Anomaly detection in industrial control systems: Leveraging adaptive multi-granularity anomaly correction for robust performance in noisy environments

Anomaly detection in industrial control systems (ICS) is critical for ensuring process reliability and protecting human life and property. Prediction-deviation-based models have attracted significant attention because they require minimal labeled anomaly data. However, prior studies have mainly focused on improving prediction accuracy, thereby increasing sensitivity to data deviations. These data deviation-based methods typically treat each time step as an individual detection unit, overlooking the interdependencies within the context of abnormal states. As a result, oscillating anomaly states are often produced, which significantly lower detection accuracy. Moreover, the uncertainty of industrial environments can introduce substantial noise or cause data shifts, posing additional challenges for anomaly detection. To tackle these challenges in complex noisy environments, we propose a multivariate time series anomaly detection method with adaptive multi-granularity anomaly correction (ATLAC). By incorporating an attention mechanism, ATLAC assigns weights to frequency-decomposed sub-series, emphasizing key frequency components while suppressing noise. We apply the TCN-LSTM architecture to model the reconstructed sub-series, using data deviation distributions to preliminarily detect anomalies. Furthermore, to account for the contextual influence of abnormal states, we adaptively set thresholds based on state information within multi-granularity windows, correcting frequent oscillations while eliminating short-term noise interference. Experimental results on both raw and noise-enhanced Tennessee Eastman datasets demonstrate that our method outperforms baseline methods in terms of detection accuracy across various scenarios, especially in noisy environments, where ATLAC significantly reduces the false positive rate by up to 13.62%. • Propose ATLAC for robust anomaly detection in noisy industrial environments. • Introduce frequency-based attention to enhance long-term noise suppression. • Design a TCN-LSTM model to capture temporal dependencies in sensor data. • Develop adaptive multi-granularity correction algorithms to address detection oscillations caused by short-term noise. • Achieve state-of-the-art performance on the Tennessee Eastman dataset and enhanced-datasets with added noise.