S-WhiteSVDD: A feature fusion approach for diesel engine performance degradation assessment using Multi-Source impulse signals

Abstract

Performance degradation assessment (PDA) is a critical component of predictive health management (PHM). The mixed multi-source impulse characteristics of diesel engine vibration signals make PDA more challenging compared to rotating machinery. To address the unique characteristics of diesel engine signals, this study proposes a Subspace-Whitening Support Vector Data Description (S-WhiteSVDD) feature fusion approach that combines knowledge-based features with deep learning features. The method tracks cross-cycle variations of multiple homologous impulses and constructs a feature subspace for each impulse. Whitening transformation ensures balanced stretching and compression of subspace data across all components, preventing features with large variances from dominating the decision boundary. This approach aligns more closely with the data manifold and enables precise control of anomaly boundaries. To overcome the challenge of significant health indicator (HI) fluctuations that hinder early fault detection, the method integrates the interpretability of knowledge-based features with the complex mapping capabilities of deep features. This fusion enhances the richness of feature representation, facilitating the detection of early fault onset. The effectiveness and superiority of the proposed method are demonstrated through both valve degradation simulations and nozzle degradation engineering case studies. The constructed HI effectively indicates component degradation. The proposed approach shows strong potential for practical engineering applications.