基于近红外光谱分析贡梨可溶性固形物全局模型的鲁棒性

贡梨是大众喜爱的水果，为研究不同检测方向对近红外在线检测贡梨可溶性固形物SSC的影响，提出全局模型并分析其鲁棒性。在贡梨六个方向上收集光谱：茎-花萼轴垂直，茎向上(A1)和茎向下(A5),茎-花萼轴和水平之间45°,茎向上倾斜(A2)和茎向下倾斜(A4),茎-花萼轴水平，茎朝向右侧光(A3),茎花萼轴水平，茎朝向带移动方向(A6)。SSC范围为9.53～14.70的150个样品分为115个标准偏差为1.05的校准集和35个标准偏差为0.93的预测集。采用偏最小二乘回归PLSR分别建立六个局部模型和一个全局模型，局部模型由各方向的115个校正集数据经过Savitzky-Golay卷积平滑、多元散射校正MSC、高斯滤波平滑GFS三种不同的预处理方法处理后使用偏最小二乘回归PLSR建立而来；用本方向校正集数据建立的局部模型验证本方向的35个预测集数据，比较这三种预处理方法后所建立的PLSR模型，结果表明经过GFS处理后建立的模型验证效果最好，因此六个局部模型和全局模型均采用GFS处理后建立的PLSR模型。全局模型是由A1, A2, A3, A4, A5和A6六个方向的690个校正集光谱数据经...

Robustness of Global Model of Soluble Solids in Gongli Pear Based on Near-Infrared Spectroscopy

Gongpear is a popular fruit. In order to study the influence of different detection directions on the online detection of soluble solid SSC in Gongpear by NIR, a global model was proposed, and its robustness was analyzed. The spectra were collected from Gongpears in six directions: stem-calyx axis vertical, stem-upward (A1) and stem-downward (A5), between the stem-calyx axis and horizontal 45°, stem-upward sloping (A2) and stem-downward sloping (A4), stem-calyx axis horizontal, stem-right light oriented (A3), stem-calyx axis horizontal, stem-band moving direction (A6). The 150 samples with SSC ranging from 9.53 to 14. 70 were divided into 115 calibration sets with a standard deviation of 1.05 and 35 prediction sets with a standard deviation of 0.93. Six local models and one global model were established by partial least-squares regression (PLSR). The local models were established by partial least-squares regression (PLSR) after 115 calibration sets of data in each direction were preprocessed by Savitzky-Golay convolution smoothing, Multiple Scattering Correction（MSC）and Gaussian Filtering Smoothing (GFS). The local model established by the local correction set was used to verify the data of 35 prediction sets in the local direction. Compared with the PLSR model established by the three pretreatment methods, the results showed that the model established by GFS processing had the best validation effect. Therefore, the PLSR model established by GFS processing was used for all the six local and global models. The global model is a Gongpears SSC model established by PLSR after GFS pretreatment from 690 calibration sets of spectral data in A1, A2, A3, A4, A5 and A6. The prediction sets in each direction verified the seven models. The verification results showed that although the prediction effect of the local model was stronger than that of the global model in the local direction, it could not be verified in other directions and the robustness was poor. Therefore, different detection directions had a great influence on the prediction effect. The global model can accurately predict the SSC of Gongpears pear in each detection direction. The global model’s correlation coefficient （R_c）is 0.828, and the root mean square error RMSEC is 0.424. The correlation coefficients （R_p） of A1, A2, A3, A4, A5 and A6 prediction sets were 0.818, 0.765, 0.799, 0.821, 0.794 and 0.824, and the root mean square errors RMSEP were 0.446, 0.525, 0.478, 0.538, 0.486 and 0.619, respectively. The R_p and R_c in six directions are close to each other and are around 0.800, the RMSEC and RMSEP are around 0.500. The results show that the global model has excellent robustness in detecting gongpear SSC in different directions.