Self-Supervised Representation Learning for Multivariate Time Series of Power Grid with Self-Distillation Augmentation

In recent years, a wealth of sensors have been configured in power grid scenarios to detect abnormal conditions so that the intelligent level of fault diagnosis could be improved. For time series data, self supervised learning can learn the data patterns to improve the ability of feature representation. However, current methods only impose constraints on the reconstructed masked time series, ignoring the impact of visible points. And the visible features in decoders have a richer knowledge, which is conducive to more effective feature. Therefore, this paper proposes a time series representation learning method based on self-supervised learning, which extends the self-distillation framework on the basis of the original mask reconstruction framework, using the encoder features of visible timing positions as students, and the decoder features of corresponding positions as teachers. By means of self distillation feature, higher level knowledge is transferred to the encoder, enabling the encoder to extract features with stronger representation capabilities for time series, thereby improving the prediction accuracy of downstream tasks.