Class-imbalanced time series anomaly detection method based on cost-sensitive hybrid network

The time series is a structured data form. In practical engineering, most of the time series is in the form of classimbalance, in which, abnormal data is rare or underrepresented. In this study, we present a cost sensitive hybrid network (CSHN) model to detect data anomaly in class-imbalanced time series. The proposed model adopts a two-flow structure that consist of a Convolutional Neural Network (CNN) and a Gated Recurrent Unit (GRU) network and combine with a cost-sensitive loss function. This model integrates the advantages of CNN and GRU. The former is provided with strong learning ability for local state features and the latter can extract the sequential features and force the decision logic of the network to pay more attention to the long-term dependence of the time series. The cost-sensitive loss function is used to solve the problem that the detection accuracy of the minority class will be inaccuracy caused by skewed data distribution. The simulation experiments are conducted on the speed and temperature datasets that come from real-world measurement and control engineering, as well as the UCR datasets. Compared to using the cross-entropy loss function, the using of the cost-sensitive loss function can improve G-means and F-measure by 3 %-5% and 5 %-9%, respectively, on the speed and temperature datasets. Experimental results demonstrate that our method is efficient for the class-imbalanced data sets, and the numerical evaluation indexes are superior to that of the comparison methods.