Abstract: To address the problem of insufficient samples obtained under small sample conditions and the decrease in diagnostic accuracy caused by ineffective feature extraction,a small sample gear fault diagnosis method combining Gramian Angular Summation Field (GASF) and MSCAM-DenseNet was proposed.The Gramian Angular Summation Field (GASF) was used to transform the multi-source vibration signals into two-dimensional features,and the multi-source features were reconstructed using Two-Dimensional Discrete Wavelet Transform (2D-DWT).Since the conventional DenseNet lacks the ability to recognize multi-scale features,a Multi-Scale Channel Attention Mechanism (MSCAM) was introduced into DenseNet,and an improved network model called MSCAM-DenseNet was proposed.Finally,the reconstructed GASF was used as the input to MSCAM-DenseNet,and after feature recognition was completed,a network classifier was used to classify the fault features.The proposed model was validated using the planetary gear dataset obtained from the laboratory and the gearbox dataset from Southeast University,and compared with other diagnostic models.Experimental results demonstrated that the proposed method achieved high fault recognition accuracy,strong generalization ability and noise resistance under small-sample and varying operating conditions.

Key words: gear, small-sample fault diagnosis, Gramian angular summation field, two-dimensional discrete wavelet transform, multi-scale channel attention mechanism

摘要： 针对小样本条件下所得样本不足,特征未能有效提取导致诊断精度下降的问题,提出一种GASF与MSCAM-DenseNet相结合的小样本齿轮故障诊断方法。首先,运用格拉姆角和域(GASF)将多源振动信号变换为二维特征,采用二维离散小波变换(2D-DWT)重构多源特征。其次,由于一般的密集连接卷积网络(DenseNet)不具备识别多尺度特征的能力,因而在DenseNet中引入多尺度通道注意力机制(MSCAM),提出一种改进网络模型,即MSCAM-DenseNet。最后,以重构后的GASF作为MSCAM-DenseNet的输入,待特征识别完成后,由网络分类器完成故障特征分类。采用实验室行星齿轮数据集和东南大学齿轮箱数据集对所提模型验证,并与其他诊断模型进行对比。实验结果证明,所提方法在小样本、变工况条件下具有较高的故障识别准确率,较强的泛化能力和抗噪能力。

关键词: 齿轮, 小样本故障诊断, 格拉姆角和域, 二维离散小波变换, 多尺度通道注意力机制