Cascaded TinyML-Based Reduction for the Anomaly Detection Model of an Industrial Combustion System
Iman Sharifirad, Jalil Boudjadar, Peter Gorm Larsen · 2025
Machine learning (ML) based anomaly detection has been thoroughly studied in the literature where different architectures and models have been developed for different industrial cyber-physical systems (CPS). However, given that most CPS control systems are deployed on embedded platforms with limited storage and computational resources, deployment and real-time functionality remain challenging. This paper applies a combination of TinyML techniques, Pruning, and quantization to reduce the computation cost and memory footprint of a transformer-based anomaly detection for an industrial combustion system without degrading the accuracy and classification performance. To assess reliability, we compare the reduced model's performance and accuracy to the original model and the state-of-the-art Cumulative Sum Control Chart (CUSUM) analytical model. The experiment results demonstrate that, while reducing the memory footprint by 70% and computation cost by 8%, the resulting model does not degrade the classification accuracy. Moreover, the reduced transformer model outperforms the CUSUM classification accuracy by up to 99%.