Abstract: The fault early warning model for centrifugal blowers established by laboratory data become invalid when used in factories. And data collecting there was difficult to build high-precision model. To address this problem， this paper proposed a transfer learning method based on AE， which had ability to build the fault early warning model suitable for factory environments quickly. Firstly， the monitoring data of centrifugal blowers collected in the laboratory was windowed resampled to establish an AE model with sparse constraints. Secondly， the factory and laboratory data were input into the AE network to obtain low-dimensional features， and  the maximum mean discrepancy（MMD）between the two low-dimensional features was minimized， then  the AE model with small learning rate was fine-tuned to complete the transfer of AE model. Finally， based on the trained AE model and early warning in dicator， the fault early warning strategy was formulated to realize fault early warning for centrifugal blowers under the actual environments of the factories. The experimental results on a centrifugal blower show that the proposed method has higher accuracy compared with other three methods. 

Key words: fault early warning, centrifugal blower, autoencoder(AE), transfer learning

摘要： 工厂实际运行环境下，基于实验室数据的离心鼓风机故障预警模型常常失效，且实际运行数据难以支撑高精度预警模型的构建。提出一种基于自编码的迁移学习方法来快速构建适用于实际运行环境的故障预警模型。首先对实验室采集的离心鼓风机监测数据进行加窗重采样，建立融合稀疏限制的自编码模型；然后将工厂和实验室数据输入自编码网络得到低维特征，最小化两者低维特征的最大均值差异，进而采用较小学习率调整自编码模型完成模型迁移；最后，基于调整后的模型与预警指标制定故障预警策略，实现工厂实际环境下离心鼓风机故障的准确预警。在某型号离心鼓风机数据集上的实验结果表明，该方法与其他三种方法相比具有更高的故障预警精度。

关键词: 故障预警, 离心鼓风机, 自编码, 迁移学习