LSTM-GAN-XGBOOST Based Anomaly Detection Algorithm for Time Series Data

• Published in: 2020 11th International Conference on Prognostics and System Health Management (PHM-2020 Jinan) pp. 334 - 339
• Main Authors: Xu, Xin, Zhao, Hongbo, Liu, Haoqiang, Sun, Hua
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.10.2020
• Subjects: Anomaly detection; Feature extraction; Gallium nitride; GAN; Generative adversarial networks; Generators; LSTM; PHM; Time series analysis; time series data; Training; XGBOOST
• ISSN: 2166-5656
• DOI: 10.1109/PHM-Jinan48558.2020.00066

Time series anomaly detection is an important and fundamental task of Prognostic and Health Management (PHM). Traditional anomaly detection algorithms can achieve the detection of shallow features when dealing with the nonstationary time series data, yet those algorithms fail to detect outliers on deep features of massive time series data. In this paper, we proposed a novel fusion algorithm named LSTM-GAN-XGBOOST based on the characteristics of artificial neural networks (ANNs) and ensemble learning (EL). The hybrid approach combines long short-term memory network (LSTM) to extract time dimensional features of time series data, generative adversarial networks (GAN) to extract deep features of normal data effectively, and extreme gradient boosting (XGBOOST) to classify the extracted features and export anomaly scores. Moreover, we proposed an anomaly detection framework of test and evaluation based on LSTM-GAN-XGBOOST to obtain the final anomaly results and evaluation indicators. The proposed algorithm shows obvious advantages in processing features extraction and anomaly detection of time series data. Experimental and classic ball bearing time series datasets have been used to testify the effectiveness of the proposed approach and its superiority over some conventional methods. The experimental results demonstrate that LSTM-GAN-XGBOOST can effectively detect the anomalies of ball bearing time series dataset, and achieves 99.1% in terms of area under ROC curve (AUC) which is a superior performance compared with conventional algorithms, and has high significance for time series anomaly detection.