Performance evaluation method for modules based on organic fusion of data-driven methods and mechanistic knowledge

Accurate performance evaluation for aeroengine modules is crucial for performance recovery and guiding repair levels. A module performance evaluation method that integrates data-driven methods and mechanistic knowledge is proposed. During the training process, mechanism knowledge is involved in the training of data-driven models, achieving an organic integration of the two, thereby improving the reliability and accuracy of performance evaluation. For the underdetermined problem of the correlation equations, module factors and non-module factors are integrated into model training, while the solution space is reduced through pretraining and joint training strategies. At the same time, the incorporation of mechanistic knowledge to constrain the feasible domain further reduces the solution space. For the purpose of improving the saliency of module efficiency features, dilated convolution is adopted to mine deep features, and a feature sharpening module is designed to improve the discriminability of module efficiency features. Finally, the proposed method is validated using real aeroengine data. The experiments show that the proposed method reconstructs gas path parameter deviation values accurately. The method also evaluates module performance very well.