Various types of sensor data can be collected by IoT. Each sensor node has spatial attributes and may also be associated with a large number of measurement data that evolve over time; therefore, these high-dimensional sensor data are inherently large scale. Detecting outliers in large-scale IoT sensor data is a challenging task. Most existing anomaly detection methods are based on a vector representation. However, large-scale IoT sensor data have characteristics that make tensor methods more efficient for extracting information. The vector-based methods can destroy original structural information and correlation within large-scale sensor data, resulting in the problem of the “curse of dimensionality,” and some outliers hence cannot be detected. In this paper, we propose a one-class support Tucker machine (OCSTuM) and an OCSTuM based on tensor Tucker factorization and a genetic algorithm called GA-OCSTuM. These methods extend one-class support vector machines to tensor space. OCSTuM and GA-OCSTuM are unsupervised anomaly detection approaches for big sensor data. They retain the structural information of data while improving the accuracy and efficiency of anomaly detection. The experimental evaluations on real datasets demonstrate that our proposed method improves the accuracy and efficiency of anomaly detection while retaining the intrinsic structure of big sensor data.

Prediction of chilling damage risk in maize growth period based on probabilistic neural network approach