Study on robust model construction method of multi-batch fruit online sorting by near-infrared spectroscopy

In the online detection of fruit samples by near-infrared spectroscopy (NIRS), the natural change of sample states or the variations of instruments will often cause a large error in predicting different batches of samples. In this study, a total of 440 tomato samples were collected in four batches with each batch of 110 samples. The Spectral and soluble solids content (SSC) of single batch were collected every other day in batch order. The multivariate statistical process control (MSPC) method was adopted to establish a stability monitor model. The robustness regression (Rob-Reg) and partial least squares regression (PLSR) were used for mixed modeling of multiple batches of samples to eliminate the variability influence of sample and instrument states. The results show that MSPC can effectively monitor the consistency of the same batch samples measured at different times or different batches. The variation of sample attributes with spectral acquisition time has dramatically damaged the adaptation of PLSR models. The Rob-Reg method can predict the SSC of the different batches of samples at different collection times. Compared with the PLSR method, the correlation coefficient of prediction (Rp) was improved from 0.61 to 0.66, and the root mean square error of prediction (RMSEP) was decreased from 0.55 to 0.44 for Rob-Reg method. The RPD of 3.85 indicated that the model is excellent. The Robust modeling method can be well applied to fruit near-infrared online detection system.

Automated summary

In this study, a stability monitor model was established using the multivariate statistical process control (MSPC) method, and a mixed modeling approach combining robust regression (Rob-Reg) and partial least squares regression (PLSR) was employed to eliminate the variability influence of sample and instrument states. The results showed that MSPC effectively monitored the consistency of the same batch samples measured at different times or different batches, and the Rob-Reg method outperformed the PLSR method in predicting the different batches of samples.