Research on CUDA-based Multidimensional Time Series Motif Mining
Jimin Wang, Shuaiwei Wang, Junru Wang · 2024
Traditional serial motif mining methods struggle to quickly identify motif information in large-scale time series data. A CUDA-based multidimensional motif mining algorithm is proposed to discover motifs in multidimensional time series. The algorithm introduces a multidimensional kNN Matrix Profile (kNN mMP) structure to address the limitations of the multidimensional mMP. Parallel computation of kNN mMP using the CUDA architecture enables rapid motif discovery in massive datasets and facilitates detection of additional motif features, such as strong and weak motifs, as well as identification of anomalous motifs. Experiments on public datasets demonstrate that using CUDA for kNN mMP structure achieves a significant speedup; on NVIDIA GeForce RTX 4060, the kNN-mSTOMP-GPU algorithm exhibits a speedup of over 220 times compared to the STOMP algorithm utilizing only kNN mMP when the sequence length is 200,000.