A hybrid deep learning and data mining approach for interpretable anomaly detection in high-dimensional time-series data

Abstract

With the continuous growth of industrial data systems, vast high-dimensional time-series data are created from sensors and networks. This research introduces a hybrid deep learning and data mining framework for accurate, interpretable, and efficient anomaly detection in industrial cyber physical systems. The Industrial Sensor Anomaly Detection Dataset, comprising SWaT and WADI benchmark datasets, is employed to evaluate model performance. Data preprocessing involves Z-score normalization and sliding window segmentation to standardize sensor readings and preserve temporal dependencies. Independent component analysis (ICA) is used for feature extraction to isolate statistically independent patterns, reducing redundancy while enhancing informative components. The proposed Quokka Swarm Optimized Efficient Deep Neural Network (QSO-EDNN) integrates QSO for optimal features selection and parameter tuning and EDNN that accurately detects anomalies from learned temporal patterns. In parallel, data mining algorithms, including Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for grouping anomalous events and data mining for discovering frequent patterns, are integrated to enhance the interpretability of detection results. Experimental evaluation reveals the robustness of proposed approach across both datasets. On the SWaT dataset, the model achieved high detection metrics with precision, recall, and F1-score 87%, while on the WADI dataset, performance ranged around 95–96%. Moreover, the result demonstrates the QSO-EDNN model’s superior performance, scalability, and reduced false alarm rate compared to traditional techniques. Overall, the proposed hybrid model offers reliable, interpretable, and computationally effective solution for real-time anomaly detection in industrial environments.