A Residual Neural Network Framework for Rotating Sound Source Localization

Imagine trying to hear a single voice in a crowded room where everyone is spinning around you. The sound waves reaching your ears are constantly shifting in pitch and timing, creating a chaotic mess that makes it nearly impossible to pinpoint who is speaking. This is the fundamental challenge engineers face when trying to locate the source of noise from rotating machinery, such as the blades of a wind turbine or the fans inside a jet engine. For decades, scientists have used arrays of microphones to capture these sounds and map where they come from. However, when the sound source is spinning, the motion creates a phenomenon known as the Doppler effect, which distorts the sound waves just as a siren's pitch changes as an ambulance speeds past. Traditional methods for untangling this distortion are often slow, computationally heavy, and prone to creating "ghost" signals—false alarms that suggest a noise source exists where there is actually nothing.

A team of researchers at Shanghai Jiao Tong University and AECC Commercial Aircraft Engine Co., Ltd. has developed a new approach to solve this problem by teaching a computer to recognize the patterns of rotating noise directly. Instead of relying on complex mathematical formulas to manually correct for the spinning motion, they trained a type of artificial intelligence called a Residual Neural Network, or ResNet, to look at the raw data from the microphones and instantly identify the true location of the sound. In a series of detailed computer simulations, this new system proved remarkably effective. It successfully located rotating sound sources with an accuracy of up to 98.83 percent, even when the noise came from multiple spinning points at once. Perhaps most impressively, the system made these decisions in less than 50 milliseconds, a speed that is orders of magnitude faster than the conventional methods currently in use, which can take several seconds to process the same information.

The core of this innovation lies in how the researchers prepared the data for the computer to learn. In standard sound analysis, scientists often use a mathematical tool called a cross-spectral matrix, which acts like a map of how sound waves from different microphones relate to one another. For stationary objects, this map is clear and easy to read. But for rotating objects, the map becomes scrambled by the Doppler effect, making it difficult for traditional algorithms to distinguish real sources from the noise. The researchers realized that a specific technique, known as mode composition beamforming, could reorganize this scrambled data into a format that preserved the essential spatial information while accounting for the rotation. They fed this reorganized data into their neural network, which is designed with a special structure that allows it to learn deep, complex patterns without getting confused or losing its way, a common problem when training very large computer models.

To test their system, the team created thousands of simulated scenarios involving rotating sound sources with varying speeds, frequencies, and intensities. They used a microphone array consisting of 16 microphones arranged in a circle, a setup that mimics the equipment used in real-world industrial diagnostics. The simulations covered a wide range of conditions, including sources spinning at speeds between 500 and 2500 revolutions per minute and emitting sounds in the high-frequency range of 4500 to 5000 Hertz. The researchers also tested the system under different geometric arrangements, placing the microphone array both parallel to and perpendicular to the plane of the rotating source. In every case, the neural network was able to identify the correct number of sound sources and their precise locations, effectively filtering out the false signals that plague older methods.

One of the most significant findings was the system's ability to eliminate "ghost" contributions. In traditional beamforming, the distortion caused by rotation often creates phantom images of sound sources at frequencies that do not actually exist. These ghosts can mislead engineers into thinking there are multiple noise problems when there is only one, or they can obscure the true location of a critical fault. The new ResNet-based framework, however, learned to ignore these artifacts entirely. In tests where the system was challenged with up to six distinct rotating sources, it correctly identified all of them without generating a single false positive. This level of precision is crucial for applications like diagnosing faults in aircraft engines, where identifying the exact source of a noise can prevent catastrophic failures.

The speed of the new method offers another major advantage for practical engineering. While the traditional time-domain method known as ROSI can take more than four seconds to analyze a single case, and the frequency-domain MCB method takes over 800 milliseconds, the neural network delivers its answer in under 50 milliseconds. This dramatic reduction in processing time means that the system could eventually be used for real-time monitoring, allowing engineers to watch the health of a machine as it operates rather than waiting for a delayed analysis. The researchers noted that the system works best when the microphone configuration matches the one it was trained on, and that generating the massive amount of training data required for the simulations took considerable time and computing power. However, once the model is trained, it operates with a consistency and speed that manual calculation cannot match.

The study also highlighted the limitations of current approaches and the specific strengths of this new framework. The researchers found that simply making the neural network deeper or more complex did not always guarantee better results; in fact, with insufficient training data, a more complex model performed worse than a simpler one. This suggests that the quality and quantity of the training data are just as important as the architecture of the network itself. Furthermore, while the system excels in the simulated environments tested, the researchers acknowledged that real-world conditions, such as unpredictable background noise or different physical setups, would require further adaptation. They proposed that future work could involve combining simulated data with real experimental data to make the system more robust and adaptable to the messy reality of industrial environments.

Ultimately, this research represents a shift in how engineers approach the problem of rotating sound. By moving away from rigid mathematical corrections and toward a data-driven learning approach, the team has demonstrated that it is possible to see through the distortion of spinning machinery with unprecedented clarity. The ability to localize sound sources with near-perfect fidelity, using a standard array of 16 microphones and without needing to know the exact physical properties of the sound's path, opens new doors for noise control and mechanical diagnostics. As the technology matures, it promises to transform how we listen to the machines that power our world, turning a chaotic swirl of sound into a clear, actionable map of where the noise is coming from.