Data Imputation and denoising using Ensemble Convolutional Autoencoder.
CPTu tests are widely used to determine the properties of soft soils, including clay, fine sand, and silt. The tests measure Cone Tip Resistance (CTR), Sleeve Friction Resistance (SFR), and Pore Water Pressure (PW). However, missing CPTu data is common due to subsurface obstructions. This project proposes an ensemble convolutional autoencoder model using Deep Image Prior (DIP) to impute CPTu data.
The CPTu data is embedded into a 2D matrix by slicing the 3D spatial data into 2D planes. Missing values are initially imputed with K-Nearest Neighbor (KNN) during data processing. Two input matrices are constructed by filling the missing values with zeros and KNN imputed values.
The task of data imputation is performed using a convolutional autoencoder. The outputs produced by two different input matrices are linearly combined by an ensemble layer.
The model was tested and compared to the baseline for all measurement types at a wide range of missing rates. The results demonstrate that the proposed method can achieve better imputation accuracy and robustness compared to the baseline methods.
The proposed CPTu data imputation method based on an ensemble convolutional autoencoder can reduce the number of tests required and improve the soil profiling accuracy in case of missing data.