As one of the most practical metal laser additive manufacturing technologies, selective laser melting (SLM) has been widely adopted in aviation, aerospace, and energy sectors due to its advantages in rapid forming of complex thin-walled components. However, the consistency issue during the forming process still limits further improvements in component quality, which is closely related to defects arising from the constant variations in melt pool size and shape. Therefore, to more effectively monitor the dynamic changes of the melt pool, a method for predicting melt pool melting state categories based on extraction of high-dimensional melt pool motion features and a long short-term memory (LSTM) model is proposed. Firstly, the U-net model is utilized to extract melt pool morphology features from melt pool images, and the distances from the melt pool centroid to its boundary are calculated and unfolded along the contour into highdimensional vectors to represent the motion features of the melt pool. Subsequently, the k-means clustering algorithm is applied to perform clustering analysis on the melt pool motion features under different process parameters, leading to the construction of four categories of melt pool melting states. Time series prediction of the melting state categories is then conducted using the LSTM model. Taking the SLM process of Inconel 625, a typical high-temperature alloy material for aviation, as an example, verification of melt pool state category prediction was conducted. The results demonstrate a prediction accuracy of 85.92%, providing a novel approach and insight for real-time monitoring and quality control in the SLM process.