基于粒子群优化算法优化反向传播神经网络构建冷藏草鱼新鲜度的近红外光谱预测模型

Establish of prediction models of near-infrared spectroscopy for freshness of refrigerated Ctenopharyngodon idella based on particle swarm optimization algorithm for optimizing back propagation neural network

中文关键词:  近红外光谱  冷藏草鱼新鲜度  总挥发性盐基氮  粒子群优化算法优化反向传播神经网络  正交信号校正法

英文关键词:near-infrared spectroscopy  refrigerated  Ctenopharyngodon idella  freshness  total volatile basic nitrogen  particle swarm optimization  back propagation neural network  orthogonal signal correction

中文摘要:

      目的 基于机器学习算法构建冷藏草鱼新鲜度的近红外光谱预测模型。方法 采集连续冷藏6 d的草鱼片的新鲜度指标, 并进行方差分析。选择受冷藏天数影响最大的指标—总挥发性盐基氮(total volatile basic nitrogen, TVB-N)进行定量预测。运用x-y距离结合的样本划分(sample set partitioning based on joint x-y distance, SPXY)方法进行数据集的划分, 并采用正交信号校正法(orthogonal signal correction, OSC)、Savitzky-Golay (SG)、一阶导数及其组合算法进行光谱预处理。再运用竞争性自适应重加权采样(competitive adaptive reweighted sampling, CARS)、连续投影算法(successive projections algorithm, SPA)、主成分分析(principal component analysis, PCA)对光谱变量进行选择和降维。最后结合偏最小二乘回归(partial least squares regression, PLSR)、反向传播(back propagation, BP)神经网络和粒子群优化算法(particle swarm optimization, PSO)优化BP神经网络(PSO-BP), 建立草鱼(Ctenopharyngodon idella)片新鲜度定量预测模型。结果 各线性和非线性模型均得到了良好的预测效果, 预测集相关系数均超过了0.95。PLSR表现较为稳定, BP神经网络虽提高了校正集预测性能, 但是预测集性能不如PLSR。PSO-BP既保证了校正集预测性能, 也提高了预测集性能。基于OSC+D1预处理和CARS变量选择后的PSO-BP模型性能最优(R2P=0.987, 预测集的均方根误差为0.108, 相对分析误差为7.778)。结论 基于PSO-BP算法和近红外光谱的定量预测模型可以很好地预测冷藏鱼肉的新鲜度指标。

英文摘要:

      Objective To establish a prediction model of near-infrared spectroscopy for freshness of refrigerated Ctenopharyngodon idella based on machine learning algorithms. Method Freshness indicators of Ctenopharyngodon idella fillets stored continuously for 6 days were collected, and variance analysis was performed. The indicator most affected by the storage days, total volatile basic nitrogen (TVB-N), was quantitatively predicted. The sample set partitioning based on joint x-y distance (SPXY) algorithm was used for dataset partitioning, and spectral preprocessing was conducted using orthogonal signal correction (OSC), Savitzky-Golay (SG), first-order derivative, and their combinations. Subsequently, competitive adaptive reweighted sampling (CARS), successive projection algorithm (SPA), and principal component analysis (PCA) were employed for spectral variable selection and dimensionality reduction. Finally, a quantitative prediction model for the freshness of Ctenopharyngodon idella fillets was constructed by incorporating partial least squares regression (PLSR), back propagation (BP) neural network, and particle swarm optimization algorithm (PSO) for optimizing back propagation neural network (PSO-BP). Results Both linear and nonlinear models showed excellent prediction performance, with correlation coefficients in the prediction set exceeding 0.95. PLSR demonstrated relatively stable performance, while the BP neural network improved calibration set prediction performance, although not as effectively as PLSR in the prediction set. PSO-BP ensured both calibration set prediction performance and improved prediction set performance. The PSO-BP model, based on OSC+D1 preprocessing and CARS variable selection, exhibited the best performance (R2P=0.987, root mean square error of prediction was 0.108, relative percent deviation was 7.778). Conclusion The quantitative prediction model based on the PSO-BP algorithm and near-infrared spectroscopy combination can effectively predict the freshness indicator of refrigerated fish meat.