【作者】 于洋；

【导师】 田海清；

【作者基本信息】 内蒙古农业大学 ， 农业机械化工程， 2025， 博士

【摘要】 青贮玉米饲料在现代畜牧业中具有重要地位,通过厌氧发酵可在较长时间内保持较高营养价值和良好适口性。然而,在实际生产与饲喂过程中,饲料暴露于氧气环境极易引发二次发酵,导致品质下降并影响适口性。为保障饲料质量,对包括p H值、有机酸(乳酸、乙酸、丙酸)含量以及含水率在内的关键发酵指标进行准确、实时监测至关重要。传统检测手段多依赖人工感官评估与理化分析,存在耗时费力,操作繁琐等问题,难以满足饲料在二次发酵过程中动态变化的实时监测需求。为此,本研究以具有普遍代表性的不同品质青贮玉米饲料为研究对象,系统分析二次发酵过程中p H值、有机酸及含水率等关键发酵指标的变化规律,结合光谱与图像层面的响应特征,深入探讨各发酵指标的动态演变趋势及其光学检测表征机理,并据此明确适配的检测技术。p H值与含水率采用近红外高光谱成像(NIR-HSI)技术实现无损检测。有机酸含量则融合近红外高光谱与计算机视觉信息实现检测。进一步,研究挖掘了表征饲料品质的光谱特征与颜色纹理图像特征,提出基于光学信息的关键发酵指标快速无损检测方法,为青贮饲料品质智能监测提供技术支撑。本研究的具体内容及结果如下:(1)利用高光谱成像(HSI)技术,构建了青贮玉米饲料在二次发酵过程中p H值检测模型。首先,分别采用SG卷积平滑、MSC等多种方法对原始光谱数据进行预处理,最终确定MSC为最优预处理方法。随后,采用CARS、SPA、DWT、VCPA-IRIV、UVE和BOSS六种特征选择算法提取关键光谱波段。在此基础上,分别构建基于PLS、SVR和ELM的预测模型,并系统评估不同特征选择方法对模型性能的影响。结果表明,BOSS算法筛选出的20个特征波段在有效降低数据冗余的同时保留了关键信息。基于BOSS特征波段构建的ELM模型(BOSS-ELM)表现最优,其预测集决定系数_PR~2为0.9241,均方根误差(RMSEP)为0.4372,相对分析误差(RPD)为3.6565,显示出良好的预测精度与泛化能力。进一步引入遗传算法(GA)、鲸鱼优化算法(WOA)和秃鹰搜索优化算法(BES)对模型优化,最终构建的BOSS-BES-ELM模型在预测集上的_PR~2提升至0.9598,RMSEP降至0.3216,RPD增至5.1448,显著增强了模型的稳定性与预测性能。基于该优化模型,进一步实现了p H值在青贮饲料中的空间分布可视化。(2)融合近红外高光谱与计算机视觉技术,构建了针对青贮玉米饲料二次发酵过程中乳酸、乙酸和丙酸含量的无损检测模型。首先,通过多种光谱预处理确定二阶导数为最优方法。随后,采用BOSS、SPA和VCPA-IRIV算法筛选光谱特征波段,并利用Lasso回归算法剔除图像特征冗余。基于筛选结果分别构建光谱与图像单模态预测模型,并通过数据级和特征级融合实现多模态联合建模。为进一步提升预测性能与鲁棒性,本研究在决策级融合中提出动态自适应加权堆叠集成模型(DAWSIM),该模型由堆叠集成模块与动态加权模块构成,前者以PLS、ELM、SVR和MLP为基础学习器,XGBoost为元学习器,实现全局信息的集成建模。后者通过随机森林(RF)对基础模型的局部预测误差进行建模,实现动态权重调整。针对传统指数函数分配权重易极端化的问题,DAWSIM模型引入Log Sum Exp函数改进权重分配,有效缓解权重失衡并增强抗噪能力。结果表明,DAWSIM优于其他方法,乳酸、乙酸和丙酸的预测集决定系数_PR~2分别为0.9728、0.8120和0.6896,RMSEP分别为1.6002、2.5268、0.9302,RPD分别为6.0664、2.3064和1.7949,显著提升了模型的预测精度与泛化能力。(3)提出一种基于深度学习的高光谱成像中多模态融合方法。首先,采用提出的BS-Net-Channel网络对原始高光谱数据进行通道级波段筛选,保留与含水率高度相关的15个关键波段,并将其与对应的空间图像组合构建为三维输入张量。随后,以轻量化三维卷积神经网络(3D-CNN)作为主干网络,引入BIFPN模块以实现多尺度特征的双向融合,并提出SE-HFAF注意力机制,用于自适应调节光谱与图像特征间的权重分配。最终构建的3D-LBIFANet模型实现了高精度预测,其预测集决定系数_PR~2达到0.9489,RMSEP为0.0461,RPD为4.4226,显著优于传统浅层建模方法。消融试验进一步验证了BIFPN融合结构与注意力机制在模型性能提升中的关键作用。(4)构建了一套青贮玉米饲料品质检测软件系统,可实现对p H值、有机酸及含水率等关键发酵指标实现快速评估。系统采用MVC架构,逻辑层集成前期训练的最优预测模型,界面层基于Py Qt5开发,支持动态图形用户界面(GUI)构建、数据可视化及交互式参数操作,显著提升了系统的可用性与用户体验。系统功能涵盖用户管理、指标检测分析与结果展示等核心模块,结合标准化处理流程与模块化设计思路,有效降低了操作门槛,增强了系统的适应性。软件测试结果表明,该系统具备良好的交互体验与功能完整性,运行稳定高效,能够为青贮饲料品质的智能化评价提供有力支撑。更多还原

【Abstract】 Silage corn plays a vital role in modern animal husbandry,as anaerobic fermentation enables it to maintain high nutritional value and good palatability over extended periods.However,during actual production and feeding,exposure to oxygen can easily trigger secondary fermentation,leading to quality deterioration and reduced palatability.To ensure feed quality,it is essential to accurately and continuously monitor key fermentation indicators—such as p H value,organic acid contents(lactic acid,acetic acid,and propionic acid),and moisture content—in real time.Traditional detection methods largely rely on manual sensory evaluation and physicochemical analysis,which are time-consuming,labor-intensive,and cumbersome,making them inadequate for the real-time monitoring of dynamic changes during secondary fermentation.To address this issue,this study focuses on silage corn of varying quality levels that are broadly