Analysis on Food Crispness Based on Time and Frequency Domain Features of Acoustic Signal

Crispness is an important indicator of crunchy food. However, it cannot be easily quantified by sensory evaluation, due to the high subjectivity of evaluators; instrument measurement of this indicator requires much manpower and time. To improve the efficiency of food crispness prediction, this paper attempts to build a rapid, convenient, and accurate crispness analysis model. Starting with the fracturing sound of crunchy food, the authors collected the fracturing acoustic signal, conducted wavelet denoising, analyzed the eigenvalues in time and frequency domains, and constructed crispness prediction models based on multiple linear regression (MLR) and neural network, respectively. Through fracturing test and acoustic test, cluster analysis was adopted to select the typical eigenvalues of acoustic signal, including the peak amplitude of power spectral density (PSD) curve, amplitude difference, and waveform index. Based on these eigenvalues, a crispness analysis model was established, and used to predict the crispness of four kinds of food, namely, potato, sweet potato, carrot, and turnip. The results show that the BP neural network had a smaller relative error than the MLR; when the threshold was 5%, the BP neural network maintained a prediction accuracy of >90%, and achieved 100% prediction accuracy on two types of food. To sum up, this paper reveals the relationship between the food chewing sound features and food quality, laying the theoretical basis for the research of food chewing sound mechanism.