Investigation into the Spectral Completion Algorithm Leveraging Dense Connection Autoencoders

To address the issue that existing deep learning-based spectrum completion methods focus on Power Spectral Density (PSD) data and are not adapted to the characteristics of Reference Signal Received Power (RSRP) data, making it difficult to balance accuracy and efficiency in sparse RSRP scenarios, a spectrum completion method based on a fully convolutional dense-connected autoencoder is proposed. An RSRP data matrix is constructed as the input, 4 missing scenarios are designed, and a mask matrix is introduced to mark the validity of data. The existing architecture is improved: the encoder is composed of dense connection blocks and transition layers, the decoder is composed of transposed convolution upsampling blocks and dense connection blocks, and an additional bottleneck layer is added between them to preserve feature expression. Masked Mean Squared Error is used as the loss function. Experiments with 3 comparison methods show that compared with the second-best fully convolutional residual autoencoder, the proposed method reduces Root Mean Square Error (RMSE) by 6.41%, increases Structural Similarity Index (SSIM) by 1.77%, and improves Peak Signal-to-Noise Ratio (PSNR) by 1.47 dB. It can support the high-precision generation of spectrum maps from sparse data in scenarios such as 5G/6G spectrum management.