Application Research Of Neural Network BP Algorithm In Food Safety

Edible synthetic pigments were widely used in pastry,jelly beverages,food industries and often mixing used,it was favor by the merchants and consumers due to it had bright color.good stability,strong tinting strength,suitable for color mixing,etc.However,with the research in-depthly,it found that most of synthetic pigments hazarded to human health,and even may cause cancer,birth defects and other serious consequences.In addition,with the cases of illegal to add food pigment disclosed by the media constantly,it caused the extensive concern of consumers and the governments pay close attention to the problem.Due to Backpropagation Artificial Neural Networks(BP-ANN)has strong computing ability to deal with many complex problems,it could sum up quantitative of the rule from the complex data set,and could be used to quantitatively predicted the result of unknown samples.This thesis mainly researched to 6 kinds of improved methods,such as standard backpropagation,momentum backpropagation,variable leamling rate backpropagation,resilient backpropagation,levenberg-marquardt,Bayesian regularization.And the improved methods were used to simultaneous forecast of 4 kinds of synthetic pigment mixed system consist of amaranth,sunset yellow,lemon yellow and carmine respectively,the calculation results were compared,and the method were as follows:Firstly,it selected the data pretreatment method and the transfer function,and principal component analysis was carried out on the data,set the training precision,trained and forecasted the network for the 6 kinds of improved BP algorithm,compared the relative prediction error for all components of the output.Secondly,through cloosen the hidden layer node number,learning rater,momentum constant,the increase of update value for each weight and bias by a factor delt_inc,the decrease of update by a factor delt_dec,the parameter mu of the initial value to optimize the network structure,trained and forecasted the network for the 6 kinds of improved BP algorithm,compared the relative prediction error for all components(RPEt),and the relative prediction error for a single component in muxtures(RPEs)of the output.Test results showed that the speed of standard backpropagation,momentum backpropagation and variable learnling rate backpropagation is slow,the RPEt value of them is high;but the speed of Resilient Backpropagation,Levenberg-Marquardt,Bayesian regularization Respectively is faster,and the RPEt of them was able to achieve small values,its network structure designed and optimized results were as follows:1.Resilient Backpropagation:the result of the RPEt is 0.28102,and the RPEs is:0.39733,0.21557.0.36865.0.34119 respectively.When the training step number was 97,it could be achieved to the accuracy.2.Levenberg-Marquardt:the result of the RPEt is 0.24501,and the RPEs is:0.55423,0.29088,0.53998,0.24162 respectively.When the training step number was4.it could be achieved to the accuracy.3.Bayesian Regularization:the result of the RPEt is 0.2298,and the RPEs is:0.24438,0.19462,0.34161,0.243 respectively.When the training step number was 10,it could be achieved to the accuracy.The conclusion was the Resilient Backpropagation,Levenberg-Marquardt,Bayesian Regularization could be used to simultaneous forecast of 4 kinds of synthetic pigment mixed system consist of amaranth,sunset yellow,lemon yellow and carmine respectively.