학술논문
基于软阈值降噪的脉冲卷积神经网络轴承故障诊断方法 / Bearing fault diagnosis method based on soft threshold denoising for spiking convolutional neural network
Document Type
Academic Journal
Author
Source
电气技术 / Electrical Engineering. 25(2):12-20
Subject
Language
Chinese
ISSN
1673-3800
Abstract
针对工业场景下滚动轴承信号易受噪声干扰,导致故障诊断准确率低和稳定性差的问题,本文提出一种基于软阈值降噪的脉冲卷积神经网络诊断方法.该方法使用软阈值滤波去噪,运用带时间标签的卷积层处理二维信号,增强动态特征提取能力.同时,通过引入IF和LIF神经元实现对时域和频域信息的联合编码,并采用替代梯度法进行端到端训练.实验结果显示,在信噪比为 6dB时,所提方法的诊断准确率达 100%,在信噪比为-6dB时诊断准确率达 77.33%,优于其他常用方法,表明所提方法在噪声下具有良好的诊断效果和稳定性.
The signals of rolling bearings are easily interfered by noise in industrial environments,which reduces fault diagnosis accuracy and worsens stability.This paper proposes a diagnostic method based on soft threshold denoising for spiking convolutional neural network.Soft threshold filtering for noise reduction is proposed in this paper.This paper uses time-tagged convolutional layers to process two-dimensional signals to enhance dynamic feature extraction capabilities.IF and LIF neurons are introduced to jointly encode time domain and frequency domain information,and the surrogate gradient method is used for end-to-end training.The results show that the diagnostic accuracy reaches 100%under the signal-to-noise ratio of 6dB,and still reaches 77.33%under the signal-to-noise ratio of-6dB.The results of this method have certain advantages compared with commonly used methods,which verifies that the proposed method has better diagnostic results and higher stability under noise.
The signals of rolling bearings are easily interfered by noise in industrial environments,which reduces fault diagnosis accuracy and worsens stability.This paper proposes a diagnostic method based on soft threshold denoising for spiking convolutional neural network.Soft threshold filtering for noise reduction is proposed in this paper.This paper uses time-tagged convolutional layers to process two-dimensional signals to enhance dynamic feature extraction capabilities.IF and LIF neurons are introduced to jointly encode time domain and frequency domain information,and the surrogate gradient method is used for end-to-end training.The results show that the diagnostic accuracy reaches 100%under the signal-to-noise ratio of 6dB,and still reaches 77.33%under the signal-to-noise ratio of-6dB.The results of this method have certain advantages compared with commonly used methods,which verifies that the proposed method has better diagnostic results and higher stability under noise.