Abstract:

Since one-dimensional feature vectors might not retain temporal feature information， but neural networks had good effects on image recognition， an image data set constructed by fault sound signals of centrifugal pumps was used to conduct centrifugal pump fault diagnosis. A Bayesian optimized multiscale DenseNet fault diagnosis method was proposed for centrifugal pump sound signals. One-dimensional time series acoustic signals were transformed into two-dimensional image through Gram angle field， and the time information and fault characteristics were preserved. Then multiscale dense blocks were used to extract image features to enhance image feature reuse. The dropout layer and L₂ regularization method were used to prevent overfitting， and Bayesian optimization algorithm was adopted to determine neural network hyperparameters. Finally， experimental verification was performed using centrifugal pump acoustic signals， and comparisons were made with other diagnostic methods. The results show that the Bayesian optimization multiscale DenseNet diagnosis model has a fault recognition rate of 99.5% for the test set.

Key words: centrifugal pump, fault diagnosis, Gram angle field, Bayesian optimization, multiscale DenseNet

摘要：

由于一维特征向量不能保留时间特征信息，而神经网络对图像识别具有良好效果，因此尝试用离心泵故障声信号构建的图像数据集开展离心泵故障诊断，提出贝叶斯优化多尺度DenseNet的离心泵声信号故障诊断方法。将一维时间序列声信号经过格拉姆角场转化为二维图像，保留其时间信息及故障特征；然后采用多尺度密集块对图像进行特征提取，增强图像特征复用；通过dropout层和L₂正则化方法防止过拟合，采用贝叶斯优化算法确定神经网络超参数，最后利用离心泵声信号进行实验验证，与其他诊断方法进行对比。结果表明，贝叶斯优化多尺度DenseNet的诊断模型对测试集具有99.5%的故障识别率。

关键词: 离心泵, 故障诊断, 格拉姆角场, 贝叶斯优化, 多尺度DenseNet