Acoustic spatial patterns recognition based on convolutional neural network and acoustic visualization.

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Steady-state analysis of acoustic problems is complex.
• Acoustic spatial patterns can be visualized and analyzed.
• Convolutional Neural Networks (CNNs) excel at image processing tasks.

Purpose of the Study:

• To apply CNNs for recognizing acoustic spatial patterns.
• To explore CNNs' capability in acoustic image processing.
• To develop a robust method for fault diagnosis using acoustic measurements.

Main Methods:

• Acoustic spatial patterns were obtained using singular value decomposition (SVD) of an acoustic radiation operator.
• Acoustic fields (real and imaginary pressure components) were rendered into images.
• CNNs were trained with numerically synthesized data, including simulated noise and incident waves.

Main Results:

• The CNN approach demonstrated more accurate and robust recognition of acoustic patterns than cross-correlation methods.
• Higher resolution acoustic images were achieved through pressure interpolation.
• The CNN scheme effectively processed two-channel acoustic field data.

Conclusions:

• CNNs offer a promising approach for fault diagnosis and condition monitoring based on spatial acoustic measurements.
• The hierarchical feature representation and nonlinear perception of CNNs enhance pattern recognition capabilities.
• The method shows robustness against uncorrelated and correlated noise.