用脉冲神经网络在边缘设备上实现低功耗振动预测维护
Spiking Neural Networks for Low-Power Vibration-Based Predictive Maintenance
- 采用循环脉冲神经网络同时完成回归与多标签分类
- 故障识别准确率超97%，关键故障零漏报，流量速度误差低于1%
- 在神经形态芯片上能耗仅为传统CPU的千分之一，适合电池供电设备
工业物联网传感器的发展使得高时间分辨率的预测性维护成为可能。针对成本效益问题，基于振动的状态监测尤为值得关注。然而，传统云端处理高分辨率振动数据会带来显著的能耗和通信开销，限制了电池供电边缘设备的应用。因此需要将智能推向传感器端。由于脉冲神经网络（SNN）具有事件驱动特性，为低功耗本地计算提供了可行路径。本文研究了一种用于工业螺杆泵的循环SNN模型，利用三轴振动数据同时实现流量、压力、泵速的回归与正常、过压、气蚀等多类故障的分类。此外，我们对比了SNN在传统（x86、ARM）与神经形态（Loihi）硬件平台上的能耗。结果表明，分类准确率超过97%，关键故障无漏报；平滑后的回归输出在流量和泵速预测上均方相对误差低于1%，接近工业传感器标准，压力预测仍需优化。能耗估算显示显著节能：Loihi平台每推理消耗0.0032 J，比x86 CPU（11.3 J/inf）和ARM CPU（1.18 J/inf）低两个数量级。
原文摘要 · Abstract (English)
Advancements in Industrial Internet of Things (IIoT) sensors enable sophisticated Predictive Maintenance (PM) with high temporal resolution. For cost-efficient solutions, vibration-based condition monitoring is especially of interest. However, analyzing high-resolution vibration data via traditional cloud approaches incurs significant energy and communication costs, hindering battery-powered edge deployments. This necessitates shifting intelligence to the sensor edge. Due to their event-driven nature, Spiking Neural Networks (SNNs) offer a promising pathway toward energy-efficient on-device processing. This paper investigates a recurrent SNN for simultaneous regression (flow, pressure, pump speed) and multi-label classification (normal, overpressure, cavitation) for an industrial progressing cavity pump (PCP) using 3-axis vibration data. Furthermore, we provide energy consumption estimates comparing the SNN approach on conventional (x86, ARM) and neuromorphic (Loihi) hardware platforms. Results demonstrate high classification accuracy (>97%) with zero False Negative Rates for critical Overpressure and Cavitation faults. Smoothed regression outputs achieve Mean Relative Percentage Errors below 1% for flow and pump speed, approaching industrial sensor standards, although pressure prediction requires further refinement. Energy estimates indicate significant power savings, with the Loihi consumption (0.0032 J/inf) being up to 3 orders of magnitude less compared to the estimated x86 CPU (11.3 J/inf) and ARM CPU (1.18 J/inf) execution. Our findings underscore the potential of SNNs for multi-task PM directly on resource-constrained edge devices, enabling scalable and energy-efficient industrial monitoring solutions.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。