【摘要】 针对传统油润滑船舶艉轴承磨损故障智能诊断中存在的振动信号实测困难、训练数据不足等问题导致模型过拟合、性能下降，以及训练网络深度增加引发的梯度消失从而影响诊断准确率等问题，提出一种基于通道注意力机制和ResNet34的融合模型(SE-ResNet34)。搭建实验平台，在前艉轴承3个测位以不同采样频率采集不同轴转速下正常状态及3种磨损(0.01、0.03、0.05 mm)故障状态的振动信号。对信号进行连续小波包变换得到时域图，经矩阵变换转换为二维灰度图像，再经归一化处理后构建训练集和测试集。所提模型经训练、测试及参数调优后用于油润滑船舶艉轴承磨损故障诊断。实验结果表明:所提模型能够准确、有效地识别故障类别，诊断准确率达99.44%;与ResNet18等其他模型相比，所提模型能够更有效地识别故障，为油润滑船舶艉轴承故障诊断模型的设计与应用提供了新思路。更多还原

【Abstract】 To address the problems in traditional intelligent diagnosis of oil-lubricated ship stern bearing wear faults,such as the difficulty in measuring vibration signals and the insufficiency of training data leading to model overfitting and performance degradation,as well as the vanishing gradient problem caused by increased training network depth that affects diagnostic accuracy,this paper proposes a fusion model (SE-ResNet34) based on a channel attention mechanism and ResNet34. An experimental platform is established to collect vibration signals under the normal state and three wear fault states (0. 0,0. 03,0. 05 mm) at different shaft rotation speeds,using different sampling frequencies at three measuring positions on the front stern bearing. The signals are transformed into time-frequency maps via continuous wavelet packet transform,converted into two-dimensional grayscale images through matrix transformation,and then normalized to construct training and testing datasets. The proposed model,after training,testing,and parameter tuning,is applied to the fault diagnosis of oil-lubricated ship stern bearing wear.Experimental results demonstrate that the proposed model can accurately and effectively identify fault categories,achieving a diagnostic accuracy of 99. 44%. Compared with other models such as ResNet18,the proposed model can identify faults more effectively,providing new insights for the design and application of oil-lubricated ship stern bearing fault diagnosis models.更多还原