Multiple parallel hidden layers autoencoder for denoising ECG signal

Abstract

Deep learning with multiple hidden layers denoising autoencoders (MHL-DAE) is commonly used to denoise images and signals through dimension reduction. Here, we explore the potential of multiple parallel hidden layers denoising autoencoder (MPHL-DAE) to denoise complex bio-signals, like electrocardiogram (ECG). A merge layer, e.g., average layer is considered as the output of the proposed model by combining the outputs of the parallel hidden layers. The parallel hidden layers in the coding layer with activation function of different scale a, e.g., are considered to capture distinct features of the input. The lower/upper number of the required hidden neurons of the coding layer are estimated using data driven approach via singular values decomposition (SVD). The results show that the proposed MPHL-DAE model achieve better/similar SNR improvement compared to MHL-DAE with suitable scale for various noise levels respectively.

Keywords

• denoising autoencoder
• multiple parallel hidden layers
• scaled hyperbolic tangent activation function