Hardware Accelerator for Shapelet Distance Computation in Time-Series Classification

Time-series classification has several important real-world applications and shapelet-based methods have emerged as highly attractive tools for this task. They are appropriate to search for time-series subsequences with high discriminative power among classes. Although these algorithms are accurate and interpretable, the task of measuring local shape similarity results in heavy computational burdens, which may limit their applicability. In this paper we address this issue by proposing a hardware accelerator to compute both Z-Score normalization and Euclidean distance. We identify these tasks as hot spots in shapelet-based TSC and propose scalable and parameterizable hardware that is suitable as a dedicated shapelet-distance engine. Results show that the proposed hardware significantly reduces the run time of the shapelet distance computation. The speedup factor increases with the shapelet length, reaching speedups of more than 5 times when compared to a software execution for shapelets with length larger than 100.