Bridging The Synthetic-To-Real Gap: A Model-Data Coevolution Approach With Stochastic Feature Decoupling For Ac Unit Fault Diagnosis

Mars: Jurnal Teknik Mesin, Industri, Elektro dan Ilmu Komputer
Vol. 3 No. 6 (2025): Desember: Mars: Jurnal Teknik Mesin, Industri, Elektro Dan Ilmu Komputer

The scarcity of real-world data in Air-Conditioning (AC) fault diagnosis necessitates the use of synthetic data; however, rule-based synthetic datasets often suffer from a significant sim-to-real domain gap. To address this, we propose a Model-Data Coevolution (MDC) framework that employs a Simulated Annealing (SA) controller to optimize augmentation parameters. We introduce a novel technique, Stochastic Feature Decoupling (SFD), which applies independent noise to raw and derived features, contrasting it with traditional Logically-Consistent Augmentation (LCA). Empirical results show that SFD significantly outperforms LCA, achieving a weighted F1-score of 0.93 and increasing NORMAL class recall to 82%. We demonstrate that by breaking deterministic feature links, SFD acts as a robust regularizer, utilizing "physically impossible" data to enhance generalization in complex real-world environments.