【作者】 李博；

【导师】 朱莉；

【作者基本信息】 东北林业大学 ， 控制科学与工程， 2025， 硕士

【摘要】 草莓作为典型的呼吸跃变型水果,采摘后生理代谢活跃,易因酶促褐变和微生物侵染等因素导致品质劣变。除了因供应链管理不当造成的损耗外,缺乏对货架期的精准评估也是造成损失的重要因素。因此商家需保证草莓的新鲜度,同时及时调整摆放周期,降低成本损失。草莓现有的品质检测方法大多依赖破坏性理化分析且时效性不足,本研究结合近红外光谱、多指标理化参数融合及智能算法优化技术,构建草莓新鲜度分级与货架期判定无损模型,旨在为生鲜农产品供应链的精准决策提供技术支撑。本研究模拟商品货架贮藏环境,采集不同储存天数下红颜草莓的近红外光谱和理化参数。结合草莓的理化参数以及人工感官评价计算每个样本的新鲜度综合得分,根据得分区间将样本新鲜度降序分为一级、二级和三级。针对光谱中的异常干扰问题,采用马氏距离联合主成分分析剔除异常样本,并使用多种预处理方法对原始光谱进行降噪。通过建立新鲜度和货架期的偏最小二乘判别(PLS-DA)、最小二乘支持向量机(LS-SVM)和极限学习机(ELM)模型选择最佳预处理方法。经验证:LS-SVM和ELM在新鲜度和货架期检测中的准确率优于PLS-DA模型,且最佳预处理方法是SG卷积平滑联合一阶导数法。为降低光谱数据的冗余性,采用竞争性自适应重加权法(CARS)、连续投影法(SPA)、无信息变量消除法(UVE)、遗传算法(GA)、变量组合集群分析(VCPA)、迭代保留信息变量法(IRIV)和套索算法(Lasso)对LS-SVM和ELM模型预测结果进行对比分析。经验证:VCPA-IRIV选择的特征最具代表性;LS-SVM模型在新鲜度分类任务中表现较好;ELM模型在货架期分类任务中表现较好。针对基于网格搜索算法的LS-SVM模型运行时间过长的问题,采用粒子群优化(PSO)、贝叶斯优化(BO)和鲸鱼优化算法(WOA)进行参数寻优,并将模型的平均运行时间以及测试集准确率进行对比。经验证:WOA-LS-SVM模型的检测效果最佳,运行时间最短。为进一步提升新鲜度检测模型的准确率并防止模型过早收敛,提出一种改进型鲸鱼优化算法(IWOA),在WOA中引入Tent混沌映射初始化种群并改进线性衰减因子和螺旋路径更新公式,并进行有效验证。针对ELM模型在货架组别day1和day2标签中分类精确率较差的问题,在Sigmoid激活函数中引入分段区间和线性补偿机制。为进一步提升ELM模型的稳定性和鲁棒性,减少ELM对随机初始权重的过度依赖,采用IWOA生成均匀的隐含层初始权重和偏置参数,引入Bagging集成学习策略增强ELM模型的稳定性。经验证:基于IWOA的集成ELM模型可有效对草莓货架期进行分类检测。综上,本研究可以为草莓的新鲜度和货架期检测提供理论依据。更多还原

【Abstract】 Strawberries,as a typical climacteric fruit,exhibit active post-harvest physiological metabolism and are prone to quality deterioration due to enzymatic browning and microbial contamination.In addition to losses caused by improper supply chain management,the lack of accurate shelf-life assessment is also a major contributing factor.Therefore,businesses must ensure the freshness of strawberries while adjusting display cycles in a timely manner to reduce cost losses.To address the limitations of existing strawberry quality assessment methods,which rely on destructive physicochemical analysis and lack timeliness,this study integrates near-infrared spectroscopy,multi-parameter physicochemical data fusion,and intelligent algorithm optimization to develop a non-destructive model for strawberry freshness grading and shelf-life prediction.The goal is to provide technical support for precise decision-making in the fresh agricultural supply chain.This study simulated the storage environment of commercial shelves and collected near-infrared spectra and physicochemical parameters of Hongyan strawberries at different storage durations.By integrating physicochemical parameters and sensory evaluations,a comprehensive freshness score was calculated for each sample.Based on these scores,samples were ranked in descending order and classified into three freshness levels:Grade 1,Grade 2,and Grade 3.To address abnormal interference in the near-infrared spectra,Mahalanobis distance combined with principal component analysis(PCA)was used to eliminate outliers,and multiple preprocessing methods were applied to denoise the raw spectra.Partial least squares discriminant analysis(PLS-DA),least squares support vector machine(LS-SVM),and extreme learning machine(ELM)models were developed to determine the optimal preprocessing method for freshness and shelf-life classification.Validation results showed that LS-SVM and ELM achieved higher accuracy in freshness and shelf-life classification than the PLS-DA model.The optimal preprocessing method was found to be Savitzky-Golay(SG)convolution smoothing combined with the first derivative method.To reduce the redundancy of spectral data,this study compared the effects of various feature selection methods on the prediction performance of LS-SVM and ELM models.The methods evaluated included Competitive Adaptive Reweighted Sampling(CARS),Successive Projections Algorithm(SPA),Uninformative Variable Elimination(UVE),Genetic Algorithm(GA),Variable Combination Population Analysis(VCPA),Iterative Retained Information Variables(IRIV),and Least Absolute Shrinkage and Selection Operator(Lasso).Validation results showed that the features selected by the VCPA-IRIV method were the most representative.Additionally,the LS-SVM model performed better in the freshness classification task,while the ELM model showed superior performance in shelf-life classification.To address the issue of excessive runtime in the LS-SVM model based on the grid search algorithm,parameter optimization was performed using Particle Swarm Optimization(PSO),Bayesian Optimization(BO),and the Whale Optimization Algorithm(WOA).The models were compared in terms of average runtime and accuracy on the test set.Validation results showed that the WOA-LS-SVM model achieved the best detection performance with the shortest runtime.To further enhance the accuracy of the freshness detection model and prevent premature convergence,an Improved Whale Optimization Algorithm(IWOA)is proposed.The improvements include introducing Tent chaotic mapping for population initialization,refining the linear decay factor,and modifying the spiral updating mechanism.The effectiveness of these enhancements was successfully validated.To address the poor precision of the ELM model in distinguishing between shelf-life groups day1 and day2,a segmented interval and linear compensation mechanism were introduced into the Sigmoid activation function.To further enhance the stability and robustness of the ELM model and reduce its excessive dependence on randomly initialized weights,the Improved Whale Optimization Algorithm(IWOA)was employed to generate uniform initial hidden layer weights and bias parameters.Additionally,a Bagging ensemble learning strategy was incorporated to improve the model’s stability.Validation results confirmed that the IWOA-based ensemble ELM model effectively classified the shelf life of strawberries.In conclusion,this study provides a theoretical basis for the freshness and shelf-life detection of strawberries.更多还原