MGAD: Mutual Information and Graph Embedding Based Anomaly Detection in Multivariate Time Series

Electronics, Journal Year: 2024, Volume and Issue: 13(7), P. 1326 - 1326

Published: April 1, 2024

Along with the popularity of mobile Internet and smart applications, more high-dimensional sensor data have appeared, these hidden information about system performance degradation, failure, etc., how to mine them obtain such is a very difficult problem. This challenge can be solved by anomaly detection techniques, which an important field research in mining, especially domains network security, credit card fraud detection, industrial fault identification, etc. However, there are many difficulties multivariate time-series data, including poor accuracy, fast generation, lack labeled capture between sensors. To address issues, we present mutual graph embedding based algorithm time series, called MGAD (mutual detection). The consists four steps: (1) Embedding where heterogeneous become different vectors same vector space; (2) Constructing relationship sensors using their each other; (3) Learning attention mechanism, predict at next moment; (4) Compare predicted values real detect potential outliers. Our contributions as follows: propose unsupervised outlier high interpretability accuracy; massive experiments on benchmark datasets demonstrated superior algorithm, compared state-of-the-art baselines terms ROC, F1, AP.

Language: Английский

Research on Outlier Detection Methods for Dam Monitoring Data Based on Post-Data Classification

Buildings, Journal Year: 2024, Volume and Issue: 14(9), P. 2758 - 2758

Published: Sept. 3, 2024

Safety monitoring of hydraulic structures is a critical task in the field engineering construction. This study developed method for preprocessing and classifying data identification gross errors structures. By utilizing linear regression wavelet analysis techniques, it effectively differentiated various waveform characteristics sets, such as Sinusoidal Wave Cyclical, Triangular Seasonal Weakly Cyclical growth types. In experiments error identification, 3σ algorithm, K-medoids Isolation Forest algorithm were applied to test data. The results showed that excelled processing Data Sets; adapted better performed well handling sets with significant anomalies or atypical fluctuations scenarios strong seasonality large fluctuations; complex Growth Sets, all three algorithms less effective, indicating potential need more advanced methods combination multiple techniques. Testing on actual further confirmed importance using specific techniques special types after set pre-classification, providing effective technical solution safety

Language: Английский

Combining normalizing flows with decision trees for interpretable unsupervised outlier detection

Engineering Applications of Artificial Intelligence, Journal Year: 2024, Volume and Issue: 141, P. 109770 - 109770

Published: Dec. 6, 2024

Language: Английский