As oil and gas exploration and development to deeper and more complex formations, the possibility of downhole risks such as overflow, leakage, and sticking occurring during drilling process is increasing, and the parameter characteristics of these downhole risks vary under different drilling working conditions. Therefore, the automatic identification of drilling working conditions is particularly important for the automatic diagnosis and early warning of underground risks. This article proposes a working condition judgment method based on coupled graph recognition of drilling parameters, establishes an intelligent recognition reinforcement learning model for drilling working conditions, and uses a discrimination tree to optimize the diagnostic process, fully leveraging the advantages of recognition algorithms based on state variables and trend variables to accurately identify drilling working conditions in real time. Nine key comprehensive logging parameters, including well depth, standpipe pressure, WOB, bit depth, turntable speed, inlet flow rate, and hook height, were selected as input features. A recognition method based on state variables and trend variables was established. By coupling the recognition results obtained from the two methods, intelligent discrimination of drilling, circulation, making up a joint, POOH (pull out of hole), and RIH (run in hole).Adopting a distributed proximal strategy optimization algorithm for optimization and improvement, the optimal values of each parameter were obtained, and the average recognition accuracy was improved to 95.09%.The algorithm model combined traditional diagnostic methods and physical model-based artificial intelligence methods has low complexity and strong interpretability, providing an effective solution for real-time intelligent recognition of drilling conditions and providing reference for the diagnosis and warning of more drilling conditions and downhole risks.