Abstract
With the rapid advancement of Internet of Things (IoT) technology, the volume of sensor data collection has increased significantly. These data are typically presented in the form of time series, gradually becoming a crucial component of big data. Traditional time series analysis methods struggle with complex patterns and long-term dependencies, whereas deep learning technologies offer new solutions. This study introduces the U-TSS, a U-net-based sequence-to-sequence fully convolutional network, specifically designed for one-dimensional time series segmentation tasks. U-TSS maps input sequences of arbitrary length to corresponding sequences of class labels across different temporal scales. This is achieved by implicitly classifying each individual time point in the input time series and then aggregating these classifications over varying intervals to form the final prediction. This enables precise segmentation at each time step, ensuring both global sequence awareness and accurate classification of complex time series data. We applied U-TSS to geomagnetic field observation data for the detection of high-voltage direct current (HVDC) interference events. In experiments, U-TSS achieved superior performance in detecting HVDC interference events, with accuracies of 99.42%, 94.61%, and 95.54% on the training, validation, and test sets, respectively, outperforming state-of-the-art models in accuracy, precision, recall, F1-score, and AUC. Our code can be accessed openly in the GitHub repository at https://github.com/wangmengyu1/U-TSS.