Detection against time series of cyber physical systems(CPS) is an important means of anomaly detection. How-ever, some existing time series anomaly detection methods often ignore the dependencies within the time series, making the predicted or reconstructed data establish poor dependencies, which in turn affects the anomaly detection performance. To address the above problems, this paper proposed a time series anomaly detection method for cyber physical systems via adaptive interactive learning and unscented Kalman filter. The method used neural networks to identify the hidden states of CPS, and then preserved the dependencies of the time series through global adaptive fusion and interactive learning. Moreover, the me-thod used the traceless Kalman filter to track the trend of the time series to enhance the robustness of the prediction process. Finally, the method evaluated anomalies by calculating anomaly scores. The average performance obtained by applying this method on three CPS datasets is 0.940 for F₁ score, 0.965 for precision, and 91.7% for recall. The experimental results show that compared with other research methods in recent years, the proposed method can better preserve the time series dependencies, extract more accurate time series features, and thus improve the prediction performance of the model, resulting in better recall and F₁ scores of anomaly detection, which has good application prospects.

[2023-05-19] Accepted Paper /doc/2023.03.0095/v1

[2023-10-05] Printed Article /doc/2023.03.0095/v2