A Novel Feature Representation Method Based on Distances Between Statistical Matrices of Acoustic Emission Waveforms

Acoustic emission (AE) has gained increasing popularity in non-destructive testing and process monitoring due to its high sensitivity to malfunctions. However, valuable information is sometimes overwhelmed by high levels of background noise, making it quite difficult to recognize the malfunctions. Therefore, it is essential to address the issue of identifying weak emission sources and obtaining an accurate representation of the original waveform. This study presents a novel feature representation method involving the calculation of distances between the statistical matrices of the original AE waveforms. In the proposed method, the waveform is considered and treated as a connection of numerous harmonics from a micro perspective. Consequently, only the amplitude and half-period of the harmonics are extracted as features to represent them, which means that data can be greatly reduced in size. Building on this, statistical matrices are proposed and defined to count the combination of amplitude and half-period to further characterize the statistical properties of the signals. Subsequently, to differentiate between statistical matrices, distances between them are calculated based on matrix norms. A standardized procedure for calculating distances between statistical matrices has been formulated. Tests conducted to detect filament absence under both single and multiple conditions, as well as for detecting warping deformation, validate the feasibility and effectiveness of the proposed method. The influence of key parameters, including segment length and grid size, is discussed. This method offers an alternative AE feature representation approach applicable in practical scenarios for process monitoring and fault diagnosis.