representative,and systematically analyzes the variation patterns of key fermentation indicators—including p H,organic acids,and moisture content—throughout the secondary fermentation process.By integrating spectral and image-level response features,the dynamic evolution trends and optical detection mechanisms of each fermentation indicator are thoroughly investigated,thereby identifying suitable detection technologies.Specifically,near-infrared hyperspectral imaging(NIR-HSI)is employed to achieve non-destructive detection of p H and moisture content,while the contents of organic acids are assessed through the fusion of NIR hyperspectral and machine vision information via joint modeling and multimodal feature analysis.Furthermore,the study extracts spectral features representing internal chemical compositions and image features such as color and texture,and proposes an optical information-based method for the rapid,non-destructive detection of key fermentation indicators.This provides a technical foundation for the intelligent monitoring of silage corn quality.The main contents and findings of this study are as follows:(1)A p H detection model for silage corn during secondary fermentation was developed using hyperspectral imaging(HSI).The raw spectral data were first preprocessed using several methods including SG smoothing and MSC,with MSC ultimately identified as the optimal technique.Six feature selection algorithms—CARS,SPA,DWT,VCPA-IRIV,UVE,and BOSS—were then applied to extract key spectral wavelengths.Based on the selected features,predictive models were constructed using PLS,SVR,and ELM,and the impact of different feature selection methods on model performance was systematically evaluated.Results indicated that the 20 wavelengths selected by BOSS effectively reduced data redundancy while retaining essential information.Among all models,the BOSS-ELM model achieved the best performance,with an_PR~2 of 0.9241,RMSEP of 0.4372,and RPD of 3.6565 on the prediction set,demonstrating strong predictive accuracy and generalization capability.To further enhance the model,optimization algorithms including GA,WOA,and BES were introduced,and the final BOSS-BES-ELM model achieved a prediction_PR~2 of 0.9598,RMSEP of 0.3216,and RPD of 5.1448,significantly improving stability and predictive performance.Based on this optimized model,the spatial distribution of p H in silage corn was successfully visualized.(2)A non-destructive detection model for lactic acid,acetic acid,and propionic acid contents in silage corn during secondary fermentation was developed by integrating near-infrared spectroscopy(NIR)and machine vision technologies.First,among various spectral preprocessing methods,the second derivative was identified as the optimal approach.Then,BOSS,SPA,and VCPA-IRIV algorithms were employed to select key spectral features,while image feature redundancy was reduced using the Lasso regression algorithm.Based on the selected features,unimodal prediction models were constructed separately for spectral and image data,followed by multimodal joint modeling through data-level and feature-level