变负载轴承故障诊断卷积神经网络模型

Fault Diagnosis Model for Variable Load Bearing in Convolutional Neural Network

针对轴承故障在实际环境中存在负载变化导致故障难以诊断的问题,提出一种基于一维卷积神经网络的变负载适应轴承故障诊断模型,卷积结构使用小卷积核卷积层堆叠的形式,训练时对输入层进行均匀分布失活率的随机失活,以提高网络的变负载适应能力,且采用全局平均池化降低模型计算量和减轻过拟合程度;在实验验证阶段,提出以两种近邻负载条件的轴承数据构成变负载数据集,充分验证轴承故障诊断的变负载适应性。实验结果表明:本文模型在各测试集均能达到96%以上的准确率且平均准确率达到98.36%,能够在变负载环境下实现准确、稳定的轴承故障诊断,具有良好的泛化能力。

Aiming at the problem that bearing faults are difficult to diagnose due to load changes in real case,a variable load adaptive bearing fault diagnosis model was thus proposed based on one-dimensional convolutional neural network.The convolution structure used a small convolution kernel with a stack of convolution layers.During training,the input layer was subjected to random deactivation of the uniform distribution rate to improve the variable load adaptability of the model,and used global average pooling to reduce model calculation and overfitting.In the experimental verification stage,it was proposed to use the bearing data of two neighboring load conditions to form the variable load data set,which could fully verify the variable load adaptability of bearing fault diagnosis.The experimental results show that the model achieved more than 96%accuracy in each test set,and the average accuracy rate was 98.36%.Therefore,the model could realize accurate and stable bearing fault diagnosis under variable load environment in a good generalization ability.