Data Reconstruction via Consensus Graph Learning for Effective Anomaly Detection in Industrial IoT

In the industrial Internet of Things (IIoT), anomaly detection is fundamental to ensuring system safety and product quality, among other things. However, the massive amount of unlabeled, high-dimensional data generated in the IIoT challenges some existing anomaly detection methods as follows: 1) the distance concentration caused by the high dimensionality of the data can cause a decrease in the detection accuracy of some of the methods; and 2) some of the methods fail to explore the intrinsic relationships between the data, resulting in less effective detection of anomalies in the data. To handle the above challenges, a framework named data reconstruction via consensus graph learning (DRCG) and two anomaly score functions are proposed. Specifically, DRCG overcomes the distance concentration problem and explores the intrinsic relationships of the data by integrating projection learning, low-dimensional embedding, and consensus graph learning into a unified objective function. Then, an iterative algorithm is designed to solve the DRCG model. By doing so, DRCG not only drives the reconstruction error of abnormal samples higher than that of normal samples, but also obtains the projection that can effectively extract the intrinsic relationship between the data. Furthermore, to identify anomalies in the data, two anomaly score functions based on the reconstruction error and projection are designed, respectively. To achieve online anomaly detection for streaming data, DRCG with the projection-based anomaly score function is extended into an online version. The effectiveness and superiority of the proposed methods have been demonstrated on four real-world datasets.