Anomaly detection algorithm of IoT devices based on semi-supervised heterogeneous ensemble learning

This paper proposes an abnormal behavior detection method based on semi-supervised heterogeneous ensemble learning to detect the abnormal behavior of IoT devices and solve the problems of data heterogeneity, weak generalization ability, and high label cost-efficiently and accurately. Firstly, the equipment is divided into three types: periodic type, trend type and random type according to the time series decomposition method and correlation coefficient. Secondly, two kinds of direct detection methods based on probability statistics and similarity measurement are adopted for the random equipment and indirect detection methods based on prediction or reconstruction before diagnosis are used for periodic and trend equipment. Finally, the optimal sub-model of each class is selected, and the heterogeneous integration model based on accuracy weighting is designed to make the anomaly detection more robust and accurate. The experimental results show that the accuracy of the LSTMs-AE + 3sigma sub-model is the highest, up to 81.09%, compared with many detection combination algorithms. The accuracy rate of the heterogeneous integration model obtained by selecting the optimal sub-models from different types of algorithms is as high as 89.25%, which is 8.16% higher than the individual sub-model. The false positive rate and false negative rate also decreased, only 5.05% and 2.08%.