A Hybrid Statistical–Transformer Autoencoder Framework for Point and Contextual Anomaly Detection in Multivariate Operational Time-Series Data

Anomaly detection in a multivariate operational time series data is a critical requirement for intelligent information systems to support reliable monitoring and decision making in the modern cloud and data center environments. However, current statistical methods are prone to overlook contextual oddities, whereas the deep learning reconstruction models may be threshold-instable and have false alarm problems. To overcome these drawbacks, in this work, the hybrid approach to the detection of anomalies based on the processing of statistical deviations (Z-score and interquartile range) and the Transformer Autoencoder for the reconstruction of the context of the sequence is proposed. The proposed approach combines statistical evidence of anomaly and contextual reconstruction error in deep by weighted fusion approach in detection of point anomalies and contextual anomalies effectively. Experiments were performed on the Server Machine Dataset (SMD) which is a collection of telemetry data from 28 heterogeneous server machines with 38 operating features each. The results show that the hybrid model is better than statistical-only model and Transformer-only model with Mean Precision = 0.295581, Recall = 0.456487, F1-score = 0.289387 and ROC-AUC = 0.882706. These results confirm the advantages of combining the power of statistical reasoning with Transformer-based contextual modelling in order to improve the robustness of anomaly detection and increase the decision reliability in operational monitoring systems.