Multi-combination fault data augmentation method of aero-engine gas path based on Extraction TimeGAN

Abstract

Listen

Translate article

Aero-engine is the core power component of aircraft, and the accurate processing of engine monitoring data is the key to ensure its reliable operation and prevent faults. In the use and maintenance of aero-engines, the multi-combination fault data acquisition of gas path system is expensive and unavailable, thus intelligent multi-combination fault diagnosis under small samples scenarios poses a significant challenge. Aero-engine gas path combination fault data exhibit strong coupling, making it difficult for data augmentation methods to produce high-quality data. To solve the above problem, a novel aero-engine gas path combination fault data augmentation method based on Extraction TimeGAN is proposed. First, original data of multi-combination faults were obtained from Gasturb. To increase the complexity of generating data, Gaussian white noise of Signal-to-Noise Ratio (SNR) 20 was injected into the original data. Then, new samples with similar characteristics were generated from the original data according to different fault categories by TimeGAN, and the original data and generated data were both mapped to the latent space by Principal Component Analysis (PCA). Finally, the optimized extraction mechanism of generated data was established, and the high-density original data was identified and divided in the latent space through Kernel Density Estimation (KDE), the high-density regional hypersphere was constructed by Support Vector Data Description (SVDD), and the generated data located inside the hypersphere is extracted as the optimized generated data. In the multi-combination fault diagnosis experiment of aero-engine gas path, Area Under the Curve (AUC) of Support Vector Machine (SVM), Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) reaches 99.74%, 99.82% and 99.85%. Under high noise conditions, the AUC of Extraction TimeGAN only decreases 0.94%. Furthermore, Under the condition of small sample, Extraction TimeGAN maintains the AUC of over 99%.