High-efficient implementation of neural networks and ANFIS using direct bitstream processing

Abstract

Intelligent information processing technologies are already widely used in the power industry. The new step in their evolution was made with application of Internet-of-Things(IoT) devices into a smart grid. A key problem of intelligent IoT device development is implementation of a data processing algorithms on a low-power on-board processing units. The paper proposes a high efficient technology for neural network implementation on a FPGA-based IoT devices. An approach based on the use of sigma-delta modulated bitstreams for signal representation is proposed, which makes it possible to reduce the size of all elements of neural network and to provide the possibility of completely parallel implementation of neural networks and adaptive network-based fuzzy inference systems (ANFIS) on the basis of existing FPGA microchips. Based on the proposed approach, the novel FPGA IP cores are developed for the implementation of direct propagation neural networks, RBF-networks and ANFIS. The FPGA resource usage efficiency of the developed solutions is demonstrated using synthesis results on modern FPGA. The proposed approach will increase performance of IoT devices, designed for smart-grid.