Study on classification of soy sauce by electronic tongue technique combined with artificial neural network

Electronic tongue as an analytical tool coupled with pattern recognition was attempted to classify 4 different brands and 2 categories (produced by different processes) of Chinese soy sauce. An electronic tongue system was used for data acquisition of the samples. Some effective variables were extracted from electronic tongue data by principal component analysis (PCA). Backpropagation artificial neural network (BP-ANN) was applied to build identification models. PCA score plots show an obvious cluster trend of different brands and different categories of soy sauce in the 2-dimensional space. The optimal BP-ANN model for different brands was achieved when principal components (PCs) were 2, and the identification rate of the discrimination model was 100% in both the calibration set and the prediction set, and the optimal BP-ANN model for different categories had the same result. This work demonstrates that electronic tongue technology combined with a suitable pattern recognition method can be successfully used in the classification of different brands and categories of soy sauce.     

利用粒子群算法优化反向传播人工神经网络模型预测熏肠中4种多环芳烃含量

构建基于粒子群优化(particle swarm optimization,PSO)算法的反向传播人工神经网络(back propagation artificial neural network,BP-ANN)预测模型,对熏肠中4种多环芳烃(polycyclic aromatic hydrocarbons,PAHs)...     

基于人工神经网络模型预测油炸外裹糊鱼块的丙烯酰胺含量

为实现对油炸外裹糊鱼块的丙烯酰胺含量的预测,采用响应面试验设计收集数据,建立以黄原胶和大豆纤维复配比例、外裹糊鱼块干燥时间、大豆油品质、油炸温度、油炸时间为输入值,油炸外裹糊鱼块的丙烯酰胺含量为输出值的反向传播人工神经网络（back propagation artificial neural network,BP-ANN）,预测外裹糊鱼块深度油炸过程丙烯酰胺含量的变化,并用训练集拟合,测试集评估模型的预测能力。结果显示,黄原胶和大豆纤维复配比例、外裹糊鱼块干燥时间、油炸温度、油炸时间对油炸外裹糊鱼块的丙烯酰胺含量均有显著影响,大豆油品质对油炸外裹糊鱼块中丙烯酰胺含量影响不显著。训练后的BP-ANN模型的相关系数R值为0.997,拟合良好,有很强的逼近能力;模型对新数据预测的误差较小,最大相对误差为5.34%,最小相对误差为0.12%,表明BP-ANN模型能准确预测油炸外裹糊鱼块的丙烯酰胺含量。     

Study on lossless discrimination of varieties of yogurt using the Visible/NIR-spectroscopy

A new approach for the rapid and lossless discrimination of varieties of yogurt by Visible/NIR-spectroscopy was put forward. Through the principal component analysis of spectroscopy curves of 5 typical kinds of yogurt, the clustering of yogurt varieties was processed. The results end to be that the cumulate reliabilities of the first two principle components (PC1, PC2) were more than 98.9%, and the first seven principle component (PC1 to PC7) were 99.97%. In addition, an artificial neural network (BP-ANN) model was set up. The first seven principles components of the samples were applied as BP-ANN inputs and the values of the type of yogurt were applied as outputs, which build the three-layer BP-ANN. With this model, the discrimination of yogurt came to be possible. The results of distinguishing the rate of the five yogurt varieties came to be satisfied. It presented that this model was reliable and practicable.     

BP人工神经网络在鱼糜挤压制品生产中的应用

采用反向传播(BP)人工神经网络和响应面法(RSM)模拟操作工艺参数(鱼糜含量、螺杆转速、III区加热温度)对双螺杆挤压机生产的鱼糜挤压制品的品质属性(持水性、膨润度、硬度和弹性)的影响,并比较了BP人工神经网络和RSM所建立的操作工艺参数与产品属性间关系模型的预测误差。试验结果表明,经训练的BP人工神经网络的模拟值和实际值的均方差(MSE)及和方差(SSE)均比RSM低,在模拟产品属性上具有更好的拟合度和准确性,采用此法确定的鱼糜挤压制品最佳工艺参数为:鱼糜含量45.70%,螺杆转速170r/min,III区温度106.2℃。     

Girth prediction of young female body using orthogonal silhouettes

Body girths are the primary dimensions needed for apparel patternmaking. Although 3D body scanning systems can provide direct girth measurements, their high price and complexity prevent widespread use in the apparel industry, especially in small-business operations. Regular 2D body images can be used to measure body widths, depths, and heights, but are not usable for direct girth measurements. In this study, we proposed a new approach to predict girths of young female body with a dedicated 2D body imaging system. In our 2D system, the frontal and side images of a subject were obtained to extract the orthogonal silhouettes, and the body widths and depths were measured at landmarks found from the silhouettes. We found that the measurements had more than 94% of the data in the error range of ±1 cm. Body girths were then predicted through the empirical equations built on the width and depth measurements. The prediction results were compared with those of the manual measurements, and examined in a perspective of apparel-manufacturing requirements. The girth prediction models established in this study were found to be effective, which could facilitate automatic girth measurements essential in apparel pattern alteration using 2D images.     

