51 / 2025-05-13 20:29:16
Anomaly Detection in Scenarios with Missing Abnormal Samples Based on Generative Flow Model and Generative Adversarial Network
Abnormal samples missing,anomaly detection,Generative Flow model,Generative Adversarial Networks
全文待审
鑫杰 龚 / 苏州大学
俊 王 / 苏州大学
双 李 / 苏州大学
伟国 黄 / 苏州大学
长青 沈 / 苏州大学
忠奎 朱 / 苏州大学
In condition monitoring of critical transportation components such as bearings, abnormal examples are extremely scarce or entirely absent, rendering conventional supervised diagnostic techniques impractical. To overcome this challenge, we propose an unsupervised anomaly detection framework that fuses Generative Flow model with Generative Adversarial Network, which is termed Glow-GAN. First, the Glow model performs invertible flow transformations and maximum likelihood estimation on extensive time-frequency matrices derived exclusively from data under healthy operation, thereby accurately modeling the healthy data distribution. A dedicated discriminator is then introduced for adversarial training, in which binary cross-entropy loss guides the model to better capture fine-grained data details. During inference, an anomaly score is computed by combining the reconstruction error with the negative log-likelihood; significant deviations from the healthy baseline trigger a fault alarm. Experimental evaluation demonstrates that the proposed Glow-GAN framework effectively separates healthy and anomalous data, enabling early bearing-fault detection. Compared to existing methods or pure GAN architectures, the proposed approach leverages the rigorous probabilistic foundations of flow models alongside the detail-enhancing benefits of adversarial training, yielding stable convergence and superior condition monitoring.
重要日期
  • 会议日期

    08月01日

    2025

    08月04日

    2025

  • 06月15日 2025

    初稿截稿日期

主办单位
中国机械工程学会设备智能运维分会
承办单位
新疆大学
移动端
在手机上打开
小程序
打开微信小程序
客服
扫码或点此咨询