Anomaly detection in time-series data is crucial for identifying abnormal behavior in domains such as healthcare, finance, and cybersecurity. This study develops a context-aware framework utilizing three deep-learning architectures—LSTM, Auto-encoder, and a Hybrid LSTM–Auto-encoder—combined with fixed, rolling, and context-adaptive dynamic thresholding. The framework jointly evaluates reconstruction deviation, temporal sequence deviation, and supervised probabilistic deviation to understand how different anomaly signals behave under varying threshold regimes. Experimental results show that while LSTM provides strong temporal modeling and the Autoencoder captures structural irregularities, their limitations are reduced when fused into a supervised hybrid classifier. The Hybrid-2 architecture delivers the most balanced and deployment-ready performance, demonstrating high precision, tolerance to operational noise, and stability across grid-search runs. The training process is formulated as an optimization problem where model parameters are learned through minimization of reconstruction and prediction losses under regularized gradient-based updates. Overall, the study offers practical guidance for selecting suitable architectures in real-world telemetry systems where anomalies vary in scale, duration, and contextual behavior.
Akaasha Asif Asif (Mon,) studied this question.