摘要
迁移学习的方法在解决齿轮箱无监督故障诊断问题上取得了极大的进展。然而,由于齿轮箱数据分布差异、噪声和干扰以及模型的局限性影响,大多方法在面对复杂的齿轮箱数据集迁移效果不佳,同时对于网络输入的可解释性研究仍然很少。提出了一种改进的域对抗网络(improve domain-adversarial neural network, IDANN)。首先,使用改进的时频网络作为特征提取器,在信号输入网络的时候提供可解释性和降噪功能;然后,在域对抗网络中添加目标域的类级对齐方法,使用两个分类器来检测靠近决策边界的目标样本,以增强迁移性能。在东南大学齿轮箱和跨座式单轨齿轮箱数据集上验证了IDANN的有效性和可靠性,并在凯斯西储大学轴承数据集上测试IDANN在噪声条件下的性能,试验证明IDANN具有优秀的诊断性能和鲁棒性。
Transfer learning method has made great progress in solving the problem of unsupervised fault diagnosis of gearboxes.However,due to differences of gearbox data distribution,noise,interference and limitations of model,most methods have poor migration performance when facing complex gearbox datasets,and there are still few studies on the explainability of network inputs.Here,an improved domain-adversarial neural network(IDANN)was proposed.Firstly,an improved time-frequency network was used as the feature extractor to provide explainability and denoising performance when signals were input into network.Then,the class level alignment method of target domain was added to IDANN,and two classifiers were used to detect target samples near decision boundary,and enhance transfer performance.The effectiveness and reliability of IDANN were verified on Southeast University gearbox and straddle-type monorail gearbox datasets,and IDANN performance under noise conditions was tested on the bearing dataset of Case Western Reserve University,USA.The test results showed that IDANN has excellent diagnostic performance and robustness.
作者
赵玲
邹杰
秦佳继
王航
ZHAO Ling;ZOU Jie;QIN Jiaji;WANG Hang(School of Information Science&Engineering,Chongqing Jiaotong University,Chongqing 400074,China)
出处
《振动与冲击》
北大核心
2025年第9期282-289,共8页
Journal of Vibration and Shock
基金
国家自然科学基金面上项目(6207023978)
重庆市教委重大科学技术研究项目(KJZD-M202400706)。
关键词
迁移学习
可解释网络
跨工况故障诊断
transfer learning
explainable network
cross-condition fault diagnosis
