HIPER-CHAD: Hybrid Integrated Prediction-Error Reconstruction-Based Anomaly Detection for Multivariate Indoor...

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Detecting subtle anomalies in noisy multivariate indoor environmental time-series data is challenging.
• Existing methods struggle to robustly distinguish between normal fluctuations and true anomalies.

Purpose of the Study:

• Introduce the Hybrid Integrated Prediction-Error Reconstruction-based Anomaly Detection (HIPER-CHAD) model.
• Improve anomaly detection by separating temporal modeling of normal behavior from probabilistic modeling of prediction uncertainty.
• Develop an anomaly score robust to noise yet sensitive to genuine abnormalities.

Main Methods:

• Utilize a Long Short-Term Memory (LSTM) network for time-series forecasting.
• Employ a Variational Autoencoder (VAE) trained on residual errors to model prediction uncertainty.
• Combine VAE reconstruction error and KL divergence for a statistically grounded anomaly score.

Main Results:

• HIPER-CHAD achieved an F1-score of 0.8571 on a real-world dataset, surpassing the LSTM Autoencoder (0.8095).
• The model maintained perfect recall, indicating no missed anomalies.
• A 20-step window optimized the F1-score to 0.884, demonstrating sensitivity analysis effectiveness.

Conclusions:

• The HIPER-CHAD model offers a reliable and accurate framework for anomaly detection in complex multivariate time-series data.
• The hybrid approach effectively handles noisy data and subtle anomalies.
• This method provides a statistically grounded anomaly score, enhancing detection capabilities.