Freshness evaluation of sea bass using multi-sensor information fusion based on olfactory visualization and NIR spectroscopy technique

Abstract: Total volatile basic nitrogen (TVB-N) is an important reference index of fish freshness. This study attempted to measure TVB-N content of sea bass using multi-sensor information fusion based on olfactory visualization and near-infrared spectroscopy technique. Sea bass samples were stored under the condition of 4℃ in refrigerator. The total number of samples was 270, among them 18 random samples every day were firstly detected by olfactory visualization detecting instrument, and then by near-infrared spectrometer. TVB-N content of these samples was measured according to the kjeldahl method. The experiment finished after 15 days because of the serious corruption of samples. Two-thirds of total samples were chosen as calibration set and the remaining samples were taken as prediction set using SPXY (sample set partitioning based on joint x-y distances) algorithm. So the sample sizes of calibration set and prediction set were 180 and 90 respectively. After sample division, a TVB-N prediction model was established based on the fused multi-sensor information. Other two models were also established based on the single-sensor information for comparing. Principal component analysis (PCA) method was used to reduce the dimension of colorimetric sensor array data. The results of PCA showed that the cumulative contribution rate of the first three principal components was 85.441%, which indicated that the first three principal components had been able to explain the vast majority of the overall information about original samples. Based on the first three principal components, a TVB-N prediction model was established by support vector regression (SVR) algorithm. The determination coefficients of calibration set (R2 c) and prediction set (R2 p) of the model were 0.762 and 0.757 respectively, while the root mean square errors of calibration set (RMSEC) and prediction set (RMSEP) were 6.012 and 6.755 mg/100g respectively. Mean centering (MC) method was used to preprocess the raw near-infrared spectrum. After preprocessing, genetic algorithm (GA) combined with partial least squares (PLS) method was carried out on the near-infrared spectrum data to remove irrelevant information as well as simplify the prediction model. The results showed that characteristic variables of near-infrared spectrum reduced from 1557 to 79 after the optimization of GA-PLS. Meanwhile, the root mean square error of cross validation (RMSECV) reduced from 12.763 to 6.585, which indicated that the remaining variables had higher correlation with TVB-N content than the original variables. Based on the informative near-infrared spectrum data, another TVB-N prediction model was established by SVR algorithm. The R2 c and R2 p of the model were 0.810 and 0.787 respectively, while the RMSEC and RMSEP were 5.385 and 6.186 mg/100g respectively. Because of the insufficiency in getting freshness information, performance of the two single-sensor models above was unsatisfactory. To improve the predictive accuracy of TVB-N content, the colorimetric sensor array data after dimension reducing and the informative near-infrared spectrum data were fused, and a multi-sensor information fusion model was established based on the fused data. The R2 c and R2 p of the model were 0.893 and 0.882 respectively, while the RMSEC and RMSEP were 4.032 and 4.585 mg/100g respectively. Compared with other two single-sensor models, the R2 p of fusion model increased by 0.095 and the RMSEP reduced by 1.601 mg/100g, which proved the superiority of fusion model. This study shows that multi-sensor information fusion based on olfactory visualization and near-infrared spectroscopy technique can be a feasible method for the evaluation of sea bass freshness.