Résumé: With the advent of 5 G, the Internet of Things (IoT) makes many basic applications possible. In IoT applications, sensors are widely deployed to collect various data. Due to the dynamic nature of the wireless channel and the harsh environment, anomalous data are often generated and the trustworthiness of the data is compromised. For the large amount of data collected by the underlying sensor network, traditional outlier detection methods often fail to reach a satisfactory points in terms of latency and accuracy, affecting the decision-making of the IoT systems. In this paper, we propose a data anomaly detection framework for edge-cloud based on iterative optimization. The model adopts a hierarchical structure, where anomaly detection is performed by edge nodes between the remote cloud and the wireless sensor networks, and online machine learning performs the training and updating of the model in the cloud. Principal component score difference is developed to obtain training data. The detection model is constructed by Multi-Kernel Support Vector Data Description(MKSVDD). The cloud updates the detection model through the iterative optimization process of principal component score difference and MKSVDD. The cluster-based data analysis framework reduces the computational and processing burden of sensor nodes. Extensive experiments on real datasets show that the data anomaly detection framework based on iterative optimization improves the accuracy of anomaly detection while maintaining the low latency of industrial sensor-cloud systems.

Langue: Anglais