Measurement Method Of Food Crispness Based On Acoustic Signal

Crispness is key quality attribute used to assess the quality and acceptability of fresh or processed food products,it can reflect the freshness,mouth feel and maturity of food.However,the evaluation method of crispness is often used by sensory evaluation,which is limited in food evaluation.In this paper,the food samples were compressed by texture analyzer and the acoustic signal from food fracture was collected.After de-noising,features of the useful signals by time domain analysis and frequency domain analysis were extracted;predict model was constructed by BP neural network method.The graphic use interface was last developed to improve the food crispness evaluation.A quick,convenient and economic method of food crispness evaluation was provided by this paper.1.Selection of test materials,method and parameters optimizationAccording to the purpose of experiment,selected samples,food breaking and voice acquisition device is designed.The test parameters were optimized by single factor tests.The parameters included the distance between food and sensor,probe selection,test mode,test speed,test distance,trigger point load,sample pretreatment method,etc.Through the experiment,the distance between food and sensor was 4 cm,probe was TA7 shear blades,test mode was single compression test mode,test speed was 1 mm/s,test distance was 6 mm,trigger point load was 0.1 N,the length,width and height of sample was 2cm*1cm*1cm,the movement direction of probe was vertical to the sample fiber;the samples were dried by oven in order to set the gradient of crispness.The drying temperature was 60 ?,time level for 5 min and 10 min.Finally,10 moisture gradient samples for potatoes,8 moisture gradients for pachyrhizus,7 moisture gradients for carrot,7 moisture gradients for white radish were obtained.The characteristic parameters extracted from mechanics curve of fracture(the first peak),curve length,curve and the coordinate axis in area,the ratio of highest peak(hardness value)with the average peak were used to analyzed the correlation with the sensory value,their correlation coefficients were 0.98?0.89,0.84 and 0.98,respectively.Therefore,the fracture was chose to evaluate food crispness as characterization parameters.2.The acquisition and denoising of acoustic signal of food fractureSimultaneously with the compression,the sounds emitted during fracturing were recorded using a microphone at a sampling rate of 44,100 Hz and then were digitized with a 16-bit analog-to-digital converter on a PC sound board.The sound colleting software was Adobe Auditon 3.0.Wavelet and spectral subtraction method were used to de-noise.The default threshold denoising method was used by Wavelet.The default threshold of the signal was obtained by the Matlab ddencmp function.8 layers wavelet decomposition of original sound signal were completed by the db2 wavelet base.For the spectral subtraction method,the first 20 frames were selected as the 'silence' frame.The indexes of SNR and RMSE were chose to compare.The results showed that spectral subtraction was superior to the wavelet denoising,by spectral subtraction denoising,SNR can reach 17.14-37.07db.3.Extraction and analysis of acoustic signal in food fracture(1)In the time domain,the 6 acoustic features were selected to analyze including the signal strength,maximum differential short-time frame energy,amplitude difference,pulse factor,waveform index and decay time.The signal intensity,maximum differential short-time frame energy,amplitude difference and waveform showed good correlation with fracture.(2)Acoustic features in frequency were extracted by FFT.The Welch method was utilized to estimate the power spectral density of acoustic signal in food breaking.The Power spectral density characteristics,the maximum peak of power spectral density,the frequency values of the maximum peak were extracted to analyze the correlation with the food fracture.The results showed that power spectral density characteristics and maximum peak of power spectral density had good correlation with the crispness of food.(3)Acoustic features in frequency were extracted by FFT HHT.According to the analysis,the energy of first 8 IMFs that got by EMD had good correlation with the crispness of food.So the first 8 IMFs were extracted by marginal spectrum analysis.The marginal spectrum characteristics,maximum peak of marginal spectrum,the frequency values of the maximum peak,were extracted to analyze the correlation with the food fracture.The results showed that marginal spectrum characteristics and maximum peak of marginal spectrum had good correlation with the crispness of food.4.Selection,modeling and food crispness prediction of acoustic features in food fracture(1)The features of time domain,FFT frequency domain and HHT frequency domain by Hierarchical clustering method were extracted.In time domain and FFT frequency domain,the maximum power spectral density,amplitude difference and waveform indicator were selected;In time domain and HHT frequency domain,marginal spectrum characteristics,amplitude difference and waveform indicator were selected.(2)Multiple linear regression models were built based on the characteristic values of time domain-FFT frequency domain and time domain-HHT frequency domain to predict the crispness of four kinds of vegetables.The results showed that the models of potato,sweet potato,carrot,white radish,potato-sweet potato,carrot-white radish were ideal,the average relative error was less than 5%;the prediction results were not ideal when synthesized the four kinds of vegetable,the average relative error was more than 5%.The average relative error of predicted results based on the feature values of time domain and HHT frequency domain were less than the average relative error based on the feature values of time domain and FFT frequency domain.(3)Models were built based on the characteristic values of time domain-FFT frequency domain and time domain-HHT frequency domain with BP neural network to predict crispness of 4 vegetables.The predicted results showed that the prediction results of potato,sweet potato carrot,white radish,potato-sweet potato,carrot-white radish were ideal;the average relative error was less than 5%.The prediction model of four kinds of vegetable was not ideal,the average relative error less than 5%.The mean errors of feature values based on time domain and HHT frequency domain predicted results were less than the FFT.(4)The 2 analysis methods of FFT and HHT and the 2 modeling methods were compared.The prediction results of BP neural network were better than multivariate linear regression for the single sample model or the synthesized model.The predicted results based on time domain-HHT frequency domain are better than the results based on time domain-FFT frequency domain.When the BP neural network was built based on time domain and HHT frequency domain and the threshold was set to 5%,the predict accuracy of other three samples could reach 100%,but the predict accuracy of carrot was 90%.Therefore,the BP neural network of acoustic features based on time domain and HHT frequency domain were used to predict crispness of the four kinds of vegetables.5.Test system for vegetables crispness by the MATLAB was developed.Interface was mainly divided into four modules,including wave image,signal analysis,characteristic value and crispness forecast,and the FFT,HHT and BP neural network technology were combined to achieve reading,pre-treat,time-analysis,frequency-analysis and feature extraction from the acoustic signal.The crispness prediction of 4 vegetables was eventually detected.