Abstract: Due to the scarcity of liquid loading samples in shale gas development field operations, resulting in low prediction accuracy and significant delay in early warning of existing liquid loading diagnostic methods, a novel early intelligent diagnostic method for shale gas wellbore liquid loading is proposed. This method combines Variational Mode Decomposition (VMD) and an improved Autoencoder (LSTM-AE). Based on minute-level production dynamic data collected by the Supervisory Control and Data Acquisition system (SCADA), VMD is applied to feature selection and reconstruction of the raw monitoring data. For model construction, an unsupervised training approach is used with normal production phase samples of shale gas wells to train the autoencoder model. Additionally, the Long Short-Term Memory (LSTM) network enhances the temporal feature extraction capability of the autoencoder for shale gas well production data samples. A discrimination threshold is set based on the reconstruction error output by the model, enabling intelligent recognition and early warning of liquid loading conditions. The experiment utilizes production dynamic monitoring data from a well with historical liquid loading records on a platform in the Weiyuan gas field of the Sichuan Basin. Using a sliding window, 5,000 normal production and liquid loading samples are generated for model training and testing. The results show that the model achieves an early warning accuracy of 97.8% on the test set, with a false alarm rate of only 2.6%. In real-time warning simulations for the test well, the model provides an early warning 31 minutes ahead of manual monitoring, demonstrating strong timeliness and the ability to provide early alerts during liquid loading events. This confirms the effectiveness of the proposed method in practical applications.