【摘要】 传统的分布式活塞式航空发动机故障诊断方法需要人工参与，诊断过程不连贯，而基于一维卷积神经网络(1D-CNN)的故障诊断存在易过拟合、提取特征信息尺度单一的现象．针对以上问题，设计了一个结合卷积注意力机制的改进残差多尺度卷积块和一个多尺度特征融合分类层．基于此，构建了残差多尺度卷积神经网络(RMSCNN)用于活塞式航空发动机点火与喷油系统的故障诊断．首先，通过改进残差多尺度卷积块提取原始时域数据的多尺度特征信息；然后对并行输出的网络结构进行降维、融合，形成多尺度特征融合层；最后，使用Softmax进行分类识别．结果表明：对于包含了点火提前角异常、喷油量异常共16类的活塞式航空发动机故障数据集，RMSCNN能够实现91.44%的故障诊断准确率，相比1D-CNN、GoogLeNet和ResNet50等常见网络，具有更好的故障诊断结果．更多还原

【Abstract】 Fault diagnosis methods for traditional distributed aero piston engines require manual involvement,and the diagnostic process is also discontinuous. Fault diagnosis based on one-dimensional convolutional neural networks(1D-CNN) tends to overfit easily and exhibits a single-scale feature extraction issue. In response to the above issues,a modified residual multi-scale convolution block combined with a convolutional attention mechanism and a multi-scale feature fusion classification layer were designed. Based on this,a residual multi-scale convolutional neural network(RMSCNN) was constructed for fault diagnosis of the ignition and fuel injection systems of an aero piston engine. Initially,multi-scale feature information from raw time-domain data was extracted by improving the residual multi-scale convolution block. Subsequently,the network structure with parallel outputs was dimensionally reduced and fused to form a multi-scale feature fusion layer. Finally,Softmax was used for classification recognition. Experimental results show that for a dataset of 16 fault categories of the aero piston engine,including abnormal ignition advance angle and abnormal fuel injection quantity,RMSCNN achieves a fault diagnosis accuracy of 91.44%. Compared to common networks such as 1D-CNN,Goog Le Net,and ResNet50,it demonstrates superior fault diagnosis performance.更多还原