<p>Geomagnetic field observation data, as a critical source for studying the seismic-magnetic relationship, is typical time series data. In recent years, it has been significantly affected by high-voltage direct current interferences. To detect these High-voltage Direct Current interference events (HVDCs) in geomagnetic field observation data, manual detection methods are currently used, which is time-consuming and labor-intensive. Theoretically, detecting HVDCs is a time series segmentation problem that presents three main challenges: (1) multi-scale interference patterns that are difficult to capture with traditional models, (2) inadequately refined feature extraction, particularly in the presence of noisy or complex data, and (3) class imbalance, as HVDCs are much rarer than normal data. To address these issues, we proposed U-SegTime, a novel end-to-end time series segmentation model based on an improved U-Net architecture. Our model employs depth-wise separable convolutions with varying dilation rates to capture multi-scale features, incorporates a squeeze-and-excitation attention mechanism to enhance feature learning, and utilizes a weighted cross-entropy loss function to address the class imbalance. Additionally, the Bayesian optimization method is applied to fine-tune the hyperparameters of the U-SegTime and improve its overall performance. The experimental results demonstrate that the proposed U-SegTime model achieved an accuracy of 95.77%, a recall of 80.77%, an F1 score of 0.8187, an AUC of 0.9730, and an mIoU of 0.8232 on the test set, outperforming the state-of-the-art methods. This study provides a new idea for automatically detecting vehicle interference, subway interference and other interference events in geomagnetic field observation data.</p>