Aiming at the problems of off-line, large time delay, and weak anti-interference ability of existing rare earth element component content models, an improved GRA-just-in-time learning algorithm(GRA-JITL-LSSVM) is proposed to establish an online component content detection model for the rare earth extraction process. Firstly, the grey relational analysis(GRA) method is used to analyze the changing trend and correlation degree between input and output variables, and the hash table is used to determine the size of the learning set to ensure the integrity of the data similarity information and the rationality of the learning set. Based on this, the least squares support vector machine(LSSVM) model is established, and the database update criterion is introduced to improve the anti-interference ability of the model. Secondly, in order to ensure the global optimization of GRA-JITL-LSSVM model parameters, a genetic algorithm with a stagnation backtracking strategy (SBS-GA) is proposed, and the convergence of the SBS-GA is analyzed and verified. Finally, the simulation experiment is carried out based on Pr / Nd extraction field data. The results show that the proposed SBS-GA can ensure the global solution of the optimization parameters. The proposed GRA-JITL-LSSVM has high real-time performance and good prediction accuracy, which can be used for the online detection of the element component content in the production site of the rare earth extraction.