High-precision anomaly detection, as the key technology of magnetic flux leakage (MFL) signal detection, is a challenging task. It is difficult to detect anomalies in MFL signals due to the variety of anomalies and the characteristics of the anomalies are easily submerged in the variation of the natural signals. To address the above issues, a feature-based unsupervised detection network (FUDet) is designed, which accomplishes the unsupervised anomaly detection task through feature discrimination and feature reconstruction. Firstly, a bidirectional discrimination module is proposed, which can input normal and anomaly feature distributions to mine the characteristics of samples, so as to enhance the ability of the model to recognize anomaly signals. Secondly, a dynamic noise generation module is designed to generate different feature distributions for each input that are consistent with the characteristics of MFL signals. This module creates an adversarial effect with the discriminator, allowing it to identify more subtle feature differences through training. Finally, a reconstruction classification module is designed to naturally reconstruct the non-normal features and normal features into normal signals, which can be used to detect anomalies by comparing the difference between the input signals and the reconstructed signals. Experimentally, the method is proved to outperform the P-AUROC of the state-of-the-art method by 3.1% under MFL signals and achieves outstanding results in MFL signal anomaly detection.

中文翻译：

---

一种基于特征的无监督检测网络的漏磁信号异常检测新方法


高精度异常检测作为漏磁（MFL）信号检测的关键技术是一项具有挑战性的任务。由于MFL信号的异常种类繁多，且异常的特征很容易淹没在自然信号的变化中，因此检测MFL信号中的异常非常困难。针对上述问题，设计了一种基于特征的无监督检测网络（FUDet），通过特征判别和特征重构来完成无监督异常检测任务。首先提出了双向判别模块，可以输入正态和异常特征分布来挖掘样本特征，从而增强模型识别异常信号的能力。其次，设计了动态噪声生成模块，为每个输入生成与MFL信号特征一致的不同特征分布。该模块与鉴别器产生对抗效应，使其能够通过训练识别更细微的特征差异。最后，设计了重建分类模块，将非正常特征和正常特征自然地重建为正常信号，通过比较输入信号和重建信号之间的差异来检测异常。实验证明，该方法在MFL信号下比最先进方法的P-AUROC提高了3.1%，在MFL信号异常检测中取得了优异的结果。