基于增强迁移卷积神经网络的机械智能故障诊断

现有的基于深度迁移学习的智能诊断方法通常对源域和目标域特征对齐来减少两者分布差异，没有考虑源域类别决策边界对目标域特征匹配的影响，针对此不足，提出了一种增强迁移卷积神经网络（Enhanced transfer convolutional neural network，ETCNN）来改进机械设备在变工况下的诊断精度和泛化能力。为充分利用标签信息和提取高维特征，构建卷积神经网络和两个独立的分类器对源域数据分别训练，用于检测远离决策边界的目标域样本。为有效减少不同决策边界处样本的误匹配，进一步构建分类损失函数和分类器判别损失函数，并引入对抗训练策略，最大两个分类器的分类差异，同时最小化源域与目标域特征分布差异，实现目标域样本与源域样本自适应匹配，从而有效改进故障的分类性能。在滚动轴承数据集上对所提方法进行充分评估，并与其他三种深度迁移学习方法：域适配网络（Domain adaptive network，DAN），多层域适配网络（Multi-layer DAN，MLDAN），以及深度对抗卷积神经网络（Deep adversarial convolutional neural network，DACNN）进行充分比较，结果表明，所提方法不仅具有良好的分类能力和泛化能力，同时明显优于其他方法。

Intelligent Fault Diagnosis for Machinery Based on Enhanced Transfer Convolutional Neural Network

Existing intelligent fault diagnosis methods based on deep transfer learning techniques usually align the features of source domain and target domain to reduce distribution discrepancy, ignoring the influences of category decision boundary on target domain samples. Therefore, an enhanced transfer convolutional neural network (ETCNN) is proposed to improve the diagnostic accuracy and generalization ability of mechanical equipment under variable operation conditions. To make full use of the label information and extract high-dimensional features, convolutional neural network with two independent classifiers is separately trained on the source domain data to detect target domain samples away from the decision boundary. To effectively reduce the mismatch of samples at different decision boundaries, a classification loss function and a classifier discriminative loss function are constructed. Furthermore, an adversarial training strategy is introduced to adaptively match samples across domains and improve the classification performance by maximizing the classification differences between the two classifiers while minimizing the differences in feature distributions. The proposed method is fully evaluated on the rolling bearing data set and compared with the other three diagnosis methods including domain adaptive network (DAN), multi-layer DAN (MLDAN) and deep adversarial convolutional neural network (DACNN). The results show that the proposed method not only presents good classification and generalization performance, but also significantly outperforms other methods.