AI-Driven Fault Detection and Predictive Maintenance in Green Electronic Systems
Аннотация
AI-driven fault detection and predictive maintenance are increasingly applied in green electronic systems; however, existing approaches often process heterogeneous sensor data independently, lack physical consistency, and incur high computational and energy costs, limiting their reliability and deployability. The solution of these challenges is extremely important in enhancing the efficiency of the system, minimizing downtime and facilitating sustainable operation within the environment that has limitations on energy. This study aimed to develop a precise, energy-saving, and privacy-conscious multi-modal learning architecture of fault detection and remaining useful life (RUL) predicting in green electronic systems. An experimental study was created based on publicly available heterogeneous datasets, which included timeseries signals, thermal/RGB images, acoustic data, and operational metadata. Multimodal Physics-aware Graph Transformer (MPGT) was developed that incorporates crossmodal attention, system topology modelling as well as physicsbased constraints and compared its performance with baseline 1D-CNN physical model and hybrid CNN-LSTM + PINN models. Evaluation included fault classification, RUL prediction, uncertainty calibration, inference latency, energy consumption, Pareto analysis, ablation studies, and confusion matrix analysis. The proposed MPGT demonstrated better fault detection performance (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{F 1}$</tex>-score <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\boldsymbol{=} \mathbf{0. 9 1}$</tex>, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{R O C}$</tex>-AUC <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\boldsymbol{=} \mathbf{0. 9 6}$</tex>) and much lower RUL prediction error (MAE <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$=11.4$</tex> cycles) than the baseline (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{F 1} \boldsymbol{=} \mathbf{0. 8 1}$</tex>, MAE <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\boldsymbol{=} \mathbf{2 4. 5}$</tex> cycles) and hybrid models (F1 <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$=0.86$</tex>, MAE <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$=18.0$</tex> cycles). The proposed MPGT demonstrated a favorable accuracy-energy trade-off, operating at <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3 8 ~ m s}$</tex> latency and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1. 1 ~ J}$</tex> per inference, while confusion matrix and ablation analyses proved that MPGT had better fault separability and multimodal, physics-conscious factors. The results indicate that physics-conscious multimodal graph-based learning is able to significantly improve the performance of predictive maintenance with less energy use. This framework provides a practical and scalable foundation for reliable, sustainable, and privacy-conscious maintenance of green electronic systems.