fusion.To further enhance prediction performance and robustness,a dynamic adaptive weighted stacking integration model(DAWSIM)was proposed at the decision level.DAWSIM comprises a stacking module and a dynamic weighting module:the former uses PLS,ELM,SVR,and MLP as base learners and XGBoost as the meta-learner to integrate global information,while the latter employs RF to model the local prediction errors of base learners for dynamic weight adjustment.To address the extremization issue caused by traditional exponential weighting functions,the Log Sum Exp function was introduced to improve the weight allocation mechanism,effectively mitigating imbalance and enhancing noise resistance.Results showed that DAWSIM outperformed other approaches,achieving_PR~2 values of 0.9728,0.8120,and0.6896;RMSEP values of 1.6002,2.5268,and 0.9302;and RPD values of 6.0664,2.3064,and 1.7949 for lactic acid,acetic acid,and propionic acid,respectively,demonstrating significant improvements in both prediction accuracy and generalization ability.(3)A deep learning-based multimodal fusion method for hyperspectral imaging was proposed.First,an improved BS-Net-Channel network was employed to perform channel-level band selection on the original hyperspectral data,retaining 15 key bands highly correlated with moisture content.These selected bands were then combined with corresponding spatial images to form a three-dimensional input tensor.Next,a lightweight3D-CNN was used as the backbone network,into which a BIFPN module was integrated to achieve bidirectional multi-scale feature fusion.Additionally,a SE-HFAF attention mechanism was proposed to adaptively adjust the weight allocation between spectral and image features.The final constructed model,3D-LBIFANet,achieved high-accuracy prediction,with an_PR~2 of 0.9489,RMSEP of 0.0461,and RPD of 4.4226 on the prediction set,significantly outperforming traditional shallow modeling methods.Ablation experiments further confirmed the critical roles of the BIFPN fusion structure and the attention mechanism in enhancing model performance.(4)A software system for silage corn quality detection was developed,enabling rapid assessment of key fermentation indicators such as p H,organic acids,and moisture content.The system adopts an MVC architecture,with the logic layer integrating the optimal predictive models trained in earlier stages.The interface layer is developed using Py Qt5,supporting dynamic graphical user interface(GUI)construction,data visualization,and interactive parameter operations,significantly enhancing usability and user experience.Core functionalities of the system include user management,indicator detection and analysis,and result display.By combining standardized processing workflows with a modular design approach,the system effectively lowers the operational threshold and improves adaptability.Software testing results demonstrate that the system offers a high-quality interactive experience and functional completeness,operating with stability and efficiency,thereby providing strong support for the intelligent evaluation of silage corn quality.更多还原