Pipe sticking significantly increases drilling costs. In directional wells, the irregular trajectory of the wellbore, combined with gravitational forces, increased the friction between the drill string and wellbore wall which elevates the risk of pipe sticking. By considering the drill string’s flexibility, this study uses a soft-string model to calculate the rotational friction coefficient in build and hold segments of directional wells. The calculations utilize real-time data, including surface well logging data such as weight on bit and rotary table torque, along with near-bit measurements like bit torque. An unsupervised learning neural network with dual autoencoders (USAD) is employed to simultaneously detect anomalies of friction coefficients in the build and hold sections of the well, obtained through inversion. Based on the detection results, this model provides early warnings for potential pipe sticking incidents and identifies the specific sections at risk. The USAD model was trained using a dataset comprising data from 16 wells. It accurately identified 84% of the 541 instances of increased friction coefficient anomalies contained in the dataset. Historical well analyses show that this method detects changes in drill string friction earlier than surface manual detection. It provides early warnings for pipe sticking and identifies high-risk sections. This approach leverages the potential of real-time drilling data and introduces a novel strategy for detecting and predicting pipe sticking.