A novel deep learning approach for denoising and classification of SDRF sensing data for breathing patterns

Abstract This study presents a deep learning-enabled framework for non-contact respiratory pattern classification using software-defined radio frequency sensing. The proposed approach builds on an SDR-based wireless channel state information acquisition system to identify three breathing patterns: normal breathing, fast breathing, and shortness of breath. To improve signal quality and class separability, the raw WCSI (wireless channel state information) data were processed through a structured denoising pipeline comprising wavelet-based filtering, DC component removal, low-variance subcarrier elimination, and normalization. Correlation matrix and power spectral density analyses were then used to compare raw and cleaned signals, demonstrating that preprocessing substantially enhances the visibility of breathing-induced channel variations. DNN (Deep Neural Network) and RNN (Recurrent Neural Network) models were trained and evaluated using five-fold cross-validation. The cleaned data produced consistently high classification performance, with both models’ achieving accuracy above 98%, while the DNN showed marginally superior precision, recall, and F1-score. The DNN confusion matrix further demonstrated near-complete separation among the three breathing classes, confirming the effectiveness of the preprocessing and feature extraction strategy. A multi-attribute decision-making analysis (MADM) based on the SAW (Simple Additive Weighting method) was also introduced to rank model performance across multiple evaluation metrics, identifying the cleaned DNN as the most robust configuration. The results confirm that SDRF (software-defined radio frequency) sensing combined with signal conditioning and deep learning can provide a low-cost, scalable, and contactless solution for respiratory monitoring. This framework offers promising potential for smart healthcare applications, including sleep monitoring, respiratory disease screening, and continuous patient observation in environments where wearable or contact-based sensors may be unsuitable.