Abstract: The time-lapse resistivity monitoring data in mines is a high-dimensional dataset that includes attributes such as recovery direction, working face extension direction, depth direction, and resistivity values. Its true distribution is unknown, and blindly applying existing dimensionality reduction methods may diminish certain high-dimensional attributes that are closely related to data quality. At present, data selection relies heavily on manual experience, resulting in a low level of automation. To address these issues, a deep learning-based feature extraction and intelligent screening method for time-lapse resistivity monitoring data in mines was proposed. First, high-dimensional discrete data containing spatial 3D coordinate information and resistivity values were subjected to dimensionality reduction to capture essential features of the data, eliminate redundancy, and achieve multi-scale feature extraction. Then, the ResNet10 convolutional neural network was used to extract 2D features from each slice and compute their structural similarity to assess the spatial continuity and smoothness of resistivity anomalies. A Transformer network was used to extract 3D features from the resistivity monitoring data. Finally, spectral clustering was applied in the feature space to perform intelligent screening of the monitoring data. The proposed method and manual selection method were used to extract features and perform quality selection on 16 monitoring datasets collected in a single day from a coal mine area. The results showed that manual selection by different personnel produced completely different results, indicating strong subjectivity, poor repeatability, lack of fixed evaluation criteria, and took an average of 30 minutes, leading to poor real-time performance. The proposed method achieved 100% consistency in the experimental results, and each selection took less than 30 seconds, indicating that the selection results were objective, stable, reliable, and fast.