To address the issue of low accuracy in estimating the State of Charge (SOC) of lithium batteries using the Unscented Kalman Filter (UKF) algorithm, a combined ELM-UKF algorithm with a state detection mechanism is proposed, leveraging the complementary advantages of Extreme Learning Machine (ELM) and UKF for estimating the SOC of lithium batteries. Firstly, the algorithm uses the relevant filtering data estimated by UKF for battery SOC as a sample set to train the ELM model. The successfully trained ELM model is then used to online compensate for the SOC estimation error of UKF, thereby achieving real-time correction of estimation deviations. Secondly, the algorithm designs a state detection mechanism for the predictive output of the ELM model to reduce the impact of overfitting in the ELM model’s predictive output on the smoothness of the SOC estimation waveform. Experimental results show that, compared to single-type algorithms, the proposed combined algorithm exhibits good robustness and generalization, effectively enhancing the estimation performance of lithium battery SOC.