When using mud logging time-series data to determine the state of a rig"s slips (setting and releasing), the threshold method and machine learning methods are limited by differences in rig models, drill string types, well types, etc. These limitations result in poor model applicability and insufficient generalization. Additionally, when the drill string depth is less than 200 meters, the recognition error increases significantly.To address the above issues, this paper proposes a method for identifying the setting and releasing states of rig slips by improving the Long Short-Term Memory (LSTM) network using an Output Feedback (OF) mechanism. First, the rig states are classified into four categories: light load, slip releasing, heavy load, and slip setting. The predicted output of the current sequence is then fed back as input to the prediction of the next sequence. Second, the loss function is improved using Time Distance Weight (TDW), resulting in a weighted loss that accounts for time-series information.Verification using drilling data shows the following results: When the drill string depth is less than 200 meters, the accuracy of the OF-LSTM model reaches 99.01%, which is 21.01% higher than that of the conventional LSTM model. When the drill string depth exceeds 200 meters, the accuracy of the OF-LSTM model is 99.22%, representing a 14.09% improvement over the LSTM model. This model can provide key technical support for the refined analysis of drilling efficiency and the optimization of rig operation processes.