Commonly Used Feature Extraction Methods for Acoustic Radiation Noise

Acoustic radiation noise feature extraction is a critical step in acoustic signal processing, widely used in fault diagnosis, target recognition, and related fields. The following implementation approaches cover several commonly used methods:

Waveform Envelope Features This method characterizes the macroscopic fluctuation characteristics of noise by extracting signal envelopes. Typically implemented using Hilbert transform to obtain analytic signals, followed by calculating their modulus as the envelope. Statistical measures like mean and variance of the envelope can reflect the energy distribution characteristics of noise. In MATLAB, this can be achieved using the hilbert() function combined with abs() operations.

Higher-Order Spectral Features Utilizes bispectrum or third-order cumulant analysis to examine nonlinear phase information while effectively suppressing Gaussian noise interference. Higher-order spectra capture higher-order statistical properties in signals, making them suitable for identifying non-Gaussian distributed noise components. Implementation often involves functions like bispecd() or custom cumulant calculation algorithms.

Fractal Dimension Features Quantifies the self-similarity of noise waveforms through box-counting dimension or Hurst exponent analysis. Fractal dimensions characterize signal complexity, such as irregularity changes in noise caused by mechanical equipment wear. Code implementation typically involves box-counting algorithms or rescaled range analysis (R/S) methods.

Singularity Exponent Features Based on wavelet transform modulus maxima analysis to examine signal singularities, using Lipschitz exponents to describe the sharpness of突变 points. Particularly suitable for transient feature extraction of impact-type noise (e.g., bearing cracks). Implementation requires wavelet decomposition followed by modulus maxima detection and Lipschitz exponent calculation.

These methods can be used individually or combined, with feature dimensions selected according to noise types, ultimately forming distinctive feature vectors for classification or analysis purposes.