基于深度信念网络的苹果霉心病病害程度无损检测

针对现有霉心病无损检测只能检测出有无病害,无法对病害程度进行判断的问题,研究并提出一种基于深度信念网络(deep belief net,DBN)的无监督检测模型。该模型由多层限制玻尔兹曼机(restricted Boltzmann machine,RBM)网络和1层反向传播(back propagation,BP)神经网络组成,RBM网络实现最优特征向量映射,输出的特征向量由BP神经网络对霉心病病害程度分类。对225个苹果样本在波长200~1 025 nm获取其透射光谱后,根据腐烂面积占横截面比例将霉心病害程度分为健康、轻度、中度和重度4种,分别用150个和75个样本作为训练集和测试集,以全光谱数据和基于连续投影算法提取的特征波长数据为输入构建病害程度判别模型,并比较DBN模型与偏最小二乘判别分析、BP神经网络和支持向量机模型的识别效果,实验结果表明,DBN模型病害判别准确率达到88.00%,具有较好的识别效果。     

基于近红外光谱技术与BP-ANN算法的豆粕品质快速检测

应用近红外漫反射光谱技术结合误差反向传递人工神经网络(BP-ANN)算法,建立豆粕品质(包括水分、粗蛋白、残油)的定量分析模型。将豆粕漫反射吸收光谱数据进行SNV、DT、SG求导、SG平滑和均值中心化处理,然后采用偏最小二乘方法(PLS)降维获取主成分,并优化选择合适的隐含层节点数、隐含层和输出层转化函数,建立校正模型,并用验证样品对校正模型进行验证。结果显示,BP-ANN法建立的水分、粗蛋白和残油的预测相关系数(R)分别为0.981、0.988、0.982,预测标准偏差(SEP)分别为0.120、0.216、0.036,均优于PLS建模方法结果,且满足传统分析方法的重复性要求,表明BP-ANN方法可用于生产过程豆粕品质的快速监控。     

基于主成分分析和人工神经网络的近红外光谱大豆产地识别

为了准确、快速地识别大豆产地,通过近红外光谱技术（NIRS）结合主成分分析（PCA）和人工神经网络技术（ANN）研究不同国家大豆内含特征,建立进口大豆产地识别模型。采用箱型图校正法,剔除阿根廷、巴西、乌拉圭、美国等4个国家166组大豆样本中12组异常样本。采用多元散射校正（MSC）、标准正态变量（SNV）、Savitzky-Golay（SG）平滑滤波等方法进行光谱数据预处理,结果表明,采用SG（3）平滑结合MSC预处理效果最好。主成分分析表明,前10个主成分的累积贡献率达到99.966%。选取主成分分析得到前10个主成分为输入向量,4个产地作为目标向量,分别采用支持向量机（SVM）、邻近算法（KNN）与人工神经网络法（ANN）建立识别模型。结果表明,采用BP-ANN建模效果最好,总体测试集准确率为95.65%,其中阿根廷准确率为100%,巴西准确率为100%,乌拉圭准确率为80%,美国准确率为100%,该模型能够实现对进口大豆生产国别的识别。     

基于声学特性的鸡蛋蛋壳裂纹检测

通过自行研制的一套禽蛋裂纹检测装置,采集并分析敲击鸡蛋产生的响应信号,检测裂纹鸡蛋。采用基于归一化最小均方算法的自适应滤波器对信号进行去噪处理。结果表明：经自适应滤波后,敲击响应信号的分辨率和灵敏度均有显著提高。提取经滤波去噪后的鸡蛋敲击响应信号功率谱的5 个特征参数,作为误差反传人工神经网络模型的输入向量进行判别。判别模型对实验鸡蛋的交互验证训练集和独立样本预测集的判别率均为97%。     

