Abstract:
The article presents a new data augmentation method for measurement systems, designed for industrial equipment condition monitoring tasks. The relevance of the study stems from the significant limitations of traditional synthetic data generation methods, which fail to adequately reproduce complex non-stationary signals with characteristic transient processes, trends, and seasonal variations observed in real industrial environments. The proposed method integrates two advanced techniques: empirical wavelet transform (EWT) and conditional generative adversarial networks (Conditional GAN). The method is implemented in three stages: (1) adaptive decomposition of raw signals into modes using EWT, (2) mode categorization with label assignment, and (3) synthetic data generation using Conditional GAN. A set of statistical metrics was used to comprehensively assess the quality of synthesized signals, including Wasserstein distance (WS), Pearson correlation coefficient (PCC), and root mean square error (RMSE). Experimental studies were conducted on real-world temperature sensor data obtained under non-stationary industrial equipment conditions. The results demonstrate a significant advantage of the proposed method over the traditional TimeGAN approach: a 17% reduction in Wasserstein distance, a 57% increase in Pearson correlation coefficient, and a 21% decrease in RMSE. These findings confirm the method’s effectiveness in reproducing key characteristics of the original signals. The developed method enables the creation of synthetic datasets required for training modern neural network models in industrial equipment diagnostics. Its practical application significantly reduces the costs associated with experimental data collection while ensuring high-quality synthesized signals, as validated by statistical metrics.
Fulltext:
PDF file (2664 kB)