Research on Prediction and Discrimination Method of Compressor Internal Stall Level Based on Time Series Feature Extraction

The accurate identification of internal stall onset in ultra-high load compressors is critical to ensuring the reliability and performance efficiency of next-generation aero-propulsion systems. Aligning with the thematic focus of Frontiers in Computer Science on intelligent modeling, predictive analytics, and complex systems computation, this work introduces a novel framework that leverages high-dimensional time-series feature extraction to address long-standing challenges in early stall detection. Traditional methods, primarily rooted in global stability analysis and empirical thresholds, often fall short in sensitivity and interpretability, particularly under transient or localized flow disturbances. They tend to overlook weakly amplified precursors due to insufficient resolution or overly generalized models, thereby compromising early warning capabilities. To overcome these limitations, we propose an integrated strategy consisting of the Selective Modal Growth Characterization Network (SMGCN) and the Incremental Margin Mapping Protocol (IMMP). SMGCN is a physics-informed neural architecture that decodes spatial and temporal growth patterns of unstable flow modes using sparse modal activation and orthogonal decomposition, ensuring both physical relevance and model compactness. IMMP then transforms these extracted features into actionable stall indicators through spatio-temporal risk localization, ferential phase tracking, and multi-resolution decomposition, culminating in accurate predictions of both stall initiation time and its spatial origin. Experimental validation demonstrates the framework’s superior capability to identify incipient instability with minimal data latency, offering robust, interpretable predictions compatible with real-time engine monitoring systems. By embedding domain-specific knowledge into deep learning pipelines, this method exemplifies a convergence of computational intelligence and aerothermal diagnostics, reinforcing the scope of the journal’s emphasis on cyber-physical systems, predictive modeling, and high-dimensional signal interpretation.