Multi-residual tensor completion for spatiotemporal data imputation

In Cyber–Physical Systems, the spatiotemporal data collected often contains many missing values, which result from uncontrollable factors like sensor failure, communication disruption, and environmental interference. The missing values can significantly degrade system performance and even jeopardize system stability. In previous studies, tensor models were considered effective in spatiotemporal data imputation, attributed to their capability to capture spatiotemporal correlations within the data. Although tensor models can effectively capture the global features of data, they cannot learn the local fluctuation characteristics well. To further enhance the tensor model's ability to capture local features, a residual iteration strategy was designed, enabling the model to learn local features from the previous round of residuals. Additionally, a multi tensor completion strategy was developed to achieve more accurate learning in each round of residuals. Combining these two strategies with tensor completion results in the Multiple Residual Tensor Completion (MRTC). We demonstrate through the visualization of imputation results that MRTC can further learn local features of the data compared to the original tensor completion model. In addition, comparative experiments were conducted on three publicly available spatiotemporal datasets, and the results showed that MRTC performed well in various missing data scenarios, outperforming the other four state-of-the-art tensor based imputation models in most cases.