铝电解生产环境极具挑战性，涉及复杂的工厂布局和设备结构，同时受到热场、流场和电磁场等多种交叉耦合因素的影响，常常出现工况不稳定和频繁故障等状况，进而引发设备损坏和人员安全事故。精确预测铝电解槽的关键参数变化并对槽况进行准确评价，能够帮助电解车间优化生产计划和控制策略。本文基于铝电解过程产生的数据和大数据挖掘技术，对铝电解槽的关键参数预测及槽况评价方法进行研究，完成了以下几项工作：（1）铝电解槽过程监控系统的架构研究。本文采用了IBM的Harmony-SE建模方法，首先确定了铝电解过程监控中的各利益相关者及其运行场景，并将这些不同利益相关者的需求转化为系统的具体需求。接着对系统功能进行了深入分析，实现了从系统需求到系统功能的有效映射。最后进行了设计综合工作，确保系统的功能与各子模块之间的对应关系得到合理配置。（2）针对铝电解槽关键参数预测问题，本文提出了一种基于多尺度特征融合的深度学习模型。以Transformer的解码器为基本框架，使用自注意力、卷积、扩张卷积、傅里叶变换4种尺度对铝电解槽生产过程数据进行特征提取并加权融合。通过交叉注意力机制提取生产过程参数和控制参数之间的关系，通过堆叠多个子模块，最后由一个MLP网络输出槽温、铝水平、电解质水平和分子比四个关键参数的预测值。实验证明本文的模型相比主流的时序预测模型能够更准确预测关键参数的变化。（3）针对铝电解槽槽况评价的问题，本文提出了一种基于集成学习和树模型的方法。根据铝电解槽数据量较小、异常值较多、耦合关系复杂等特点，本文选择使用基于树模型的分类算法训练铝电解槽程槽况评价模型，并使用贝叶斯优化方法对其超参数进行寻优。单一树模型存在学习能力有限、容易过拟合的问题，本文用集成学习策略将决策树、随机森林、GBDT、XGBoost和LightGBM五种不同的树模型融合。实验证明本文的方法能够准确地评价铝电解槽的槽况。（4）为了解决基于机器学习的槽况评价模型可解释性较差的问题，本文引入了铝电解槽健康度的概念，根据决策树输出的铝电解槽参数重要度和机理知识筛选了4个重要参数，用多项式回归的方法提出了铝电解槽健康度的经验公式，丰富了铝电解槽评价体系。

Aluminium electrolysis is a challenging production environment, involving complex plant layouts and equipment structures, as well as multiple cross-coupling factors such as thermal, flow and electromagnetic fields, often resulting in unstable conditions and frequent failures, which can lead to equipment damage and personnel safety accidents. Accurate prediction of changes in key parameters of aluminium electrolysis cells and accurate evaluation of cell conditions can help electrolysis shops to optimize production plans and control strategies. In this paper, based on the data generated by the aluminium electrolysis process and big data mining technology, the key parameters of aluminium electrolyzer prediction and tank condition evaluation methods are studied, and the following tasks are completed: 1.An architectural study of an aluminium electrolyzer process monitoring system. In this paper, IBM‘s Harmony-SE modelling approach was adopted to first identify the various stakeholders in aluminium electrolysis process monitoring and their operation scenarios, and to translate the requirements of these different stakeholders into system-specific requirements. Then an in-depth analysis of the system functionality was carried out to achieve an effective mapping from system requirements to system functionality. Finally, the design synthesis work was carried out to ensure that the system functions and the correspondence between the sub-modules were properly configured. 2.Aiming at the problem of predicting key parameters of aluminium electrolyzer, this paper proposes a deep learning model based on multi-scale feature fusion. Taking Transformer‘s decoder as the basic framework, features are extracted and weighted and fused to the aluminium electrolyzer production process data using four scales: self-attention, convolution, dilation convolution and Fourier transform. The relationship between production process parameters and control parameters is extracted through the cross-attention mechanism, and the predicted values of four key parameters, namely, tank temperature, aluminium level, electrolyte level and mole ratio, are finally outputted by an MLP network by stacking multiple sub-modules. Experiments demonstrate that the model in this paper is able to predict the changes of key parameters more accurately than the mainstream time series prediction model. 3.Aiming at the problem of cell state evaluation of aluminium electrolysis tank, this paper proposes a method based on integrated learning and tree model. According to the characteristics of aluminium electrolytic tank with small data volume, more outliers and complex coupling relationship, this paper chooses to use the classification algorithm based on tree model to train the tank condition evaluation model of aluminium electrolytic tank program, and uses Bayesian optimization method to find the optimality of its hyper-parameters. A single tree model has the problem of limited learning ability and easy overfitting, so this paper uses an integrated learning strategy to integrate five different tree models: decision tree, random forest, GBDT, XGBoost and LightGBM. Experiments prove that the method in this paper can accurately evaluate the tank conditions of aluminium electrolysis tanks. 4.In order to solve the problem of poor interpretability of the tank condition evaluation model based on machine learning, this paper introduces the concept of aluminium electrolytic tank health, screens four important parameters according to the importance of aluminium electrolytic tank parameters output from the decision tree and the knowledge of the mechanism, and proposes an empirical formula for the health of aluminium electrolytic tanks by polynomial regression, which enriches the evaluation system of aluminium electrolytic tanks.