Modulation Recognition Method with Dual-Branch Feature Fusion Based on Lightweight CNN and Dual-Attention Network

In the invisible ocean that surrounds us, a constant stream of information travels as radio waves. These waves carry everything from emergency broadcasts to the data that powers our smartphones, but they are often hidden in plain sight, disguised by the static of the natural world. To make sense of this chaotic environment, engineers rely on a process called automatic modulation recognition. Think of it as a translator for the airwaves; its job is to listen to a signal and instantly identify what kind of code it is using. Different signals use different methods to pack information, much like how different languages use different sentence structures. Some signals are simple and robust, while others are complex and efficient, packing more data into the same space. The challenge arises when these signals are weak or buried under noise, such as the interference from other electronic devices or the natural static of the atmosphere. In these difficult conditions, the subtle differences between signal types can vanish, making it nearly impossible for traditional systems to tell them apart. This is a critical problem for military reconnaissance, spectrum monitoring, and the future of wireless communication, where knowing exactly what a signal is can be the difference between success and failure.

A team of researchers at Chongqing Polytechnic University of Electronic Technology and Chongqing Open University has developed a new approach to solve this problem, specifically targeting the confusion that happens when signals are weak. Their method, detailed in a recent study, moves away from trying to force a single, massive computer model to do everything at once. Instead, they designed a system that works like a two-step screening process, combining speed with deep, careful analysis. The researchers recognized that not all signals are created equal. Some signals, like those used in basic frequency modulation, have distinct shapes that are easy to spot even in a noisy room. Others, particularly high-order digital signals, are incredibly similar to one another, differing only in tiny, subtle patterns of amplitude and phase that are easily lost when the signal is weak. To handle this diversity, the team built a dual-branch network. The first branch is a lightweight, fast-moving system designed to quickly scan all incoming signals and filter out the ones that are easy to identify. This branch acts as a rapid sieve, catching the obvious signals and clearing the path for the second branch.

The second branch is where the real work happens for the difficult cases. Once the easy signals are set aside, the system focuses its full attention on the six types of high-order signals that are most likely to be confused with one another. This specialized part of the network uses a dual-attention mechanism, which functions like a pair of focused eyes that know exactly where to look. It examines the signal from two different perspectives simultaneously: one looking at the timing and sequence of the data, and the other looking at the strength and distribution of the signal's components. By weighing these different features, the system can pick out the faint, critical details that distinguish one complex signal from another, even when they are buried under significant noise. The final step involves a decision-making module that listens to both branches. It compares the confidence levels of their answers and selects the most reliable result, ensuring that the final identification is the best possible guess based on the available evidence.

The results of this approach were tested across a wide range of signal strengths, from very weak signals that are barely detectable to strong, clear transmissions. The researchers found that their dual-branch method achieved an overall recognition accuracy of 86.1% across the entire testing range. This performance was particularly impressive in the low-signal conditions that usually defeat other systems. In the critical range where signals are weak, specifically between -8 dB and 0 dB, the new method achieved an average recognition rate of 71.8%. This is a significant improvement over existing models, which struggled to reach similar levels of accuracy without the benefit of this specialized two-step architecture. The study showed that while a single, simple network could identify the easy signals well, it failed miserably with the complex ones. Conversely, the specialized network alone was excellent at the complex signals but confused by the simpler ones. Only by fusing the two together did the system achieve a balance that worked for all twelve types of signals tested, including various forms of phase and amplitude modulation.

The researchers also compared their method against several other leading deep learning models used for this task. In these comparisons, their approach consistently outperformed the others, particularly in the low-signal environment where the difference in performance was most pronounced. The new method was able to maintain stability and accuracy where other models saw their performance drop sharply. This suggests that the strategy of separating the task into a broad, fast screening and a deep, focused analysis is a more effective way to handle the complexity of real-world electromagnetic environments. The system does not require complex manual adjustments or prior knowledge of the specific signals, making it adaptable to changing conditions. By successfully distinguishing between signals that are easily confused, even when they are weak, this research offers a more reliable tool for understanding the invisible world of radio communications. It demonstrates that by mimicking a human-like process of quick scanning followed by careful inspection, machines can become much better at hearing the signal through the noise.