ime series data are widely used in various domains and applications, such as emotion recognition, sleep stage detection, and machine remaining useful time detection. Unlike other modalities, time series data are more abstract and complicated, with strong temporal dependencies and complex patterns. Such characteristics make the process of labeling time series data a complex, time-consuming, and costly process, requiring specialised expertise. Among methods dealing with unlabelled time series data, multimodal learning methods stand out by incorporating information from multiple modalities. Using modality-specific and shared information, multimodal methods are expected to be better at using the unlabelled time series data to improve model performance. In this project, we are exploring the best practices of applying multimodal learning with unlabelled time series data.