基于三维荧光光谱快速识别冷榨一级核桃油品牌和产地

为寻找一种能快速识别核桃油品牌和产地的方法,以来自4个品牌3个产地的冷榨一级核桃油为研究对象,分批次收集了300个样品,用荧光检测仪进行样品扫描,分两个时间段采集三维荧光光谱,间隔为1个月。对收集的共600组光谱数据运用主成分分析（PCA）进行特征提取,同品牌或同产地的样品分别选取主成分1组成新的数据集,达到数据降维,再结合偏最小二乘判别（PLS-DA）和人工神经网络判别（BP-ANN）化学模式识别方法,对应构建核桃油的品牌识别模型和产地识别模型。结果表明：PLS-DA和BP-ANN对核桃油的品牌和产地的识别率都能达到100%。因此,三维荧光光谱与PLS-DA和BP-ANN方法结合,可用于快速识别核桃油的品牌和产地。     

基于三维人体扫描模型的衬衣版型生成方法

针对衬衣定制自动生成版型与体型匹配误差不可控问题,提出一种基于三维人体扫描模型的衬衣版型生成方法。首先,扫描目标人体获得个性化的三维人体模型,通过对模型截面曲线的几何特征分析,在三维人体模型上提取个性化体型特征点;使用服装制版标准人台,构建一个衬衣的标准三维基础模型;然后,使用保刚性变形算法,以个性化体型特征点为约束,实现该衬衣标准三维基础模型的变形;最后,使用基于质点-弹簧系统的三维模型展平算法,根据变形后的衬衣三维模型各个版片生成对应的二维版片,添加经实验验证的版片边缘线缩放量,生成适应目标体型的个性化衬衣版型。实验结果表明,该方法能够生成合体的衬衣版型,基于数字化模型和自动化处理算法,实现衬衣版型的自动化生成,保证个性化制版的精度,并提高制版效率,为服装智能生产制造提供技术解决方案。     

食品添加剂高效液相色谱分析的神经网络建模研究

以七种食品添加剂的高效液相色谱分析数据为基础,建立了一个预测保留时间的人工神经网络模型。模型采用BP网络的基本结构和算法,含有一个隐层的双层拓扑结构。确定了隐层节点数的最佳取值范围,不仅可以满足模型对仿真精度的要求,而且可以使模型的训练速度保持在合适的范围内,避免了过多的隐层节点数导致网络冗余和收敛速度下降。模拟结果表明,基本BP算法训练网络具有很好的稳定性,预测结果与实验数据有良好的一致性。     

Application of linear/non-linear classification algorithms in discrimination of pork storage time using Fourier transform near infrared (FT-NIR) spectroscopy

To address the rapid and nondestructive determination of pork storage time associated with its freshness, Fourier transform near infrared (FT-NIR) spectroscopy technique, with the help of classification algorithm, was attempted in this work. To investigate the effects of different linear and non-linear classification algorithms on the discrimination results, linear discriminant analysis (LDA), K-nearest neighbors (KNN), and back propagation artificial neural network (BP-ANN) were used to develop the discrimination models, respectively. The number of principal components (PCs) and other parameters were optimized by cross-validation in developing discrimination models. Experimental results showed that the performance of BP-ANN model was superior to others, and the optimal BP-ANN model was achieved when 5 PCs were included. The discrimination rates of the BP-ANN model were 99.26% and 96.21% in the training and prediction sets, respectively. The overall results sufficiently demonstrate that the FT-NIR spectroscopy technique combined with BP-ANN classification algorithm has the potential to determine pork storage time associated with its freshness.     

应用BP神经网络估算服装铺料长度

针对传统估算方法存在事后性和经验依赖性等缺点,提出应用BP神经网络来估算铺料长度。该方法避开了先打板,再排料,后估算等过程,直接从订单的信息分析入手,通过实验从中提取出有关铺料长度的影响因素,并将其作为输入参数,建立估算铺料长度的BP模型,实验证明该模型具有较好的预测效果。在实际应用中,如将企业实际生产的样本输入模型,通过训练可使估算长度更加接近于企业生产的实际铺料长度,为实现快速准确地估算面料消耗,设计出优化的分床裁剪方案提供指导。     

应用粗糙集和支持向量机的熔喷非织造布过滤性能预测

为预测熔喷非织造布的过滤性能,提出基于属性约简和支持向量机的预测方法。运用粗糙集理论在ROSETTA 环境下对含有9 个参数的熔喷非织造纤网结构参数全集进行约简,得到6 个各含3 个参数的约简集。分别将参数全集及各个约简集作为输入建立基于支持向量机（SVM）和ＢＰ 神经网络（BP-ANN）的28 个过滤性能预测模型,运用交叉验证法进行模型结构参数优化。结果表明：以含厚度、纤维直径和孔径的约简集为输入,基于ＳＶＭ模型预测准确度最高;其对过滤效率和过滤阻力的预测精度均超过98%,且CV 值均小于2%,表明这3 个参数是影响熔喷非织造布过滤性能的核心要素;基于SVM 模型的预测准确度总体优于基于BP-ANN模型的。     

