Well leakage accidents occur frequently during drilling operations in ultra-deep oil and gas wells, which may lead to serious safety hazards if not detected promptly. Therefore, early intelligent detection of well leakage is critical to ensure drilling safety. To address the issues of the scarcity of well leakage data samples and the differences in the response patterns of drilling parameters under different drilling conditions, this study proposes an early intelligent detection method for well leakage based on the adaptive fusion of multi-drilling parameter trends. First, an intelligent recognition model for drilling conditions was built using the Random Forest algorithm (RF). Secondly, by eliminating the influence of drilling conditions on the feature parameters of well leakage and combining the trend values of multi-drilling parameters, an adaptive calculation method for the Leakage Risk Coefficient (LRC) is constructed. Finally, the effectiveness of this method was verified by testing on measured well leakage data. The research results show that the accuracy of the intelligent recognition model for drilling conditions reaches 97.9%, which can be used to optimize well leakage feature parameters. This method is less affected by changes in drilling conditions and parameters, making it applicable throughout the entire drilling process. It can issue an early warning signal for well leakage 16 min earlier than the mud pit gain method, providing valuable time to take appropriate measures.