The application of image processing technology based on deep learning in recognition of dropper state is gradually rising. However, the lack of abnormal samples and the particular shape of the dropper significantly increase the difficulty of training the general classification network. In this paper, we propose a convolution autoencoder aiming at the unbalanced data set based on the VGG classification network. It adopts semi-supervised learning to avoid the disadvantages of unbalanced data sets. We embed the memory mechanism in the network, which improves capability of state identification further, and use lightweight module ghost reduces the model parameter. In the balanced test set, the recognition accuracy of the algorithm is 92%, which has achieved good performance in the related applied research.