Blade fault is a catastrophic failure of gas turbines. In order to improve the reliability of blade during operation, condition monitoring is one of the effective methods. However, there always exist two problems with blade monitoring: 1) challenges in warning of the occurrence of blade failure in advance and 2) difficulties finding the location after blade failure. In this article, we solve these problems by excavating the characteristics of blade-related signals in broadband casing vibration with advanced signal processing methods and machine learning technology. First, a novel method–Sparse Harmonic Product Spectrum (SHPS)–is proposed to accurately calculate blade passing frequency from gas turbine broadband casing vibration. The SPHS relies on Fourier transform, and its calculation utilizes the physical relationship between fundamental frequency and blade passing frequencies. Combining Vold-Kalman filter with adaptive parameter optimization process (AVKF), the blade-related vibration can be separated from casing vibration even in strong noise. Analysis of simulated casing vibration signal is used to verify the effectiveness and superiority of proposed method. Based on blade-related vibration, we build a gas turbine blade condition model in an unsupervised learning manner. The model can excavate potential blade failures earlier and more accurately than conventional threshold methods. Then, three coefficients are constructed according to the blade-related vibration characteristics to identify the blade fault location in multi-stage system. Moreover, the effectiveness of the proposed blade diagnosis framework is verified using actual industrial gas turbine blade fracturing failure data.
|Number of pages
|Measurement: Journal of the International Measurement Confederation
|Early online date
|29 Mar 2023
|Published - 15 Jun 2023