همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

DiTA-RUL: Diffusion–Transformer Augmentation for CMAPSS

نویسندگان :

Yasamin Tafakor¹ Ali Jahan² Reza Tavakoli³ Siavash Ahmadi⁴ Babak Khalaj⁵

1- Computer Science Department, Amirkabir University of Technology 2- Computer Engineering Department, Amirkabir University of Technology 3- Mathematical Sciences Department, Sharif University of Technology 4- Electronics Research Institute, Sharif University of Technology 5- Electrical Engineering Department, Sharif University of Technology

کلمات کلیدی :

Remaining Useful Life (RUL)،Prognostics and Health Management (PHM)،Diffusion Model،Transformer،Data Augmentation،CMAPSS

چکیده :

Accurately estimating the Remaining Useful Life (RUL) is critical for improving the reliability and safety of industrial systems. The CMAPSS dataset, which provides run-to-failure engine trajectories, is widely used as a benchmark for RUL prediction. However, limited training samples and the challenge of capturing complex degradation patterns make effective data augmentation essential in this field. In this work, we propose a diffusion-based data augmentation framework that employs a latent structure incorporating wavelet decomposition, attention mechanisms, and sequence modeling. We then apply a diffusion model within this latent space to generate augmented sensor signals. In the backward phase of the diffusion model, a transformer-based network reconstructs realistic engine degradation signals using the learned latent representations. The generated data not only capture the statistical and temporal characteristics of true sensor signals but also improve RUL prediction compared to many existing approaches. We demonstrate the quality of the augmented data both by visual comparison with actual signals and by showing their effectiveness in boosting RUL prediction performance.