TITLE
A HYBRID LSTM-AUTOENCODER-ISOLATION FOREST DEEP LEARNING MODEL FOR SMART GAS LEAKAGE ALERT SYSTEM
AUTHOR(S)
Mustafa Kemal TEZCAN*, Umit KURSUN
ABSTRACT
This study proposes a new enhanced hybrid deep learning model, combining LSTM and Autoencoder architectures with a non-parametric Isolation Forest (IF) filter, to address the high false alarm issues of traditional systems. The model's anomaly threshold was determined using the statistically 99th percentile (0.0649). The system's robustness was validated using a noisy synthetic time-series dataset. Experimental results establish the foundational superiority of the Hybrid Model: while the single Autoencoder and LSTM benchmarks detected 187 and 189 anomalies, respectively, the standalone Hybrid Model successfully filtered over 280% of that noise, detecting only 49 anomalies. The subsequent integration of the Isolation Forest filter, trained on a multi- dimensional error vector, provided a final, critical safety layer, further reducing the count to 47. This confirms the reliability of the hybrid approach and introduces a robust, multi-layered validation system, paving the way for future implementation on local IoT hardware for real-world validation.
DOI
How to cite this article:
Mustafa Kemal TEZCAN*, Umit KURSUN, A HYBRID LSTM-AUTOENCODER-ISOLATION FOREST DEEP LEARNING MODEL FOR SMART GAS LEAKAGE ALERT SYSTEM, UNITECH – SELECTED PAPERS - 2025