Probabilistic Methods for Predictive Maintenance and Beyond: Graph and Human-in-the-Loop Machine Learning

Thumbnail Image
Journal Title
Journal ISSN
Volume Title
School of Science | Doctoral thesis (article-based) | Defence date: 2024-02-15
Degree programme
61 app. 65
Aalto University publication series DOCTORAL THESES, 25/2024
Probabilistic methods are key tools for machine learning problems. Even so, there remain many applications where they cannot be applied due to their limitations. These limitations may include the lack of methods for a particular data format (e.g., manifolds, texts, or graphs), data unavailability, or the inability to work collaboratively with human experts. Inspired by problems in predictive maintenance (PdM), this thesis introduces a set of machine learning solutions that are more generally applicable. It begins with applied tasks in cable networks, data centers, and other telecom applications and indicates the crucial limitations of current approaches: the absence of (i) probabilistic methods for spatio-temporal graph problems, (ii) practical human-in-the-loop methods that learn from detailed domain experts' feedback, and (iii) systems for synthetic temporal data creation that enable secure sharing of sensitive data between parties. Moreover, even if such methods become available, it is important to describe how those methods can be used in an end-to-end system for predictive maintenance covering both the modeling and operations sides. This thesis analyses and resolves these issues. The first issue, the lack of probabilistic methods for graph and spatio-temporal graph data, was resolved by connecting graph kernels with stochastic partial differential equations (SPDEs). This method results in a variety of kernels suitable for machine learning problems on graphs, including Mat\'ern, stochastic heat, and stochastic wave graph kernels. The second issue, the lack of human-in-the-loop methods with domain experts' explicit feedback, was resolved by developing a decision rule elicitation mechanism and its domain adaptation properties. The method is grounded in human decision-making mechanisms and has been tested in several user studies. It leads to a simple yet effective method for working with domain experts. Next, synthetic data generation was resolved by introducing an open-source software framework called TSGM. This framework effectively generates synthetic time series data and provides a toolkit for evaluation. This work also examined the various approaches to the generation and evaluation of synthetic data. Finally, the methods proposed in the thesis resulted in successful real-world implementations tested on several large-scale cases with our industrial partner Elisa Oyj. Furthermore, those implementations led to five submitted patents, one of which has already been granted. This thesis discusses the aforementioned results, places them into a broader perspective, and provides possible avenues for future research.
Supervising professor
Kaski, Samuel, Prof., Aalto University, Department of Computer Science, Finland and The University of Manchester, United Kingdom
machine learning, deep learning, Gaussian processes, stochastic differential equations, generative models, predictive maintenance, human-in-the-loop, open source
Other note
  • [Publication 1]: Alexander V. Nikitin, S. T. John, Arno Solin, and Samuel Kaski. Non-separable Spatio-temporal Graph Kernels via SPDEs. In International Conference on Artificial Intelligence and Statistics, Online, May 2022.
    DOI: 10.48550/arXiv.2111.08524 View at publisher
  • [Publication 2]: Alexander Nikitin and Samuel Kaski. Decision Rule Elicitation for Domain Adaptation. In 26th International Conference on Intelligent User Interfaces, Online, April 2021.
    DOI: 10.1145/3397481.3450682 View at publisher
  • [Publication 3]: Alexander Nikitin, and Samuel Kaski. Human-in-the-Loop Large- Scale Predictive Maintenance of Workstations. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington DC USA, August 2022.
    DOI: 10.1145/3534678.3539196 View at publisher
  • [Publication 4]: Alexander Nikitin, Letizia Iannucci, and Samuel Kaski. TSGM: A Flexible Framework for Generative Modeling of Synthetic Time Series. Submitted to a conference, February 2023.
    DOI: 10.48550/arXiv.2305.11567 View at publisher