Bingjie Li, Chen Huang, Teng Zhou, Wenqiang Zhang, Lei Zhang, Ben Zhao
We present an image-based approach for early detection of surge precursors in a centrifugal compressor. One-dimensional pressure signals are mapped to Gramian angular field images, from which Gray Level Co-occurrence Matrix correlation is computed as the primary indicator using fixed windowing and gray-level quantization. On a single test rig at 60 000 and 70 000 rpm, the correlation trace rises prior to measured surge onset and follows the approach–surge–recovery evolution (e.g., a surge interval near 19.41–19.51 s at 60 000 rpm), enabling a baseline-referenced threshold for monitoring. To stabilize short-window fluctuations, a lightweight exponentially weighted moving average (α = 0.1) produces smoother yet responsive trajectories without resorting to black-box training. In the stable operating map, correlation exhibits a V-shaped variation with decreasing mass flow—coherent at overload, weakest near peak efficiency, and strengthening again as surge is approached—providing a physically interpretable trend. Within this scope, the study is a compressor-focused proof-of-concept; extension to other unsteady-flow problems is plausible but will require problem-specific tuning and independent validation.