PB-DDPM: A Pontoon Bridge Denoising Diffusion Probabilistic Model for Missing Data Reconstruction

Abstract

The passive and active data missing in data acquisition reduces data availability and brings difficulties in subsequent data processing. To solve this problem, a data reconstruction method based on pontoon bridge diffusion model has been proposed. This method transformed the missing position into a 0-1 mask tensor. By element-wise multiplication, missing values were replaced by the Gaussian noise to construct new data for model inputting according to the mask. And then, the missing data were reconstructed under the condition of observed data. Compared to other methods that divide the original data into missing and observed parts while the data dimensions remaining unchanged to make up inputs, this approach reduced the dimension of input data and simplified the correlation between inputs. Meanwhile, using the original parameters of noise strategy as the anchors and adjusting the uncertainly coefficients by in-situ replacement, the pontoon bridge mode reduced the difference between distributions of practical reconstruction and theoretical calculation to enhance the relevance of training process and generating process. Furthermore, the input data of the training process is augmented with various masks to simulate data missing modes. Experimental results demonstrated that data augmentation techniques enhanced the model’s ability to handle missing data, and the pontoon bridge diffusion model could effectively improve the quality of reconstructed missing data.