refer to [2].
Timestamp Air Temperature Salinity …
2020-03-18 6.507 32.363 …
2020-03-25 7.484 32.313 …
2020-04-01 8.461 32.263 …
2020-04-08 9.291 32.251 …
Table 1
Example of time-series data from the marine use case.
will evaluate the sampling methods for two different scenarios. This research is relevant for
uncovering phenomena and underlying processes in natural and life sciences. Ziolkowski et
al. have already shown a first application of process mining on times series for data from
oceanography [3]. My work will continue in the same direction and build on it.
This PhD proposal is conducted within the Marispace-X project, which in this way presents
the scenario and real-time data. The purpose of the project is to develop a cloud-based platform
to improve data exchange and efficient processing of maritime data. The technical cloud
foundation relies on the GAIA-X framework. The data includes time-series data acquired from
distributed sensors from underwater locations, alongside data from single sensors mounted on
research vessels and stations.
4. Related Work
Herbert et al. proposed a methodology for generating synthetic time series data for process
analytics [4]. However, this approach lacks the ability to specify the effects present in the data.
To address this limitation, my proposed approach provides more flexibility in configuring the
effects represented in the data, as well as the complexity level. An approach that uses process
mining for time series data from smart products was presented by Eck et al. [5]. The approach
applies human activity recognition on the data collected by the smart products and subsequently
event logs are generated for process mining.
A challenge when using process mining methods can be the uncertainty in the data [6, 7].
Pegoraro et al. introduced a concept and a tool for not deleting and losing data, but rather
incorporating the uncertainty into a resulting model [6, 7]. Another challenge stemming from
the uncertainty in time series data is the presents of imprecise timestamps. The evaluation
of such partially ordered events has already been addressed by Lu et al. [8]. Process mining
for marine time series was introduced by Ziolkowski et al. by using a clustering algorithm for
generating a event log from the time series [3]. Subsequently, a process mining algorithm was
used to mine a process model.
5. Conclusion
This PhD proposal presents a pipeline for discovering process models from distributed time-
series data. For the development and evaluation of novel methods, synthetic data is generated
in varying complexity and noise intensity. Ultimately, this will be used for developing a method
for mapping activities onto distributed time-series data. Furthermore, sampling approaches will
be evaluated for two different scenarios within the pipeline. In the end, we want to be able to
apply sampling and activity recognition on real marine time series data in order to generate
event logs and subsequently use this for process mining.
6. Acknowledgement
I would like to express my gratitude to my PhD supervisor, Agnes Koschmider, for her invaluable
guidance and support in writing my PhD proposal. I would also like to thank Milda Aleknonytė-
Resch, for her insights and feedback on my PhD proposal. Furthermore, this project has received
funding from the German Federal Ministry for Economic Affairs and Climate Action under the
Marispace-X project grant no. 68GX21002E.
References
[1] A. Koschmider, N. Oppelt, M. Hundsdörfer, Confidence-driven communication of process
mining on time series, Informatik Spektrum 45 (2022) 223–228.
[2] F. Fonger, M. Aleknonytė-Resch, A. Koschmider, Mapping time-series data on process
patterns to generate synthetic data, in: CAiSE Workshops 2023, Lecture Notes in Business
Information Processing, Springer, 2023. To appear.
[3] T. Ziolkowski, R. Schubert, M. Renz, A. Koschmider, Process Mining for Time Series Data,
Technical Report, 2022. doi:10.1007/978- 3- 031- 07475- 2 .
[4] T. Herbert, J. Mangler, S. Rinderle-Ma, Generating Reliable Process Event Streams and
Time Series Data based on Neural Networks, volume 421, 2021, pp. 81–95. URL: http:
//arxiv.org/abs/2103.05462. doi:10.1007/978- 3- 030- 79186- 5_6 , arXiv:2103.05462 [cs].
[5] M. L. van Eck, N. Sidorova, W. M. P. van der Aalst, Enabling process mining on sensor
data from smart products, in: 2016 IEEE Tenth International Conference on Research
Challenges in Information Science (RCIS), IEEE, Grenoble, France, 2016, pp. 1–12. URL:
http://ieeexplore.ieee.org/document/7549355/. doi:10.1109/RCIS.2016.7549355 .
[6] M. Pegoraro, Probabilistic and non-deterministic event data in process mining: Embedding
uncertainty in process analysis techniques (2022). arXiv:2205.04827 .
[7] M. Pegoraro, M. S. Uysal, W. M. P. van der Aalst, PROVED: A Tool for Graph Representation
and Analysis of Uncertain Event Data, in: D. Buchs, J. Carmona (Eds.), Application and
Theory of Petri Nets and Concurrency, volume 12734, Springer International Publishing,
Cham, 2021, pp. 476–486. URL: https://link.springer.com/10.1007/978-3-030-76983-3_24.
doi:10.1007/978- 3- 030- 76983- 3\_24 , series Title: Lecture Notes in Computer Science.
[8] X. Lu, D. Fahland, W. M. P. van der Aalst, Conformance Checking Based on Partially
Ordered Event Data, in: F. Fournier, J. Mendling (Eds.), Business Process Management
Workshops, volume 202, Springer International Publishing, Cham, 2015, pp. 75–88. URL: http:
//link.springer.com/10.1007/978-3-319-15895-2_7. doi:10.1007/978- 3- 319- 15895- 2\_7 ,
series Title: Lecture Notes in Business Information Processing.