【摘要】 针对现有方法在处理训练样本较少的数据集时易出现过拟合现象的问题，将图卷积神经网络引入柴油机故障诊断领域，并结合量子粒子群优化算法，建立一种基于QPSO-MC-GCN(Quantum Particle Swarm Optimization-Multi-channel-Graph Convolutional Network)的故障诊断方法。该方法搭建了一种邻接矩阵，将时序振动数据转换为图数据，实现多个测点样本特征的有效融合；利用QPSO对多通道图卷积神经网络(MC-GCN)的关键参数学习率和热核函数宽度进行寻优，以提高模型的泛化能力；在传统图卷积神经网络(GCN)的基础上建立双头权值矩阵以提取更丰富的深层特征，并引入一维最大池化层进一步控制过拟合现象。对实测柴油机振动信号的分析结果表明，该方法针对试验所设定故障类型的诊断准确率优于文中的对比方法，尤其是在低标签比的情况下优势更明显。更多还原

【Abstract】 Here, aiming at the problem of existing methods being easy to have over-fitting when processing data with fewer training samples, the graph convolution neural network(GCN)was introduced into the field of diesel engine fault diagnosis.Combined with the quantum particle swarm optimization(QPSO) algorithm, a fault diagnosis method based on QPSO-multi-channel-GCN(QPSO-MC-GCN) was established. In this method, an adjacency matrix was built to convert timeseries vibration data into graph data, and realize effective fusion of sample features of multiple measuring points.QPSO was used to optimize the learning rate of key parameters and the width of thermal kernel function of MC-GCN, and improve the generalization ability of the model.Based on the traditional GCN, a double-headed weight matrix was established to extract richer deep features, and a 1-D maximum pool layerwas introduced to further control over-fitting phenomenon. The analysis results of measured diesel engine vibration signals showed that the diagnosis accuracy of QPSO-MC-GCN method for fault types set in testsis superior to QPSO method’s and GCN method’s, especially, in the case of low tag ratio, QPSO-MC-GCN method’s advantage is more obvious.更多还原