自适应注意力LSTM-ResNet下的滚动轴承故障诊断

Fault diagnosis of rolling bearing under adaptive attention LSTM-ResNet

滚动轴承信号具有复杂性和非线性的特点,对特征提取和故障分类带来挑战。为解决上述问题,文章提出一种自适应注意力LSTM-Resnet(long short-term memory residual network)下的滚动轴承故障诊断方法。首先设计双向LSTM组特征提取模型,获取复杂运行条件下的滚动轴承特征;然后,提出自适应注意力LSTM-ResNet完成特征学习,并自适应调整模型中关键特征的权重;最后,采用全局平均池化(global average pooling,GAP)方法结合Softmax模型缓解模型过拟合并完成故障分类。在数据集中完成滚动轴承故障分类,实验结果表明:文章方法的滚动轴承故障诊断准确率相对于SVD-ResNet方法和宽卷积模型更高,并且能在标记样本数量较少和噪声环境下均达到较高的检测准确率,具有更高的准确性和更强的鲁棒性。

The complexity and nonlinearity of rolling bearing signals pose challenges to feature extraction and fault classification.To solve these problems,in this paper,an adaptive attention long short-term memory residual network(LSTM-ResNet)is proposed.The bidirectional LSTM group feature extraction model is designed to obtain the rolling bearing features under complex operating conditions.Then,an adaptive attention LSTM-ResNet is proposed to complete feature learning and adjust the weights of key features in the model.Finally,global average pooling(GAP)method is used to combine softmax model to alleviate overfitting and merge the model to complete fault classification.The rolling bearing fault classification was completed in the data set.The experimental results show that the proposed method has a higher detection accuracy than SVD-ResNet method and wide convolutional model,and can achieve a higher detection accuracy under the condition of less marked samples and noise.with higher accuracy and stronger robustness.