Improving Predictive Maintenance in Bakery Digital Twins: A Hybrid Autoencoder-LSTM Framework with Adaptive Anomaly Thresholding
Year: 2025
Type: conference-paper
Abstract: The implementation of Digital Twin (DT) systems in food manufacturing is pivotal for realizing Predictive Maintenance (PdM) under Industry 4.0. However, the reliability of such systems is frequently compromised by harsh operational environments that accelerate sensor degradation (drift), leading to high false alarm rates. To address this, we propose a novel Hybrid Autoencoder-LSTM (HAE-LSTM) framework optimized for multivariate IIoT time series. Unlike static thresholding, our proposed Adaptive Anomaly Threshold (AAT) mechanism dynamically modulates detection sensitivity by correlating reconstruction error with a calculated degradation feature vector, effectively decoupling long-term sensor drift from genuine process anomalies. Experimental validation on a real-world dataset from a bakery enterprise demonstrates that the HAE-LSTM achieves an F1-score of 0.92, significantly outperforming baseline Isolation Forest (0.71) and SARIMA (0.66) models. Furthermore, the framework reduces false positive alarms by over 60%, enabling the Digital Twin’s optimization algorithms to secure a 24.4% reduction in production cycle time.
Institution: Al-Farabi Kazakh National University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7169914851