MVFusion-TSC: A multi-view fusion image-based network for time series classification

Abstract

Time series classification (TSC) is a supervised task that aims to categorize time series data into predefined classes by identifying meaningful patterns within their temporal structure. Current methods face numerous challenges in handling time-dependent modeling, capturing long- and short-term relationships, and feature extraction. To address these issues, we propose a time series classification framework, MVFusion-TSC, based on multi-view fusion image representation (MVFI) and a dual-branch feature fusion network (DFFNet). This framework uses the MVFI method to simultaneously encode local texture variations and global structural trends, and uses DFFNet, where one branch focuses on fine-grained local patterns while the other captures holistic temporal structures, effectively revealing and extracting both short- and long-term temporal dependencies. First, MVFI method is employed to convert time series data into multi-view fusion (MVF) images that integrate structural, color, and texture information. This enables the time-varying amplitude features of time-series data to be represented as structural and texture patterns in images, which better capture temporal dependencies and improve the performance of deep learning models. Secondly, DFFNet network is constructed, including a local feature extraction branch, a global feature extraction branch, and a gating fusion module, to achieve the extraction, fusion, and classification of both global and local temporal dependency features. Experiments on 25 benchmark time-series datasets show that the proposed method outperforms the current state-of-the-art methods in key metrics such as accuracy, win rate, and ranking, verifying its effectiveness and advantages in time-series classification tasks. The proposed method can provide a new research perspective for time-series classification, with significant theoretical value and practical application potential.