Time series-to-image encoding for saturation line prediction using channel and spatial-wise attention network

Abstract

As an essential indicator of the structural health status of dams, saturation lines are often utilized to reflect the safety of tailings ponds. Effective monitoring of saturation lines during mining operations can significantly reduce the risk of tailings pond failure. However, the existing mathematical models based on shallow machine learning-driven methods are greatly limited in their ability to capture complex time series information. Moreover, these models only capture spatial dependencies but neglect channel adaptability when dealing with 1-D raw data sequences. To address this issue, we propose a channel and spatial-wise attention model to extract both rich visual patterns and time series dependence in the encoded images. Specifically, our model employs a novel two-stage prediction method. Firstly, after capturing refined stable trend characteristics using discrete wavelet transform, we encode the time series saturation line data into 32×32×3 RGB images. Secondly, to further capture visual patterns, channel-wise information is detected using features from different channels in each block of our backbone. The performance of our proposed model is demonstrated to outperform four other state-of-the-art deep neural networks on six real-world datasets of saturation lines.