Anomaly Detection of Small Time Series Data Based on Improved Generative Adversarial Networks

Abstract

Small sample time series anomaly detection, as an important part of time series research, has attracted extensive attention and research in both academia and industry. The superior performance of deep learning for time series anomaly detection is largely due to the large number of training samples. However, the problem of difficult data collection leading to inaccurate modelling is common in practice. The solution to the problem of small-sample problem of anomaly detection on time series data, this paper proposes a small-sample time series data anomaly detection method ADGAN based on adversarial learning, which firstly uses the generative adversarial network as the basic framework with different network structures, in which the generative network integrates the TCN and the self-attention mechanism to achieve better data reconstruction results, and then the single-layer LSTM is used as the discriminative mechanism. The single-layer LSTM is used as the discriminative network, and the model can effectively detect anomalies in small-sample time series data through the improved GAN network structure. The experimental results on the NAB dataset show that this method has certain advantages in improving the detection accuracy and efficiency.