An energy-efficient spiking neural network accelerator on FPGA for real-time health monitoring of rotating machinery

Abstract

Real-time health monitoring of rotating machinery is critical for ensuring the safe operation of industrial systems. However, intelligent diagnostic algorithms with high computational complexity typically rely on high-performance computing platforms, limiting their deployment on resource-constrained edge devices and resulting in high diagnostic latency and energy consumption. Spiking neural networks (SNNs) benefit from their event-driven computational mechanism and exhibit potential energy-efficiency advantages in long-term health monitoring. Therefore, this article proposes an efficient diagnostic framework based on a field-programmable gate array (FPGA)-accelerated SNN. To enhance feature representation from dynamic measurement signals, a hybrid learning strategy for SNNs integrating local spike-timing-dependent plasticity and global backpropagation through time is introduced. At the hardware architecture level, a fully parallel Poisson encoding array is designed to achieve single-cycle spike conversion of sensing signals. Furthermore, a distributed addition engine based on near-memory computing and a neuron computation unit developed with shift logic are introduced to minimize off-chip memory access and eliminate the use of digital signal processor blocks. Experimental results on the Zynq-7020 platform demonstrate that the proposed accelerator achieves a power consumption of 175 mW, outperforming existing lightweight convolutional neural network-FPGA acceleration schemes. The diagnosis latency for each signal sample is 0.209 ms, which is 3.1× lower than that of GPU-based implementations. In addition, on our independently collected fault dataset, the performance degradation in terms of F1-score, precision, and accuracy is maintained within 0.3%, demonstrating the reliability of the proposed framework for practical health monitoring deployment.