An Unsupervised Anomaly Detection Scheme based on Memory Mechanism in Cloud-edge Collaboration Scenario

Abstract

In order to timely reject defective products in high-speed assembly lines, the task of anomaly detection is widely used in industrial production process. Recently, unsupervised anomaly detection has made tremendous developments as more scholars have started to focus on this field. The model detects abnormal samples whose data distribution deviate from the normal data distribution. However, due to the problem of data island in machine learning, it is difficult to achieve global optimization for models trained on local datasets. In this work, a memory network based algorithm for cloud-edge collaborative anomaly detection is proposed to break down data barrier. Our proposed method obtains an anomaly detection model using local datasets trained at each edge side, and the local model updates the parameters in the cloud to obtain the global model, which decoupling the data retrieval and the training of the global model. In addition, due to the inconsistent data distribution, we design the memory module to retain the data features from each edge in order to further improve the detection capability of the global model. The memory module from each edge is iteratively updated in the cloud with the training of the model, so that our final global anomaly detection model retains the information about data distribution features from different domains of different edges.