Multi-scale LSTM forecasting with attention-based feature fusion for anomaly detection in multivariate monitoring data

Engineering systems such as power equipment and industrial processes are monitored through multivariate time series whose anomalies appear at different temporal scales, from short fluctuations to deviations that persist over a whole operating cycle. A detector that reads a single input window has to commit to one of these scales. This paper presents MAF-LSTM, a forecasting-based anomaly detection framework in which parallel long short-term memory (LSTM) branches read input windows of different lengths, a scale-level attention module fuses their representations into a one-step-ahead forecast, and the forecasting residuals are scored with the Mahalanobis distance estimated on normal validation data. Experiments on smart meter data, two sets of high-voltage converter modulator pulses and a simulated chemical process show that MAF-LSTM achieves the highest AU-ROC and AU-PR of the five compared methods on all four datasets, exceeding the best baseline by 3.20 and 5.33 percentage points on average. An ablation study on the smart meter data quantifies the contribution of each component and shows that multi-scale modeling is most beneficial for anomalies with pronounced within-day fluctuation, where it matches or exceeds the best single-window model on each anomaly type without requiring the window length to be tuned.