Gene recombination-guided convolution neural network for early fault diagnosis of aero-engines

Deep learning has been widely applied in fault diagnosis of the aero-engine, yet it faces significant challenges under complex operating conditions. These challenges include monitoring signals with extremely weak fault features and strong background noise, as well as high similarity in characteristics between normal and faulty states. Such scenarios trap traditional deep learning models in local optima due to their excessive reliance on error backpropagation during model training. Inspired by phenotypic trait measurement and gene recombination in biology, the gene recombination-guided convolution neural network is proposed in this paper for the precise diagnosis of subtle aero-engine faults. First, the signal phenotypic trait metric is introduced to quantitatively assess and sort the feature extraction ability of the convolution kernel. The high-performance and low-performance convolution kernels are distinguished. Second, similar to gene recombination, the weights of high-performance convolution kernels are randomly combined to generate new filial generation kernels. This process improves the ability to extract fault features and enhances the model’s robustness against noise. Finally, the proposed method is validated using the real-world monitoring data of the aero-engine. Experimental results demonstrate that it can efficiently extract weak fault features and achieve accurate diagnosis. In contrast experiments with mainstream and state-of-the-art methods, the proposed method achieves superior performance in diagnosing early faults.