Research On Nondestructive Testing Of Product Quality Based On Hyperspectral Imaging Technology

The quality of product in various fields has a great influence on the production,processing,consuming and export.Traditional quality detection methods,such as expert evaluation and physical and chemical experiments,are time-consuming,inefficient and destructive,and are no longer suitable for applying in large-scale production.Therefore,rapid,accurate and non-destructive qualitative and quantitative research on the internal and external quality of products has become a hot development trend in the field of scientific research.Hyperspectral imaging technology and its data processing methods were studied in this thesis and they were applied to analysis the quality of the product qualitatively and quantitatively.Firstly,the theoretical knowledge of data processing methods was studied,and a series of effective hyperspectral data processing methods and models were constructed.By analyzing interference information such as noise with different characteristics and migration in hyperspectral data,spectral correction was performed by multivariate scatter correction and convolution smoothing methods.To solve the problem of large dimension and strong correlation of hyperspectral data,several optimization methods of characteristic variables were summarized to realize the compression of spectral data.The principle and applicability of multivariate statistical methods were summarized and analyzed,and introduced into the process of hyperspectral analysis and identification to meet the application requirements of different objects in different fields.Secondly,the non-destructive qualitative analysis of two kinds of samples(wheat grain and paraffin wax)with obvious differences in appearance,internal characteristics and sample size was conducted by using the hyperspectral imaging technology combined with the above theoretical research.For the paraffin samples,the feature selection method of genetic algorithm-partial least squares(GA-PLS)and the feature extraction method of principal component analysis(PCA)were used to reduce the dimension of the data respectively,and the influence of spectral data dimension reduction method on the accuracy of the model recognition was studied.The qualitative analysis results of paraffin grade based on support vector machine(SVM),probabilistic neural network(PNN)and extreme learning machine(ELM)recognition models were compared and analyzed.The results show that the ELM model constructed by using PCA to reduce the dimension of the data is the best one for detecting the level of paraffin,with an accuracy rate of 91.7%.To solve the problem that the conventional data processing methods were ineffective in identifying grainy and highly similar wheat seeds,a new feature mining and optimization method was proposed in this thesis.Experiment shows that this method can fully mine the effective features of hyperspectral data and improve the recognition accuracy.Finally,the non-destructive quantitative analysis was conducted on samples(apples)by using hyperspectral analysis method.Due to the non-specificity,interaction sensitivity and complexity of the components affecting apple taste information among different bands of apple hyperspectral data,the screening of characteristic wavelengths of hyperspectral data was difficult in the establishment of the association model,and the detection effect was not good.By analyzing the characteristics of the data,a competitive adaptive weighting method(CARS)was proposed to eliminate redundant information from hyperspectral data,and 43 characteristic wavelengths corresponding to acidity and 22 characteristic wavelengths corresponding to sweetness were selected.A support vector regression model(PSO-SVR)based on particle swarm optimization was established.The results show that this technology can realize the quantitative prediction of sourness and sweetness of the apple.The correlation coefficients(R~2)of sourness and sweetness are 0.81 and 0.887 respectively,and the predicted root mean square error(RMSEP)are 0.03 and 0.018 respectively.In summary,combined with practical application,this thesis used hyperspectral imaging technology to propose a targeted technical means for different spectral data characteristics of different samples,and the non-destructive qualitative and quantitative analysis of product quality was finally realized.