Detection Of Multicomponent Content In Wheat Flour Based On Near Infrared Spectroscopy

The content of protein,moisture and ash are important index for wheat flour quality evaluation.In this paper,the rapid detection method of multicomponent content in wheat flour was studied by near infrared spectroscopy.100 wheat flour samples are collected and the physicochemical values and near-infrared spectral data are measured.The spectral data are preprocessed and divided into three sample sets.Adaptive Extreme learning machine(ELM)and variable convolution one-dimensional convolution neural network(1D-CNN)models are proposed to improve modeling efficiency and prediction accuracy.The main research contents of this paper are as follows.(1)An adaptive ELM modeling method is proposed to improve the stability and prediction accuracy.Particle swarm optimization(PSO)is used to optimize the initial weight and bias.The model accuracy,over fitting degree and running time are comprehensively evaluated based on analytic hierarchy process(AHP)to determine the optimal number of ELM hidden layer nodes.The determination coefficients(_PR~2)of adaptive ELM model for protein,moisture and ash in wheat flour are 0.9115,0.9192 and 0.7252 respectively,which are significantly higher than those of traditional ELM which are 0.8330,0.8678 and 0.6755.However,the prediction accuracy of ash content is relatively unsatisfactory,and its average content is only0.9%.(2)The convolutional neural network(CNN)algorithm of deep learning is applied to the modeling of near-infrared spectral data,and an optimized 1D-CNN model with stronger feature extraction ability is proposed.The classical 2D-CNN model structure is transformed into 1D-CNN which is more suitable for spectral data.The optimized 1D-CNN model is established by simplifying the 1D-CNN structure,wavelet compression with L2 and Early Stopping compound regularization methods.The _PR~2 of protein,moisture and ash are 0.9546,0.9570 and 0.8450 respectively,which improves the prediction accuracy compared with adaptive ELM,in particular,the prediction accuracy of ash with less content has been greatly improved.(3)An 1D-CNN model with variable convolution kernel is established,which makes full use of spectral features to simplify modeling and improve modeling efficiency and prediction accuracy.The spectral derivative method is used to extract each absorption peak of wheat flour spectrum.The larger depth convolution kernel is used to extract the characteristic information.There is less spectral data in each segment,and the number of hyper-parameters combinations of variable convolution kernel 1D-CNN is reduced to 83.33%compared with optimized 1D-CNN leading to the modeling time greatly shortened._PR~2 of protein,moisture and ash in wheat flour are increased to 0.9734,0.9602 and 0.8486 respectively,the modeling efficiency and accuracy are improved.