A Hybrid Lightweight Model with Enhanced Interpretability for Surge Detection in Aero-Engine Compressors

Compressor surge detection, as a critical issue of aero-engines, often faces limitations with physics-based methods due to their reliance on expert-driven feature engineering and empirical thresholding rules. While deep learning offers superior feature extraction, it is hindered by low interpretability and high computational demands for real-time deployment. This paper aims to resolve these issues by proposing a novel framework that integrates a Lightweight Deep Support Vector Data Description (LWDSVDD) network with an Enhanced layer-wise relevance propagation (Enhanced-LRP) structure and dynamic threshold strategy. This new Enhanced-LIP algorithm identifies top-ranking temporal and spectral features by evaluating relevance score along with sensitivity for the surge phenomenon, ensuring precise and minimal extraction of critical features. Subsequently, these features are processed by the LWDSVDD which essentially embeds depthwise separable convolutions into traditional Deep Supported Vector Description, decomposing standard convolutions to depthwise and pointwise operations to achieve lightweight design. The joint effect of enhanced-LRP and LWDSVDD design enables a significant reduction in computational cost with minimal impact on detection accuracy. The proposed model was validated on several types of full-scale multi-stage compressors by deploying it in a portable edge-computing system, successfully demonstrating not only robustness to external interferences but also real-time surge warning capability with substantial lead time for reliable active anti-surge control.