Artificial neural networks in the prediction of fraud in integral milk powder by adding whey powder

ABSTRACT: This research was performed to ascertain the most suitable Artificial Neural Network (ANN) model to quantify the degree of fraud in powdered milk through the addition of powdered whey via regular standard physicochemical analyses. In this study, an evaluation was done on 103 samples with different quantities of added whey powder to whole milk powder. Using Fourier Transform Infrared Spectroscopy the fat, cryoscopy, total solids, defatted dry extract, lactose, protein and casein were analyzed. The hyperbolic tangent transformation function was used with 45 topologies, and the Holdback and K-fold validation methods were tested. In the Holdback method, 75% of the database was employed for training, while 25% was used for validation. In the K-fold method, the database was categorized into five equal sized subsets, which alternated between training and validation. Of the two methods, the K-fold method was proven to have superior efficiency. Next, analysis was done on three models of multilayer perceptron networks with feedforward architecture. In Model 1, the input layer contained all the physicochemical analyses conducted, in model 2 the casein analysis was excluded, and in model 3 the routine analyses performed for dairy products was done (fat, defatted dry extract, cryoscopy and total solids). From Model 3 an ANN was derived which could satisfactorily predict fraud calculated from using the routine and standard analyses for dairy products, containing 64 nodes in the hidden layer, with R2 of 0.9935 and RMSE of 0.5779 for training, and R2 of 0.9964 and RMSE of 0.4358 for validation.摘要：本研究旨在确定最适配的人工神经网络（Artificial Neural Network, ANN）模型，以通过常规标准理化分析量化乳清粉添加型奶粉的掺假程度。本研究针对103份添加不同比例乳清粉的全脂奶粉样品开展评估，采用傅里叶变换红外光谱（Fourier Transform Infrared Spectroscopy, FTIR）对脂肪、冰点、总固形物、脱脂干提取物、乳糖、蛋白质及酪蛋白进行检测分析。本研究采用双曲正切变换函数构建45种网络拓扑结构，并测试了预留法（Holdback）与K折交叉验证（K-fold）两种验证策略：预留法中将75%的数据集用于模型训练，剩余25%用于验证；K折交叉验证法则将数据集划分为5个大小均等的子集，交替用于模型训练与验证。经对比，K折交叉验证法的验证效率更优。随后，本研究针对3种前馈架构的多层感知器网络模型开展分析：模型1的输入层包含全部开展的理化检测指标；模型2剔除了酪蛋白检测项；模型3采用乳品行业常规检测指标（脂肪、脱脂干提取物、冰点及总固形物）。基于模型3得到的人工神经网络可基于乳品常规与标准理化检测结果，对奶粉掺假程度进行可靠预测：该模型隐藏层包含64个节点，训练集的决定系数（Coefficient of Determination, R²）为0.9935、均方根误差（Root Mean Square Error, RMSE）为0.5779，验证集的决定系数（Coefficient of Determination, R²）为0.9964、均方根误差（Root Mean Square Error, RMSE）为0.4358。