Water beam mark is a common problem in slab heating, which causes quality defects on strip steel. At present, it is difficult to recognition the water beam mark and the workload is heavy in industry. In order to solve these problems, proposed a recognition algorithm of hot rolled strip steel water beam mark based on semi-supervised learning model of improved Denoising Auto-encoder. Random noise was added to each layer of the coding layer of Denoising Auto-encoder, a classification layer was added after the hidden layer and fake labels are added to training data. These methods made the model own the ability of semi-supervised. In this paper, by extracting the feature of temperature difference of strip temperature data at the outlet of roughing mill, this feature was used to train the model. The experimental results showed that compared with BP, Deep BP, DBN and LSTM, the classification accuracy of proposed model is 5.0% - 10.0% higher than other models when the number of tags proportions is small, and when the number of tags proportions is large, the accuracy of proposed model is up to 93.8%.