For PHM(Predictive Health Management) in an industrial scenario, motor vibration anomaly detection requires a large number of labeled data. The problem is that the anomaly vibration signal data is scarce and difficult to be obtained, especially the problem of insufficient data in the building and deployment of multi-machine models. This paper proposes a Meta-learning method based on deep unsupervised learning(Autoencoder). We use non-labeled and a small amount of vibration signal data which is converted into 52 key physical and statistical features to build models of different motors. By inputting these 52 features into designed models, other 52 features can be reconstructed as output. We use input and output features to calculate reconstruction error which is used as the criterion for the performance of models. We also use accuracy as the criterion for the performance of models. In the actual industrial situation, the improvement of accuracy by our proposed method is about 33.50% better than the method only with unsupervised learning for the new sensor model with few vibration data.