基于卷积的时态CBR快速检索算法及应用

探讨了如何增强CBR对一种常见的时态信息,即时间序列数据的检索能力;分析了已有的基于傅里叶频谱分析的时间序列检索算法应用于CBR时遇到的问题,并根据时态CBR检索的需要,提出了一种新的基于循环卷积和傅里叶变换时间序列检索算法.理论分析和数值实验结果都证明,提出的算法在检索效率上有一定的优势.将采取这种检索方法的时态CBR应用于时间序列的预测问题中,取得了较好的预测效果且具有较高的预测效率.

CBR algorithm supporting time series data

This paper focused on the retrieval algorithms of a special kind of CBR system in which cases were composed of time-series data. Introduced the classical algorithm used for processing similarity queries on time series data, This algorithm was based on the fact that DFT preserved the Euclidean distance in the time or frequency domain, and only the first few elements of the frequency sequence were significant, so the retrieval process could only use these significant elements to compute similarity degree. However, this algorithm had several disadvantages limiting its usage in CBR retrieval, so developed a new algorithm using batch method to compute the similarity degree. It was based on the observation that the original problem could be transformed to a convolution problem, and the circular convolution could be computed more efficiently using FFT. Theoretical analysis and experiment results prove that this algorithm is efficient and robust. The presented algorithm furnished the CBR with the ability to process cases consist of time-series data, developed a time series prediction algorithm based on CBR and the experiment results proved its efficiency.