Classification of Multivariate Time Series Signals Using Self-Supervised Representation Learning for Condition Monitoring

In the analysis of real-world data, two significant challenges often arise: high-dimensional signals and their temporal interactions. To address these issues and identify transitions in process conditions through end-to-end learning, we propose a self-supervised representation learning framework that conceptualizes signals as images. This straightforward approach classifies imaged signals as time-specific conditions by maximizing mutual information within a domain-specific feature space. We applied this methodology to two labeled open datasets and one unlabeled real-world process dataset, yielding promising results.