一种在线时间序列预测的核自适应滤波器向量处理器

A Kernel Adaptive Filter Vector Processor for Online Time Series Prediction

针对信息物理融合系统中的在线时间序列预测问题,该文选择计算复杂度低且具有自适应特点的核自适应滤波器(Kernel Adaptive Filter,KAF)方法与FPGA计算系统相结合,提出一种基于FPGA的KAF向量处理器解决思路。通过多路并行、多级流水线技术提高了处理器的计算速度,降低了功耗和计算延迟,并采用微码编程提高了设计的通用性和可扩展性。该文基于该向量处理器实现了经典的KAF方法,实验表明,在满足计算精度要求的前提下,该向量处理器与CPU相比,最高可获得22倍计算速度提升,功耗降为1/139,计算延迟降为1/26。

To address the online time series prediction problem in CPS（Cyber-Physical System） system, both KAF（Kernel Adaptive Filter） with low computation complexity and adaptive characteristic and FPGA computing system are employed. A novel FPGA implementation of vector processor targeting KAF algorithm is proposed. The parallelized datapath and multi-stage pipeline are introduced to enhance the performance and reduce the power consumption and latency. The microcoding technology is further employed to improve the reusability and extensibility. The classical KAF algorithms are implemented based on the vector processor. Experiments results show that the proposed vector processor improves the execution speed by factors of 22, the power consumption decrease to 1/139, while the latency decrease to 1/26 compared with a CPU, on the condition that the precision meets the requirement.