Time Series Representation Learning for Industrial Process Characterization

Time Series Representation Learning for Industrial Process Characterization

The M-Group at KU Leuven Bruges Campus is an interdisciplinary research team focusing on intelligent and dependable mechatronic systems, combining research expertise from the departments of Computer Science, Electrical Engineering and Mechanical Engineering. One of the key research tracks focusses on the application of Artificial Intelligence and Machine Learning in real-world industrial settings on the one hand. The objective of the proposed PhD position is to investigate how novel techniques for time series representation learning on sensor data origination from complex and dynamic industrial processes can be leveraged for process characterization across multiple time series data mining tasks, such as anomaly detection, anomaly diagnosis and root cause analysis. The successful candidate will be offered to opportunity to pursue a PhD in Computer Science at KU Leuven, and will also be embedded within the Declarative Languages and Artificial Intelligence (DTAI) lab (https://dtai.cs.kuleuven.be), which pursues excellence in an explicit and synergistic combination of fundamental and applied research on machine learning and artificial intelligence. Doctoral training is provided in the framework of the Leuven Arenberg Doctoral School (https://set.kuleuven.be/phd). The PhD will be supervised by Prof. Mathias Verbeke and Mr. Louis Carpentier.

Context
Industry 4.0 has equipped industrial processes with large numbers of sensors, enabling continuous monitoring of machine and process behaviour. The resulting data allow the process to be modelled, which in turn yields analytical insights for optimising the process or for detecting and preventing faults. However, industrial processes are becoming increasingly complex, with many interacting components, varying operating conditions, and hundreds of monitored variables, which makes modelling the process as a whole increasingly difficult.

Existing approaches rely mostly on deep neural networks to learn representations of the data, which are then leveraged in downstream tasks. While effective in terms of predictive performance, these models provide little understanding of the process itself, which is crucial in real-world applications. Post-hoc explanations do not solve this issue, as they only approximate the model's reasoning and are not guaranteed to faithfully reflect its behaviour. Furthermore, deep neural networks typically require large datasets to learn complex behaviours, which are rarely available in real industrial applications.

Project
The central research challenge of this PhD position is to automatically learn explainable representations from high-dimensional industrial time series data, and to leverage these representations across multiple time series data mining tasks. Within this PhD research, you will leverage motifs, i.e., representative subsequences, which act as the building blocks of the explainable time series representations and thereby offer a holistic view of the industrial process. To this end, you will design new algorithms for mining motifs in multivariate time series and for learning relationships among the discovered motifs. These algorithms will focus on three aspects. First, they will balance scalability and representativeness, i.e., the computational cost of building the motif-based representation against the level of detail and accuracy with which the motifs describe the process. Second, they will incorporate weak supervision, so that domain experts can steer the algorithms, and hence the motif-based representation, without needing to provide exhaustive labels. Finally, they will be simple to use, requiring minimal configuration with few hyperparameters whose meaning is intuitive to domain experts. The resulting representations will be evaluated in downstream tasks such as anomaly detection, anomaly diagnosis, and root cause analysis.

The PhD position is part of an industrial research project with the aim of acquiring foundational knowledge about complex industrial systems and generating analytical insights that can be used to improve the production process. This PhD is embedded in the methodological core of the project, and is responsible for developing the representation learning methods on which the downstream analyses are built. The methods developed throughout the PhD will be validated on real-world industrial use cases leveraging operational data provided by the industrial partner in the project.

Profile

Profile

We are seeking a highly motivated, enthusiastic, and communicative researcher, with a proactive and creative attitude who is eager to explore innovative solutions. If you recognize yourself in the story below, then you have the profile that fits the project and the research group:

• I have a Master's degree in Computer Science, Artificial Intelligence