New FreSH Framework Boosts Time Series Classification Accuracy

A new framework called FreSH (Frequency-Segmented Hierarchical Multi-Expert Framework) has been developed to tackle the challenges of Multivariate Time Series Classification (MTSC). MTSC requires models to effectively identify complex patterns in data that change over time, often across various scales, while remaining computationally efficient. Existing methods frequently struggle with learning fine-grained representations, especially when dealing with imbalanced datasets or real-world constraints. FreSH addresses these issues by introducing an adaptive, multi-scale analysis approach for temporal signals. It allows different aspects of the time series data to be modeled in a complementary and coordinated fashion, combining specialized local analysis with a broader understanding of the overall context. This design enables the framework to achieve strong representational capabilities without incurring excessive computational costs. The framework also features an adaptive fusion strategy, which dynamically emphasizes the most informative components of the input data, and a more robust optimization objective to enhance learning stability across diverse sample difficulties and class distributions. Extensive testing on 30 benchmark datasets and real-world vibration data demonstrated that FreSH consistently surpasses current state-of-the-art methods in classification accuracy, while also reducing model size and improving efficiency.