噪声标签下注意力特征混合的旋转机械故障诊断

Fault diagnosis of rotating machinery with attentive feature mixup in noisy labels

实际工程中,受人为标记或数据预处理等原因影响,旋转机械故障数据集易出现噪声标签,导致故障诊断模型性能降低,故提出注意力特征混合的旋转机械故障诊断方法。首先,构建残差神经网络(ResNet)提取样本中的时频特征,通过随机分组和特征交互构建正确标签样本组、部分噪声标签样本组和噪声标签样本组;其次,引入注意力机制计算各样本组内样本相关性对各组样本分配权值,得到能区分部分噪声样本组中噪声标签样本的差异性权值;然后,根据权值对每组样本进行混合(Mixup),通过对噪声标签样本插值并在反向传播中更新注意力层参数降低噪声标签样本所占比例;最后,利用在线标签平滑(OLS)统计模型预测信息更新软标签,通过降低噪声标签样本对模型损失更新的影响,进一步抑制噪声标签样本组的负面影响。在不同程度的噪声标签干扰下的旋转机械故障数据集上进行实验验证,检测精度均达到95%以上,证明了所提方法的有效性。

In practical engineering,the rotating machinery fault data set is prone to noise labels due to human labeling or data preprocessing,which could lead to the degradation of fault diagnosis model performance.Therefore,the attentive feature Mixup method of rotating machinery fault diagnosis is proposed.First,a residual neural network(ResNet)is established to extract time-frequency features from the samples.The correct label sample groups,partially noisy label sample groups,and noisy label sample groups are constructed through random grouping and feature interaction.Secondly,an attention mechanism is introduced to calculate the correlation between samples within each sample group,and assign weights to each group of samples.The different weights that can distinguish noisy label samples within partially noisy sample groups are achieved.Then,Mixup is performed on each group of samples according to their weights,which can interpolate noisy label samples and update the attention layer parameters during backpropagation to reduce the proportion of noisy label samples.Finally,the online label smoothing(OLS)is used to update the model′s prediction information by reducing the negative impact of noisy label samples on the model loss update and further suppressing the effects of noisy label sample groups.Experiments on the rotating machinery fault dataset with label noise interference show the effectiveness of the proposed method.