MFCD:A Deep Learning Method with Fuzzy Clustering for Time Series Anomaly Detection

The problem of unsupervised anomaly detection in time series is very challenging. In recent years, deep learning based methods have been widely used. However, existing methods still struggle to effectively detect certain specific types of anomalies such as collective anomalies. Therefore, in this paper, we propose a novel method based on reconstruction. We apply fuzzy clustering to the deep features of time series and calculate the sum of distances between sample points and all cluster centers, and then train the model by combining the reconstruction error. In addition, an anomaly criterion applicable to fuzzy clustering called Fuzzy Anomaly Distance (FAD) is devised to further amplify the difference between anomalies and normal points. We named the model MFCD, and its average anomaly detection F1 score on 7 datasets (including 4 real-world applications) is 96.07%, which is significantly better than previous state-of-the-art methods.