Rotating sound sources located without propagation functions: 16 microphones, a prediction in under 50 milliseconds

Rotating source localization now lets a residual network read the source distribution directly from a cross-spectral matrix, without prior propagation functions or microphone positions. Earlier time-domain ROSI and frequency-domain MCB both fail to fully remove ghost contributions, spurious sources that appear alongside real ones, under unconstrained array geometry. In a simulation comparison with 16 microphones, 14 modes, 1500 rpm and 5000 Hz, MCB identified only three of five sources, ROSI produced an extra predicted source, and ResNet matched the true source count, positions and intensities.

The network input is not a conventional cross-spectral matrix but one built in MCB form, stacking each microphone spectrum after frequency shifts by the rotational angular frequency. Conventional cross-spectral matrices suit stationary sources and degrade under rotation, so an input representation that embeds rotational information is needed. The matrix is formed from 16 microphone signals and 14 acoustic modes, its real and imaginary parts are standardized separately, and ResNet18 outputs a one-dimensional intensity vector over a 10×10 scanning grid.

Training set size and network depth jointly set accuracy, and more depth does not compensate for too little data. With only 83,200 training samples, ResNet18 reached 52.43% accuracy because the joint frequency-rotation parameter space was insufficiently covered. Raising training samples to 123,200 lifted ResNet18 to 89.32%, and ResNet34 reached 94.31% at the same scale; under non-uniform intensity and a perpendicular array, ResNet34 reached 98.83% with 711,000 training samples.

Results rest on numerical simulation: source frequencies of 4500 to 5000 Hz, rotational speeds of 500 to 2500 rpm, 2 to 6 sources, 16 microphones, 14 modes, and a 10×10 scanning grid. Both parallel and perpendicular array placements were tested; in the perpendicular case ResNet18 and ResNet34 reached 98.66% and 98.83% with 711,000 training samples. When the array configuration matches training and all parameters fall within the randomized ranges, the model predicts within 50 milliseconds.

Generating training data remains the most time-consuming stage, taking possibly weeks on a personal computer and hours with high-performance computing; single-precision storage only halves data volume. Models are strongly specialized, so changes in microphone array configuration or in the array-source plane relation require new simulated data. The physical mechanism behind the network's predictions remains opaque. If the scanning frequency differs from the true frequency by an integer multiple of the rotational frequency, the heatmap may report non-existent sources.