Adaptive Clustering Framework for Time-Series Data Using Enhanced FastDTW
Jiahao Tao, Qun Wu, Liang Zhao, Liang Chen · 2025
Clustering time-series data presents considerable challenges, driven by its inherent high dimensionality, dynamic variability, and nonlinear temporal dependencies. To overcome these obstacles, this study introduces an Adaptive Clustering Framework for Time-Series Data, built upon the foundation of Adaptive Constraint FastDTW (ACFastDTW) and further extended with the enhanced ACFastDTW-SALKM. The ACFastDTW algorithm utilizes a CatBoost-regressed adaptive windowing mechanism, which significantly boosts both the accuracy of distance calculations and computational efficiency. Building on this foundation, ACFastDTW-SALKM employs a self-attention LSTM autoencoder for extracting critical features and combines it with K-means clustering to achieve precise segmentation of time-series data. Through validation on both synthetic and industrial datasets, the framework showcases substantial improvements in clustering quality and computational speed compared to traditional methods. These results underscore its capability to handle the demands of dynamic and complex time-series clustering tasks with remarkable effectiveness.