Denoising Autoencoder for ECG Signals Based on Frequency Domain Alignment and Multi-Scale Transformer

Electrocardiogram (ECG) signals are widely utilized for cardiovascular disease monitoring. However, these signals are often susceptible to various types of noise during acquisition, which can significantly impair subsequent disease monitoring and diagnosis. Under high-intensity noise interference, existing denoising methods face challenges in effectively extracting the essential features of the P-QRS-T waves in ECG signals, and a frequency domain gap persists between the denoised signals and the original signals. To address these issues, we propose a denoising autoencoder model based on Multi-Scale Transformer (MST) and frequency domain alignment. To enhance the model's capability to extract P-QRS-T wave features, both the encoder and decoder are constructed using MST equipped with multi-scale window attention modules. These modules enable the capture of both local details and global features of the P-QRS-T waves in ECG signals across different window scales, thereby enhancing the model's denoising performance. To mitigate the frequency domain gap, we further introduce a Multi-Resolution Focused Frequency Domain Loss (MRFFL) function, which optimizes the model's fitting of frequency domain components. We conducted inter-patient testing on the QT database, and the experimental results demonstrate that the signal-to-noise ratio (SNR) of the denoised signals achieved by our model reaches 20.82 dB, marking an improvement of 10.8%.