Enhancing Axial Compressor Stall Warning with Time-Frequency Generative Adversarial Networks: A Novel Approach in Advanced Signal Processing

• Main Authors: Qiu, Xiaohong, Wu, Jiawei, Guo, Chaochao, Wu, Xin
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Adversarial training; Axial compressor; Fault monitoring; Feature extraction; Generative adversarial networks; Predictive models; Time series analysis; Time-frequency analysis; Training
• ISSN: 1948-9447
• DOI: 10.1109/CCDC62350.2024.10588069

In the domain of axial compressor stall warnings, existing methods largely focus on time or frequency domain analyses, leading to excessive dependence on signal quality, increased rates of false alarms, and merely partial extraction of characteristics signaling an imminent stall. To overcome these challenges, it is imperative to develop a model that can swiftly and effectively identify early anomalies that signal an impending stall within a complex time series. Such a model ideally enables a thorough extraction of early stall indicators. This paper introduces the Time-Frequency Generative Adversarial Networks (TFGAN), representing the inaugural innovative fusion of adversarial training with time-frequency analysis. Utilizing the Wigner-Ville distribution function, the TFGAN achieves enhanced fidelity in information exchange between the time and frequency domains. Through adversarial training, the TFGAN discerns the characteristic states of various sequences following time-frequency analysis, precisely identifying conditions that suggest pre-stall precursors, thus fulfilling the goal of accurate stall warnings. Empirical results show that the TFGAN model effectively captures the complex time-frequency relationships present in signal data, facilitating the identification of subtle temporal changes across different features. This enables the swift extraction of stall characteristics, anomaly detection, and stall prediction within complex time series data. When compared with a range of generative models, the proposed method markedly outperforms experimental benchmarks in generative accuracy and predictive capability, as demonstrated by both qualitative and quantitative analyses.