A Multi-scale Multivariate Time Series Classification Method Based on Bag of Patterns

Multivariate time series classification (MTSC) has become a crucial challenge with widespread implications in diverse fields, ranging from astronomy to medical analysis. The primary hurdle in MTSC lies in effectively integrating multi-dimensional information, setting it apart from univariate time series classification (UTSC). In response to these challenges, we propose an innovative solution-a multi-scale multivariate time series classification model. This model harnesses a multi-scale feature extraction network and a bag-of-patterns method to comprehensively learn morphological and local features across various scales. Notably, our method excels at integrating information from corresponding positions across different dimensions, a critical capability for distinguishing MTSC from multiple UTSC scenarios. Our proposed method demonstrates remarkable accuracy on UEA Archive dataset when compared to existing methods. This success underscores the effectiveness of our approach in addressing the inherent complexities of MTSC, offering a promising solution for precise and robust classification in real-world applications.