模型无关元迁移学习用于空间滚动轴承寿命阶段识别

Model-agnostic meta-transfer learning for life stage identification of space rolling bearings

针对变工况下空间滚动轴承寿命阶段识别时因样本分布差异较大、可训练用寿命阶段样本较少以及不同寿命阶段样本数量不均等所造成的寿命阶段识别准确率较低的问题,提出模型无关元迁移学习(Model-Agnostic Meta-TransferLearning,MAMTL)用于空间滚动轴承寿命阶段识别。在MAMTL中,将模型无关元学习和迁移学习相结合以实现多任务同步平行训练从而代替传统的迭代训练,多个任务损失函数利用不同工况下无类标签样本和历史工况下少量有类标签样本共同更新MAMTL网络参数,以寻求网络参数的全局最优解,这使MAMTL具有更好的泛化能力,因此MAMTL在较少历史工况有类标签训练样本情况下比传统迁移学习具有更好的域适配性;在MAMTL中构建新型原型网络以将历史工况每一类别的样本表示为一个原型,通过计算当前工况待测样本与原型的相似度完成当前工况待测样本分类,且该分类过程无需参数学习,因此可防止样本不均等情况下对于不同类别样本识别精度差距较大和在少量有类标签训练样本情况下网络出现过拟合的问题,从而更好提高分类精度。MAMTL的以上优势使得它可利用空间滚动轴承历史工况下的少量、非均等已知寿命阶段的训练样本对当前工况待测样本进行较高精度的寿命阶段识别。空间滚动轴承寿命阶段识别实例验证了该方法的有效性。

Considering the low accuracy of life stage identification caused by the large difference of sample distribution,the small number of life stage training samples and the unequal number of samples in different life stages,Model-Agnostic Meta-Transfer Learning(MAMTL)is proposed to identify the life stage of space rolling bearings.In MAMTL,model-agnostic meta-learning and transfer learning are combined to achieve multi-task synchronous parallel training instead of traditional iterative training.Multi-ple task loss functions in MAMTL use unlabeled samples under different working conditions and a small number of labeled samples under historical working conditions to jointly update the network parameters of MAMTL to seek global optimal solutions of them,which makes MAMTL have better generalization ability so that MAMTL has better domain adaptability than traditional transfer learnings when there are few labeled training samples under historical conditions.Moreover,a new prototype network is construct-ed in MAMTL to represent the samples of each class in historical working conditions as a prototype.Thus,the testing samples un-der current working conditions are classified by calculating the similarity between the testing samples and the prototypes,and the classification process does not need parameter learning,which can prevent the problem of large difference in identification accuracy of different classes under the unequal number of samples in different classes and the over fitting problem of network in the case of few labeled training samples,and then better improve the classification accuracy.The above advantages of MAMTL enable it to use few and unequal training samples known for the life stages under historical working conditions of space rolling bearings to per-form high-precision life stage identification for the current testing samples.The life stage identification examples of a space rolling bearing demonstrate the effectiveness of the proposed MAMTL-based life stage identification method.