【摘要】 针对航空活塞发动机振动信号中表征故障的关键特征易被噪声淹没，现有模型难以全面刻画复杂信号的挑战，提出了一种基于卷积神经网络(convolutional neural network, CNN)-Transformer交互融合网络的航空活塞发动机进排气故障诊断方法。首先，设计的CNN-Transformer双分支并行结构充分发挥各自优势，分别从原始振动信号中提取局部细节特征和全局时序特征。在此基础上，引入交叉注意力特征交互融合模块，通过注意力权重分配动态关联两类特征的关键信息，实现局部与全局特征的深度融合。最终，全面刻画出表征发动机进排气故障的振动特征，从而实现高精度的故障诊断。试验结果表明，该模型在不同工况下的三个数据集上均实现了99.40%以上的测试准确率，即使在噪声干扰下依旧保持了良好的诊断性能，与现有的诊断模型相比，具有更强的泛化性和鲁棒性。更多还原

【Abstract】 Here, aiming at challenges of key features characterizing faults in vibration signals of aviation piston engine being easily overwhelmed by noise and difficulty for existing models to fully characterize complex signals, a fault diagnosis method for intake and exhaust of aviation piston engine was proposed based on convolutional neural network(CNN)-Transformer interactive fusion network. Firstly, the designed CNN-Transformer dual-branch parallel structures could fully utilize their respective advantages to extract local detail features and global time sequence features from the original vibration signals. Then, a cross attention feature interaction fusion module was introduced to dynamically correlate key information of two types of features with assigning attention weights, and realize deep fusion of local and global features. Ultimately, vibration characteristics to characterize engine intake and exhaust faults were comprehensively characterized to realize high-precision fault diagnosis. The test results showed that the proposed model can realize a testing correctness rate of over 99.40% on 3 datasets under different working conditions, and keep good diagnosis performance even under noise interference; compared with existing diagnosis models, the proposed model has stronger generalization and robustness.更多还原