<![CDATA[Maintaining vaccines within the 2°C–8°C range throughout cold chain storage and distribution is essential, since temperature excursions can degrade their biological potency. The growth of IoT-enabled sensors now allows continuous temperature data collection, opening the door to automated anomaly detection via time-series analysis. This research compares several unsupervised approaches for spotting temperature anomalies in an IoT-based pharmaceutical cold chain setup, benchmarking Isolation Forest against three deep learning architectures: Autoencoder, LSTM, and LSTM-Attention. Using roughly 8,640 temperature readings collected at 5-minute intervals over 30 days, the data were normalized with Min-Max Scaling and structured into sequences via a sliding window technique. Performance was assessed using precision, recall, and F1-score, alongside MAE and RMSE for prediction accuracy. Results showed Isolation Forest outperforming the other models (precision: 0.686, recall: 0.418, F1-score: 0.520) while also being the fastest to train and run. The deep learning models underperformed, likely limited by dataset size, making Isolation Forest the more practical choice for balancing detection accuracy with computational cost.]]>
Copyrights © 2026
