Meta-learning for Multivariate Signals

Title Meta-learning for Multivariate Signals
Summary Apply meta-learning algorithms to unlabelled time-series data to solve machine activity recognition problems.
Keywords Meta-learning, domain adaptation, neural network, time-series data, activity recognition
TimeFrame 2022 Fall - 2023 Summer
References [[References::[1] Finn, Chelsea, Pieter Abbeel, and Sergey Levine. "Model-agnostic meta-learning for fast adaptation of deep networks." International conference on machine learning. PMLR, 2017.

[2] Ganin, Yaroslav, et al. "Domain-adversarial training of neural networks." The journal of machine learning research 17.1 (2016): 2096-2030.

[3] A. Fischer, A. B. Bedrikow, S. Kessler and J. Fottner, "Equipment data-based activity recognition of construction machinery," 2021 IEEE International Conference on Engineering, Technology and Innovation (ICE/ITMC), 2021, pp. 1-6.

[4] Xianjie Gao, Maolin Shi, Xueguan Song, Chao Zhang, Hongwei Zhang. “Recurrent neural networks for real-time prediction of TBM operating parameters.” Automation in Construction, Volume 98, 2019.]]

Prerequisites A solid background in neural networks; good knowledge of Python for implementing deep learning algorithms.

Courses: Learning System; Deep Learning

Supervisor Kunru Chen, Anna Vettoruzzo, Mohamed-Rafik Bouguelia
Level Master
Status Open


In the machine learning community, meta-learning or learning to learn indicates the ability of the models to accumulate experience from multiple domains and exploit it to achieve fast adaptation to a new domain. Multiple attempts have been made in the past for solving few-shot supervised problems, where only a few labelled data are available, e.g., see [1]. However, it is difficult and expensive to obtain labelled data in many real-world applications, opening the way to unsupervised meta-learning. Let’s take for example the machine activity recognition (MAR) problem where the aim is to extend the knowledge learned for classifying different activities to different countries, warehouses, or machines. This problem could be addressed with domain adaptation methods, like DANN [2], but the time required for adaptation and the resource for training the algorithm is huge. Therefore, exploiting meta-learning to solve MAR problems could lead to a significant improvement in the field.

In this project, time-series data are collected in warehouses from different countries (i.e., Sweden, Norway, Italy, and Czech) during normal operations of forklift trucks. These data are unlabelled, and they consist of 262 multi-variate features acquired at 10 Hz for 2 months. The goal is to recognize three forklift activities performed by trucks from working places. A processing operation has been implemented to assign pseudo-labels to the data at the scope of recognizing three different classes: driving (D), loading (L), and other (O). The project aims to apply meta-learning to the unsupervised setting with the scope of exploiting the knowledge acquired from Sweden, Norway, and Italy to predict the correct activity on Czech data.

Research questions:

1-How to achieve domain adaptation for solving MAR problem?

2-How to apply meta-learning with real-world time-series data?

3-How to extend the knowledge acquired from a set of domains to a new unlabelled domain?

4-What are the challenges of this method for MAR?

If you want more information about this topic, you can contact us with and, or pass by our office at F504a.

Note: the project is a part of the collaboration between Halmstad University and Toyota Material Handling Sweden AB.