基于MDFF与ISSA的滚动轴承故障声发射诊断

AE Diagnosis of rolling bearing faults based on MDFF and ISSA

针对滚动轴承早期、复合故障难以准确诊断与智能诊断模型超参数确定严重依赖专家先验知识问题,提出一种基于多维深度特征融合(multi-dimensional depth feature fusion, MDFF)与改进麻雀搜索算法(improved sparrow search algorithm, ISSA)的滚动轴承故障声发射诊断方法。用一维卷积与线性瓶颈反向残差二维卷积神经网络构建多输入卷积神经网络(convolution neural network, CNN)结构的诊断模型,模型输入为滚动轴承声发射信号及其小波时频图,提出基于布伦纳梯度和信噪比的质量指标,在108种小波基中筛选出最佳时频图以提升输入数据质量。接着,采用特征金字塔网络将模型的一、二维低层与高层特征融合,建立深度融合的诊断模型。然后,将交叉混沌映射、自适应权重及融合的随机游走策略引入麻雀搜索算法中,以自适应获取MDFFCNN最优超参数。试验表明,对比近期多个主流智能诊断算法,所提方法可避免人工选择诊断模型超参数,对滚动轴承早期尤其复合故障具有更高的诊断精度和稳定性,模型诊断过程的智能化水平得到了进一步提高。

Here,to solve problems of diagnosing early and compound faults of rolling bearing being difficult and determining super-parameters of intelligent diagnosis model being heavily dependent upon expert prior knowledge,an acoustic emission(AE)diagnosis method for rolling bearing faults based on multi-dimensional depth feature fusion(MDFF)and improved sparrow search algorithm(ISSA)was proposed.Firstly,a diagnostic model of multi-input convolution neural network(CNN)structure was constructed by using ID and 2D CNN with linear botdeneck reverse residual error.The model input was AE signals of rolling bearing and their wavelet time-frequency diagram.A quality index based on Brenner gradient and signal-to-noise ratio was proposed,and the optimal time-frequency diagram was selected from 108 wavelet bases to improve the quality of input data.Then,the feature pyramid network was used to fuse the ID and 2D low-level features and high-level ones of the model to establish a diagnostic model of deep fusion.Furthermore,cross chaotic mapping,adaptive weighting and fused random walk strategy were introduced into SSA to adaptively obtain MDFFCNN optimal super-parameters.Tests showed that compared with several current mainstream intelligent diagnosis algorithms,the proposed method can avoid manual selection of super-parameters of the diagnosis model;it has higher diagnostic accuracy and stability for early rolling bearing faults,especially for compound faults;the intelligent level of the model diagnosis process is further improved.