基于角域重采样和领域对抗网络的滚动轴承故障迁移诊断方法及实验分析

Fault transfer diagnosis method and experimental analysis of rolling bearings based on angular domain resampling and domain-adversarial networks

该文提出一种基于角域重采样和领域对抗神经网络的电机滚动轴承跨工况故障迁移诊断方法。首先对不同工况下的时域振动信号进行角域重采样,用以降低不同转速下振动信号的时频差异;然后利用领域对抗学习策略提取出源域与目标域数据中的领域不变特征,进一步减小不同工况间的数据分布差异。该文还搭建了电机滚动轴承故障诊断实验平台,针对6种跨工况迁移诊断任务开展了验证实验。实验结果表明,所提方法的故障平均迁移识别率高达95.08%。该故障诊断方法实验研究涉及信号处理、深度学习等领域知识,有助于学生掌握基本原理,锻炼理论联系实际的能力。

[Objective]Rolling bearings of motors are susceptible to failures owing to severe working environments and load fluctuations.A delay in dealing with this will lead to economic loss or even endanger personal safety.In recent years,deep learning has been broadly applied in rolling bearing fault diagnosis.However,traditional methods require training and test data to observe the same distribution,constraining their diagnostic ability under diverse operating conditions.To solve this problem,this paper presents a transfer learning approach that fuses angular domain resampling and domain-adversarial neural networks to lessen the distribution inconsistencies of data between dissimilar operating conditions and to achieve the cross-working-condition fault transfer diagnosis of rolling bearing faults.Moreover,the study of the proposed approach and the development of the associated experimental program aim to deepen students’understanding of signal processing and artificial intelligence theory and promote their learning enthusiasm.[Methods]First,the time–domain vibration signals and rotational speed pulse signals of the motor under diverse rotational speed conditions are synchronously gathered,and the rotational speed information is used to determine the rotational speed change curve of the motor.Next,the time–domain vibration signals are resampled in the angular domain according to this curve to acquire the angular domain vibration signals under diverse rotational speed conditions.This step aims to lessen the effect of rotational speed adjustment on the time–frequency properties of vibration signals and lessen the time–frequency variance of vibration signals under diverse rotational speed conditions.Then,the angular domain vibration signals under diverse rotational speed conditions are set as the source and target domains,respectively.The domain-invariant features in the source and target domain data are obtained using the domain-adversarial learning strategy,which reduces the data distribution variances between dissimilar rotational speed conditions.The domain-invariant properties of the labeled data in the source domain are utilized to train the classification network and enhance the fault classification capability of the network.This method facilitates satisfactory performance even in the case of unlabeled data in the target domain,enabling the cross-working-condition fault diagnosis of rolling bearings in electric motors.To confirm the effectiveness of the proposed method,an experimental platform for motor rolling bearing fault diagnosis is established,and six cross-working-condition transfer diagnosis tasks and four comparison experiments are designed.[Results]The experimental results reveal the following:1)The variances in the periodic shock intervals in the time–domain vibration signals under diverse rotational speed conditions are substantially decreased by angular-domain resampling.2)The fault identification accuracy of the proposed method is enhanced by over 20%compared with that of the ResNet method,which does not use the transfer learning strategy.3)Compared with the TFA method that uses the wavelet packet time–frequency analysis technique,the fault recognition rate is enhanced by more than 15%.4)The fault recognition rate of the proposed method is increased by over 10%compared with the DAN approach that uses the MMD kernel as the metric function.5)The transfer recognition accuracy of the proposed method exceeds 93%in each transfer diagnosis task.[Conclusions]The time–frequency variance of vibration signals under diverse rotational speed conditions is remarkably reduced by resampling the time–domain vibration signals in the angular domain.It is combined with the domain-adversarial neural network,which further reduces the variance of data distribution among diverse working conditions.According to the above enhancements,the mean fault transfer recognition rate of the proposed approach reaches 95.08%,which accomplishes the expected objective.Moreover,the research of the whole fault diagnosis approach requires the knowledge of multiple disciplines,such as signal processing and deep learning.The application of the proposed method and experimental program to teaching will aid in broadening students’horizons,stimulate their interest in learning,and provide a teaching case for the development of interdisciplinary and integrated talents.