Early warning method for compressor aerodynamic instability using phase-locked averaging filtering and moment function neural network

Diagnosis and early warning of compressor instability in aero engines are among the current research hotspots and challenges in the field of aero engines. To address the issues, such as capturing non-integer order frequency components in dynamic pressure signals, limited feature dimensionality, complex stall mechanisms, and the difficulty in quantifying evolution trajectories of compressor instability, a compressor aerodynamic instability early warning method is proposed based on phase-locked averaging filter and moment function neural network, using a multi-stage high-speed compressor as the research object. The method first employs phase-locked averaging filter to extract non-integer order frequency disturbance features under high-load conditions. Subsequently, an instability early warning model based on moment function neural network is constructed, which utilizes moment functions to capture global statistical features of instability and local detail features of asymmetric separation and intermittent pulses in early-stage weak signals. Next, Box-Cox transformation is introduced to eliminate heterogeneity among higher-order moment features, and multi-layer perceptron network layers are adopted to achieve early warning and diagnosis of compressor instability. Finally, the effectiveness of the proposed method is validated through test data at different rotational speeds. Results demonstrate that the method accurately characterizes the evolutionary laws of higher-order moment feature spaces, enabling efficient visualization, identification, and separation of instability precursors and steady-state data. Compared with the traditional surge detection method on the compressor rig, it can provide instability warning up to 4.8 s in advance.