Research On NIR Model Updating Method With Application In Food Detection

The food industry is vital to human beings.Its development level is an important index to reflect people's living standards and social development.With the continuous development of automation,the food production in China has completed the transformation from manual workshop production to mechanized production.However,the major detection method is still offline.Food detection is vital to production and quality control.For online production adjustment,new food detection methods which can obtain real-time results are urgently needed.Recently,near-infrared(NIR)technology has been widely used in quality detection of agricultural products and food products due to advantages like short analysis time,no sample pretreatment,non-destructive,no pollution and low cost.The performance of NIR detection is highly determined by model quality.However,in food production,the model is often influenced by the varying environment and raw materials.So,it is often necessary to remodel or update the model.The objective of this paper is to apply NIR online detection technology to food detection by improving the adaptive ability of NIR model,reducing the maintenance cost,intelligentizing production,and optimizing food production.The key research points are as follows:(1)In food detection,NIR model performance will gradually deteriorate due to the varying environment and raw materials.In order to maintain the accuracy of NIR model,it is necessary to include new sample information into model continuously.The recursive partial least squares(RPLS)is a widely used model updating algorithm.In NIR detection,RPLS always employs full spectrum for modeling,which is not only inefficient,but also easy to be influenced by non-target factors.This paper proposed a new updating algorithm based on RPLS with characteristic wavelengths.The proposed algorithm is applied in the detection of Chinese yellow wine and wheat grain and significantly improves detection accuracy.(2)In food detection,the initial database should cover as many varieties as possible,so as to enhance the robustness of NIR model.However,local model based on similar samples may perform much better than global model in the quality prediction.This paper introduces just-intime(JIT)framework into NIR detection and proposes a novelty similarity calculation method.What's more,the maximum information coefficient(MIC)is employed to carry out more accurate similarity measurement.The performance of NIR model is effectively improved by applying the proposed algorithm to ethanol detection of Chinese yellow wine.(3)Furthermore,considering that both recursive algorithm and JIT can improve the prediction accuracy,we integrate RPLS into the JIT framework to deal with the two problems simultaneously.In addition,the traditional moving window method may gradually forget the initial database,which is essential to model robustness.Thus,this paper proposed a modified recursive locally weighted partial least square(m-RLWPLS)algorithm to balance adaptability(the introduction of new information)and stability(maintain the initial database integrity)simultaneously.Meanwhile,MIC weighed spectrum instead of the original spectrum is employed in the JIT part.All the improvements increase the prediction accuracy observably.