基于低场核磁共振的白酒品牌判别分析

为探求一种白酒品牌判别的方法,基于低场核磁共振（LF-NMR）技术,综合运用主成分分析（PCA）、偏最小二乘法-判别分析（PLS-DA）、反向传播人工神经网络（BP-ANN）方法对6种浓香型白酒品牌共300个样本进行模式识别分析,解析了对不同品牌浓香型白酒进行判别的可行性。结果表明,运用PCA方法对样品进行识别,无法区分白酒品牌;运用PLS-DA的方法对白酒样品进行识别,训练集的识别率约为99.5%,预测集识别率约为96.7%;运用BP-ANN的方法对白酒样品进行识别,训练集识别率约为99.5%,预测集识别率约为98.9%。结果表明,PLS-DA方法和BP-ANN方法对浓香型白酒样品的区分成功,表示将低场核磁共振方法应用到浓香型白酒的品牌判别中是可行有效的。     

光谱预处理对太赫兹光谱预测猪肉K值的影响

采用太赫兹（terahertz,THz）光谱分析技术无损检测猪肉的新鲜度K值,但水会强烈吸收THz波,从而严重影响THz波对肉的检测。考察预处理方法对削弱水的干扰、提升THz光谱检测猪肉K值的模型性能的影响。分别采用多元散射校正、标准正态变量变换、一阶微分、二阶微分4?种预处理方法对衰减全反射光谱进行预处理,基于反向传播人工神经网络回归算法建立猪肉K值的THz光谱预测模型,比较研究4?种预处理方法后的模型预测精度。研究表明：一阶微分预处理方法效果最好,能够消除光谱基线漂移,提高光谱质量。与原始光谱相比,模型的预测集相关系数（Rp）从0.34提高到0.75,预测集均方根误差从20.24%降低到14.36%。因此,选择合适的光谱预处理技术对提高模型预测精度非常重要,采用一阶微分预处理后的THz光谱数据建立反向传播人工神经网络模型能够无损检测猪肉的新鲜度K值。     

Near-Infrared Spectroscopy for Classification of Oranges and Prediction of the Sugar Content

A nondestructive method for the classification of orange samples according to their growing conditions and geographic areas was developed using Vis/Near infrared spectroscopy. The results showed that the NIR spectra of the samples were moderately clustered in the principle component space and pattern recognition wavelet transform (WT) combined artificial neural network (BP-ANN) provided satisfactory classification results. Additionally, a partial least square (PLS) method was constructed to predict the sugar content of certain oranges. It showed excellent predictions of the sugar content of oranges, with standard error of prediction (SEP) values of 0.290 and 0.301 for Shatangju and Huangyanbendizao, respectively.     

Prediction of body lipid change in pregnancy and lactation

A simple method to predict the genetically driven pattern of body lipid change through pregnancy and lactation in dairy cattle is proposed. The rationale and evidence for genetically driven body lipid change have their basis in evolutionary considerations and in the homeorhetic changes in lipid metabolism through the reproductive cycle. The inputs required to predict body lipid change are body lipid mass at calving (kg) and the date of conception (days in milk). Body lipid mass can be derived from body condition score and live weight. A key assumption is that there is a linear rate of change of the rate of body lipid change (dL/dt) between calving and a genetically determined time in lactation (T') at which a particular level of body lipid (L') is sought. A second assumption is that there is a linear rate of change of the rate of body lipid change (dL/dt) between T' and the next calving. The resulting model was evaluated using 2 sets of data. The first was from Holstein cows with 3 different levels of body fatness at calving. The second was from Jersey cows in first, second, and third parity. The model was found to reproduce the observed patterns of change in body lipid reserves through lactation in both data sets. The average error of prediction was low, less than the variation normally associated with the recording of condition score, and was similar for the 2 data sets. When the model was applied using the initially suggested parameter values derived from the literature the average error of prediction was 0.185 units of condition score (+/- 0.086 SD). After minor adjustments to the parameter values, the average error of prediction was 0.118 units of condition score (+/- 0.070 SD). The assumptions on which the model is based were sufficient to predict the changes in body lipid of both Holstein and Jersey cows under different nutritional conditions and parities. Thus, the model presented here shows that it is possible to predict genetically driven curves of body lipid change through lactation in a simple way that requires few parameters and inputs that can be derived in practice. It is expected that prediction of the cow's energy requirements can be substantially improved, particularly in early lactation, by incorporating a genetically driven body energy mobilization.