| China is the largest consumer of pork in the world,and the total production reached 85.4 million tons in 2016.With the improvement of living standards,people propose more requirements for the quality of meat.However,the mixed quality of pork in the market and the phenomenon of "water injection" have seriously affected the health of consumers and the development of pork industry.Traditional detection methods have drawbacks as they are time-consuming,unable to realise real-time detection,and hence cannot satisfy the practical requirements for quality monitoring.Hence,nondestructive,rapid and real-time analytical techniques and practical and efficient detection device are in urgent required to improve the quality control level of pork industry.In this paper,the quality attributes including color,pH,tenderness,water holding capacity,fat,protein,water content,and total volatile base nitrogen(TVB-N)of pork meat were used as detection index.Based on the visible and near-infrared spectroscopy technology,the influence of spectral signal preprocessing,anomalous samples elimanation,and feature wavelengths selection on the modelling results were discussed.A fusion method for dual-band spectra and adaptive model updating method were proposed to improve the accuracy and applicability of the models.On this basis,portable and online detection device were developed,which were suitable for different occasions.Predictive models for multiple quality attributes were established,and experimental verifications were also performed.The main contents and results of the paper were summarized as follows:1.The effects of spectral signal preprocessing,anomalous samples elimanation,and characteristic wavelength selection on the modeling results were studied to improve the model stability.Savitzky-Golay smoothing,derivative,and standard normal variable transformation were used individually or in combination to explore their influence on the modeling results.The influence of anomalous samples on the model accuracy of the model was discussed,and Mahalanobis distance and Monte Carlo method were compared.Then feature wavelengths selection methods including competitive adaptive reweighted sampling algorithm(CARS),Random Frog algorithm,Monte Carlo-uninformative variable elimination-successive projections algorithm(MC-UVE-SPA)and two-dimensional correlation spectroscopy were compared.The results show that the optimal spectral pretreatment methods were not always the same for different parameters to be predicted.Anomalous sample elimanation can help improve model performance and predictability,and the Monte Carlo method has greater advantages than Mahalanobis distance.The CARS method has more obvious advantages in stability and prediction accuracy than other methods.2.A fusion method for dual-band spectra was proposed to deal with the limited detection capability of single-band spectra and the inability to capture dual-band continuous spectra simultaneously.Based on visible-near-infrared spectroscopy(350?1100 nm and 1000?2500 nm)and(350?1000 nm and 950?1700 nm),two sets of hardware systems were constructed.Direct translation method,regional feature weighting fusion method,linear fitting method and parabola fitting method were used for spectral fusion.Then the above methods were compared from the fusion effect of spectral curve and modeling results.The results show that the characteristics of spectroscopic instruments should be considered when fusing dual-band spectra.For the spectral fusion of 350?1100 nm and 1000?2500 nm,the parabola fitting method yielded better result with the correlation coefficient in the prediction set(Rp)for TVB-N of 0.9212 and 0.9487 before and after spectral fusion.For the dual-band spectra of 350?1000 nm and 950?1700 nm,the regional feature weighted fusion algorithm performed better with Rp of 0.9085 and 0.9227 for TVB-N before and after spectral fusion.3.An adaptive model updating method was proposed to improve the prediction ability of established calibration model for different batches of samples.The method had two functions:supplementing and updating the basic data for modeling,and establishing online prediction updating model for external tested samples.Different batches of pork samples were used as the research objects,and the neighborhood distance was employed to decide whether to add the sample into the calibration set to update the basic data.A similarity function was built based on the Euclidean distance and the spectra information divergence-spectra angle metric,and a similarity factor was determined to select the modeling neighborhood for local model establishment of TVB-N.Compared with the results before establishing the updated model,the correlation coefficient R increased from 0.8365 to 0.9291,and the prediction error decreased from 4.3191 mg/100g to 2.6120 mg/100g,which improved the prediction accuracy of the model.4.A portable device for simultaneous detection of multiple parameters of fresh pork quality is developed,and optimal models for quality parameters were established,which realized real-time detection of pork.Based on dual-band spectrometers,the effect of distance between light source and detector was discussed.A point light source and ring light source were designed and compared based on the coefficient of variation,signal-to-noise ratio,and spectrum area change rate.The results show that the ring light source had higher stability and the spectra could provide more comprehensive information for meat.Based on the light source,the hardware unit and the software control program were developed,and the dual-band spectral data could be acquired,fused,processed,displayed,and saved in real time.Prediction models for(L*,a*,b*),pH,TVB-N,fat,protein,water,tenderness and cooking loss were then established with Pp of 0.9347,0.9123,0.9625,0.9312,0.9493,0.9505,0.9510,0.9271,0.9001,and 0.9177.The results showed that the developed portable device could realise on-site and real-time detection of multiple quality attribtes of meat with detection speed of 2 seconds/sample.5.An online detection device for multiple quality parameters of pork quality was developed,and predictive models for multi-parametes were established to realize non-destructive detection of pork quality.Two types of light source systems,namely direct-illuminated halogen lamp and straight-tube annular light guide,were designed.In terms of light intensity uniformity and stability of the acquired spectra dynamically,the latter was found to have better performance.The effect of variable length and thickness of samples on prediction accuracy was eliminated to ensure that the spectra of the middle area can be acquired at the optimal distance.Predictive models for color(L*,a*,b*),pH,TVB-N,fat,protein,water,tenderness and cooking loss were established with Rp of 0.9644,0.93 10,0.9025,0.9235,0.9225,0.9608,0.9472,0.8923,0.9452,and 0.9125.It showed that the developed device could realise real-time detection of multiple quality attributes with detection speed of 3.5 seconds/sample.This paper provides a new technology and method for simunateous and real-time detection of multiple quality parameters of pork meat.The accuracy and applicability of prediction models for pork were improved,and the development of practical detection devices enable the simultaneous and real-time detection of multiple quality parameters of pork. |
PDF Full Text Request