时频特征提取与差异性强化及其在变转速工况下行星齿轮箱模式识别中的应用研究

Extraction and Discriminability Enhancement of Time-frequency Features and its Application of Planetary Gearbox Pattern Recognition under Variable Speed

时变工况下行星齿轮箱的故障诊断仍然是一个热点问题。时频分析可以揭示非平稳信号的时变频率成分,但结果易受主观因素的影响。此外,不同类别的时频特征存在特征重叠现象,对智能诊断模型的性能产生不利影响。针对上述问题,提出了一种对抗自监督推理模型(Adversarial auto-supervised inference,AAsI),旨在实现时频特征自适应提取并强化不同类别特征间差异。首先,利用时频分析方法获得时频图像作为编码器输入。使用高斯混合分布构建可视化和分类良好的样本作为解码器的输入。之后,结合类别标签信息,探索一种对抗性机制,用于独立训练编码器和解码器,从而减少计算干扰,使提取的特征与分类良好的样本相似,同时以最小化均方误差为约束重构原始信号。最后,通过特征训练和测试Softmax分类器。该方法通过行星齿轮箱实验数据集进行了验证。结果表明,AAsI能够自适应提取时频特征并强化特征间差异,性能优于对抗学习推理、自动编码器和变分自动编码器。此外,在该数据集上基于不同时频分析方法AAsI准确率均高于98%。

It is still a topic issue to diagnose faults of planetary gearboxes under time-varying running conditions.Time-frequency analysis can reveal time-varying frequency components of nonstationary signals but easily be affected by subjective factors.Besides,overlapping features occur in the time-frequency features set,resulting in the adverse performance of an intelligent model.To address the above issues,an AAsI model is proposed,which aims at extracting the time-frequency features adaptively and enhancing their discriminability.Firstly,the time-frequency analysis method is utilized to obtain time-frequency images of signals as input for the encoder.A GMD is used to construct visualized and well-classified samples as input for the decoder.After that,an adversarial game is explored to train the encoder and decoder independently with label information,reducing computing interference and making the extracted features similar to the well-classified samples,and reconstructing the raw signals through the MSE constraint.Finally,a Softmax classifier is trained and tested by the features.This method is validated via a planetary gearbox data set.The results indicate that AAsl is valid for extracting the features adaptively and enhancing their discriminability,and outperforms ALI,AE,and VAE.Besides,the accuracies of AAsI based on different TFA are more than 98%over the data set.