Multi-scale metric learning neural network with adaptive graphs for few-shot gas path fault diagnosis of variable cycle engine

Variable Cycle Engine (VCE) operates over a broad flight envelope. Traditional data-driven fault diagnosis models struggle to achieve high accuracy over the entire envelope, since gas path fault samples are scarce in some flight conditions. Besides, an engine to be diagnosed may suffer from data scarcity, since fault data cannot be shared across engines with significant performance differences. Moreover, VCE is a strong nonlinear system that aggravates the difficulty of extracting fault features from gas path parameters. To address these problems, a Multi-Scale Metric Learning Neural Network with Adaptive Graphs (AG-MSML) is proposed. AG-MSML employs an adaptive graph module based on VCE space structure to extract multi-scale features containing spatial-physical and global-local information. Meanwhile, this module utilizes the graph attention mechanism to model parameter coupling variations caused by variable-geometry actuators, enhancing AG-MSML’s adaptability to different working modes. Besides, the Variable-Activation-Function (VAF) of the learnable metric module is variable according to multi-scale features, allowing for more effective capture of hidden information in the data. Therefore, the VAF-based metric learner can better model multi-scale relationships and improve its capability to distinguish between different classes. Few-shot task simulations demonstrate that AG-MSML has better gas path fault diagnosis accuracy and robustness than other networks.