Monitoring of Industrial CPS Behaviour through Stream Processing
Description
Industrial Cyber-Physical Systems (ICPS) are designed and intended to run for long intervals. The real-world experience however is far from this requirement. A plethora of root causes could result in interruptions, which ultimately translate to extra cost and service disruption. Example causes can be listed as software bugs, component wear and tear, substandard input, environmental condition fluctuations, lack of calibration, and so forth.
A rich body of literature exists on the topic of anomaly detection/identification and predictive maintenance, our focus subjects. Our particular techniques of choice are the use of advanced data processing and Machine Learning (ML). We aim to take advantage of such techniques in an online fashion, i.e., at runtime.
Task
The expectation is to generate machine-specific signatures with proper cataloguing, as well as performing online ML inference. Sample traces will be provided. The student could consider a simple continuous data source, emulating trace generation. Alternatively, real computational and I/O workloads could be considered. The potential tooling for this project is diverse. The student is required to try out, compare, select and deploy a suitable setup [1, 2, 3]. Possible candidates are: Apache Kafka, Apache Flink, InfluxDB.
For ML tasks, we rely on the PyTorch library. While we do have methodologies to follow for postmortem data analysis [4], performing similar tasks in an online fashion is rather challenging. Though we strive for definitive results, whether detection or identification, there will be two main constraints: timeliness and data limits, which go hand in hand. Online processing can only consider limited buffer sizes, i.e., immediate historical data, which is induced by the timeliness requirement. We simply cannot wait too long and miss the deadline for a useful detection result. The opposite is equally valid too. Lengthy processing of large data, even if available, will take long. Techniques to save historic trends, whether independent or integrated within the ML model, are valuable.
Application
The results of this study may be integrated in the research project ZORRO, as well as other industry collaborations. The outcome is expected to boost the state-of-the-art for anomaly detection and predictive maintenance topics, alongside the relevant tooling.
While not mandatory, we encourage and support our students in writing conference/workshop papers based on Master projects.
Level
MSc
Contact
Uraz Odyurt, u.odyurt@utwente.nl
Alex Chiumento, a.chiumento@utwente.nl
Berend Jan van der Zwaag, b.j.vanderzwaag@utwente.nl
References
[1] Stonebraker, 2005, The 8 Requirements of Real-Time Stream Processing. URL: https://doi.org/10.1145/1107499.1107504
[2] Almeida, 2023, Time series big data - a survey on data stream frameworks, analysis and algorithms. URL: https://doi.org/10.1186/s40537-023-00760-1
[3] Fragkoulis, 2024, A survey on the evolution of stream processing systems. URL: https://doi.org/10.1007/s00778-023-00819-8
[4] Odyurt, 2022, Improving the robustness of industrial Cyber–Physical Systems through machine learning-based performance anomaly identification. URL: https://doi.org/10.1016/j.sysarc.2022.102716