Walnut crack detection based on EEMD and acoustic feature optimization

Yunnan walnut crack detection based on an acoustic method was investigated with a focus on the pre-processing of acoustic signals and feature optimization. The acoustic signals were generated by the impact of walnuts on an impact plate, and acoustic signals at different impact locations were collected by observing the walnut impact points with a high -speed camera. Ensemble empirical mode decomposition (EEMD) was performed on original signals, then effective components were selected for signal reconstruction. Time domain, frequency domain and information entropy features were extracted and the method customized by previous scholars were extracted. The distinguished index (D.I) method was used to evaluate the ability of features to distinguish between intact and cracked walnuts, different threshold parameters were set to screen different feature combinations, and the correlation between the features was investigated. Features with significantly low D.I values, and features with low D.I values in highly correlated feature groups were eliminated. The classification accuracy of 99.17% was achieved by using the arithmetic optimization algorithm (AOA) optimized least squares support vector machine (LSSVM) model. EEMD preprocessing and feature dimensionality reduction help to simplify the classification algorithm and reduce the computation effort for the online system.