基于条件归一化流的并行时空注意力的无监督异常检测模型

Conditional Normalization Flow-Based Unsupervised Anomaly Detection Model with Parallel Spatio-Temporal Attention

摘要: 预测性维护作为提升工业生产效率的重要环节，其核心挑战在于准确识别故障并发出早期预警。在实践中，工业生产中记录的数据具有两大特征：数据规模庞大与标签缺失。因此需建立针对多元时间序列的无监督异常检测模型。本文提出一种新型无监督异常检测模型——基于条件归一化流的并行时空注意力无监督异常检测(PAF)模型。该模型基于“异常数据远少于正常数据”的假设，结合高斯分布理论，将异常数据定位于低概率密度区域。基于此假设，采用条件归一化流变换数据分布并进行密度估计，进而设定异常评分实现有效判别。该模型的核心特征在于运用并行注意力结构捕捉原始数据的时空信息，作为条件归一化流的前置信息条件。相较于现有相关研究，PAF展现出更优异的检测性能。最终，我们通过两个公开数据集验证了PAF模型的先进性和有效性。

Abstract: Predictive maintenance, as a crucial component for improving industrial production efficiency, faces a core challenge in accurately identifying faults and issuing early warnings. In practice, the data recorded in industrial production have two main characteristics: multivariate data and missing labels. Therefore, there is an urgent need to establish unsupervised anomaly detection models for multivariate time series. This paper proposes a novel unsupervised anomaly detection model—the Parallel Spatiotemporal Attention Flow-based (PAF) model. This model is based on the assumption that “anomalous data are far less frequent than normal data” and, combined with Gaussian distribution theory, positions anomalous data in low-probability density regions. Based on this assumption, conditional normalizing flows are used to transform the data distribution and perform density estimation, thereby setting anomaly scores to achieve effective detection. The core feature of this model is the use of a parallel attention structure to capture the spatiotemporal conditions of the original data as prior information for the normalizing flow. Compared with existing related studies, PAF demonstrates superior detection performance. Finally, the effectiveness and advancement of the PAF model are validated using two publicly available datasets.