基于卷积神经网络的多工况多传感滚动轴承实时监控方法  

Real Time Status Monitoring Method for Rolling Bearing with Multiple Working Conditions and Multi-Sensor Based on Convolutional Neural Network

作　　者：陈昌川[1] 朱嘉琪 魏琦[3] 尹淑娟 乔飞[3] 赵超莹 CHEN Changchuan;ZHU Jiaqi;WEI Qi;YIN Shujuan;QIAO Fei;ZHAO Chaoying(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunication,Chongqing 400065,China;Baotou Teacher’s College,Inner Mongolia University of Science&Technology,Baotou Inner Mongolia 014030,China;Department of Electronic Engineering,Tsinghua University,Beijing 100084,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]内蒙古科技大学包头师范学院,内蒙古包头014030 [3]清华大学电子工程系,北京100084

出　　处：《传感技术学报》2024年第7期1162-1171,共10页Chinese Journal of Sensors and Actuators

基　　金：新疆维吾尔自治区重点研发计划项目(2022B01008-3);国家重点研发计划项目(2018YFB1702501);重庆市研究生教育教学改革研究一般项目(yjg213079);陕西省重点研发计划项目(2023-YBNY-222)。

摘　　要：针对工业环境中广泛在多工况下多滚动轴承实时状态监测的需求和部署环境受限的挑战,提出一种基于卷积神经网络(Convolutional Neural Network,CNN)的面向多传感器滚动轴承运行状态监控方法。该方法将两个不同工况下的一维时间序列数据集以均方根(Root Mean Square,RMS)指标标注,并通过将一维时间序列多传感器数据重构为二维空间张量的形式输入卷积神经网络训练。最后利用层融合和16比特量化优化,将网络部署到FPGA上,用以解决CNN的计算开销。实验结果表明,在结合了两种不同工况的数据集下,网络测试推理准确度依然高达99.24%,比多层感知机实现高10.48%,比多层感知机结合支持向量机的实现高2.91%,该算法对于新加入的数据集也有较强的鲁棒性,经过重训练,新加入的数据集准确率可以达到99.17%。基于FPGA部署优化的网络的峰值能效为76.217GPOS/W,为CPU实现的33.09倍,GPU实现的5.39倍。其中,16比特精度部署的网络测试精度相较32比特精度实现仅降低0.001%。Aiming at the challenge to monitor the status of multiple rolling bearings under multiple working conditions and constrained deployment environment,a status monitoring method for rolling bearings based on convolutional neural network with multi-sensors is proposed.Two one-dimensional time series datasets with multi-sensors under different working conditions are annotated based on RMS index,and constructed as two-dimensional spatial tensors.Then different networks are trained by using the reconstructed datasets respectively.Finally,the network is deployed on a FPGA with layer fusion and 16 bit quantization optimization methods to address the computation overhead.The result shows that the network achieves 99.24%test accuracy in the dataset that combined with two datasets with different working conditions,which is 10.48%higher than that of the multi-layer perceptron implementation and 2.91%higher than that of multi-layer perceptron and support vector machine.The algorithm also shows robustness for newly-add dataset,the accuracy of newlyadded dataset reaches 99.17%with retraining.The energy efficiency for the network with deployment optimization methods based on FPGA is 76.217GPOS/W,which is 33.09x that of CPU implementation and 5.39x that of GPU implementation.The test accuracy of 16 bit implementation is only 0.001%lower than that of the 32bit implementation.

关 键 词：滚动轴承 多工况 卷积神经网络 FPGA 部署优化 

分 类 号：TH133.3[机械工程—机械制造及自动化] TP183[自动化与计算机技术—控制理论与控制工程]