Abstract: Considering the power transformer is one of the key components of the power system, its reliability directly affects the stability and safety of the entire system. Therefore, it is important to analyze the transformer's operating status in real-time and perform an accurate fault diagnosis. A semi-supervised manifold embedding (SSME) method for online transformer fault diagnosis was presented in this paper in response to the problem that it is difficult to collect transformer operating status data, and the unreliability of fault analysis models due to the lack of fault data. This model is developed based on the concentration characteristics of five different gases in the transformer oil, including H₂, CH₄, C₂H₆, C₂H₄ and C₂H₂. The gas concentration sample data were obtained from online monitoring in order to analyze the operation status of the transformer. In this model, the status detection of 700 online monitoring data can be performed within 0.1 seconds, thereby meeting the requirements of online analysis. Experimental comparisons of different transformer fault diagnosis models are performed to determine its prediction accuracy and efficiency. According to our experimental results, the proposed method has higher fault diagnosis accuracy than the classical deep belief network (DBN), K-Nearest Neighbor (KNN), and Random Forest (RF) methods. This method can serve as an effective reference for the safe operation of power transformers.