To address the challenge of real-time weight transfer recognition in directional well sliding drilling, an intelligent multi-level identification method based on measurement-while-drilling (MWD) real-time data is proposed. A four-level weight transfer evaluation system based on weight-on-bit (WOB) transfer ratio is established. By analyzing the response characteristics of downhole WOB and vibration signals, a multi-dimensional feature space is constructed from statistical domain, frequency domain, and temporal evolution perspectives. A comprehensive importance evaluation strategy is adopted to select 10 core features. To meet the strict requirements of real-time performance and lightweight design for downhole closed-loop control, a shallow random forest recognition model is designed, utilizing class weight methods to handle sample imbalance and well-based data partitioning strategy to ensure model generalization capability. Based on measured data from 5 directional wells in a western oilfield, the model achieves 90.2%accuracy and 0.900 Macro-F1 score on the independent well test set, with a recall rate of 87.6%for complete weight transfer. The model is successfully deployed on an ARM Cortex-M4 processor with 52 KB storage space and 355 milliseconds inference time, meeting all downhole hardware constraints. The consistency between the model decision logic and weight transfer physical mechanism is verified through interpretability analysis. The research results can be directly applied to intelligent on-off control of downhole active control devices such as hydraulic oscillators, reducing response time from minute-level of traditional surface control to within 5 seconds, which has significant engineering value for improving drilling efficiency and reducing downhole risks.