A Deep Neural Network-Based Method for the Separation of Superimposed Signals

In the current era of digital interconnectivity, signal processing has become increasingly complex, especially in fields such as communications, biomedical sciences, and audio analysis. Traditional signal processing methods, though effective in some scenarios, often falter when faced with the non-linearity, correlation, and noise that characterize complex, real-world signal environments. In response to these challenges, this study presents a novel deep learning-based approach for separating superimposed signals, leveraging the sophisticated capabilities of Deep Neural Networks (DNNs). Our innovative method transforms a fully connected neural network by configuring each node to represent a sinusoidal function. This design allows for the creation of a composite signal through the superposition of multiple sinusoidal functions. A key feature of our method is its ability to automatically adjust the parameters of these functions, thus accurately capturing the original signal and effectively decomposing it into its constituent sinusoidal components. This strategy marks a significant shift from traditional methods, harnessing the neural network's potential to independently learn from data and move away from reliance on predefined signal models. The effectiveness of our technique is corroborated by extensive experiments, which demonstrate the network's high accuracy in disentangling superimposed signals. These findings are supported by both quantitative metrics and visual analyses, highlighting the promise of our method for complex signal processing tasks.