摘 要

在铝电解工业的生产中，时刻检测铝电解槽工作状态,并给出合理的判别，是提高铝电解工业生产品质及自动化水平的重要环节。采用计算机技术和信息化技术对铝电解生产过程进行监控，分析采集数据中包含的丰富信息，发现生产过程中与实时槽况相互作用的的信息，提取这些信息去分析铝电解槽的实时槽况，分析铝电解槽现所处的状态，预报该状态下可能会出现的故障，及时维修，保证正常生产。因而，如何运用计算机技术及信息化技术监测铝电解槽的工作状态，并作出诊断有着重大的现实意义。

本文首先根据铝电解生产原理及其故障模式诊断原理，选择铝电解槽的槽电阻信号作为监控对象，把槽电阻按频率划分为10个区间，然后以其频谱幅值大小为特征，组成10维的特征向量，3维的输出向量。按照上述输入和输出的特征向量大小，BP神经网络的输入层节点数目对应槽电阻的10维特征向量，故其大小为10个，输出层节点数目对应输出状态的数目，故其大小为3个，而后根据公式计算神经网络隐含层的节点数目。传统的BP神经网络有着易陷入局部最小值并且全局搜索速度低的问题，本文试着将遗传算法及BP神经网络结合，利用遗传算法搜索神经网络的输入层与隐含层、隐含层与输出层各节点的连接初始权值及阈值进行寻优，以便加速神经网络训练时的学习速度，增强其稳健性及泛化性能。对于本文的铝电解槽况预测系统，划分为6个功能模块，分别是系统管理模块，数据显示模块，曲线动态显示模块，故障预测模块，数据查询，打印模块及帮助模块，并利用Visual C 、SQL Server2008和MATLAB联合编程进行实现。最后，对比BP神经网络及遗传算法优化的BP神经网络两者的训练速度及训练时间。结果显示，遗传算法优化后的BP神经网络的训练次数减少14%，收敛时间缩短19s。然后利用10个测试样本，对遗传算法优化后的BP神经网络模型开始测试，正确率达到90%，基本满足工程应用的需求。拟定本系统的测试环境及方案，对铝电解槽况预测系统进行测试，并进行软件的操作，证明软件能对铝电解槽况进行预测。

摘要：电解槽；BP神经网络；预测系统；Visual C ；MATLAB

Abstract

Moment in the production of aluminum electrolytic industry, detection of aluminum electrolytic cell work status, and gives a reasonable criterion, is to improve the aluminum electrolytic industry an important part of the production quality and the automation level. Using computer technology and information technology for aluminum electrolysis production process monitoring and analysis of collected data contains abundant information, found that interact with real time slot conditions during the production of information, to extract the information to analyse the real time slot of the aluminum electrolytic cell conditions, analysis of aluminium electrolytic cell is now of the state, forecast the condition may be failure, timely maintenance, guarantee the normal production. Therefore, how to use the computer technology and informationization technology to monitor the work state of aluminum electrolytic cell and make the diagnosis is of great practical significance.

At first, this paper according to the principle of electrolytic aluminium production principle and fault diagnosis model, select aluminum electrolyzer tank resistance signal as monitoring objects, the tank resistance according to the frequency is divided into 10 interval, then is characterized by its frequency spectrum amplitude size, composed of 10 dimensional feature vector, the output of the 3 d vector. Size according to the characteristics of the input and output vector, the BP neural network input layer node number of the corresponding slot resistance of 10 dimensional feature vector, so its size is 10, the output layer node number corresponding to the number of the output state, so the size of three, and then according to the formula to calculate the node number of the hidden layer neural network. The traditional BP neural network has easily plunged into local minimum and global search the problem of low speed, this paper try to combine genetic algorithm and BP neural network, neural network using genetic algorithm to search the input layer and hidden layer, hidden layer and output layer connection initial weights and threshold value of each node to carry on the optimization, in order to accelerate the learning speed when the neural network training, to enhance the robustness and generalization performance. For aluminum electrolytic cell condition prediction system, this paper is divided into six function modules, system management module, data display module, dynamic curve display module, fault prediction module, data query, print module and help module, and by using Visual c , SQL Server2008 and MATLAB programming for implementation. Finally, the training speed and training time of BP neural network, which is optimized by BP neural network and genetic algorithm, are compared. The results showed that the optimized BP neural network was reduced by 14% and 19s in convergence time. Then, using 10 test samples, the optimized BP neural network model of BP neural network was tested, and the accuracy reached 90%, which met the requirement of engineering application. Draw up the test environment and the scheme of this system, the testing of aluminum reduction cells, the forecast system, and carry on the operation of the software, proved that the software can forecast of aluminum reduction cells.

Key words: electrolytic cell; The BP neural network; Prediction system; Visual c ; MATLAB

目 录

摘 要 I

Abstract II

1 绪论 1

1.1 研究背景及意义 1

1.2 国内外研究现状 2

1.2.1 国外研究状况 2

1.2.2 国内研究状况 2

1.3 故障预测技术的分类 3

1.3.1 基于解析模型的故障预测方法 3

1.3.2 基于信号处理的故障预测方法 3

1.3.3 基于知识的故障预测方法 3

1.4 铝电解故障预测方法 3

1.5 研究内容 4

1.6 论文的结构安排 5

2 铝电解槽故障预测系统总体设计 6

2.1 铝电解槽生产的原理 6

2.2 槽电阻信号的计算 6

2.3 铝电解槽的故障分类 6

2.3.1 阳极效应 7

2.3.2 铝液波动 7

2.3.3 冷槽热槽 7

2.4 系统的总体设计方案 7

2.5 主要开发工具 9

2.6 铝电解槽况预测系统模块化设计 9

2.7 本章小结 10

3 故障预测子系统的设计 11

3.1 故障预测方法的选择 11

3.2 神经网络结构的选择 11

3.3 基于神经网络故障模型的建立 13

3.3.1 输入和输出向量的选择 13

3.3.2 输入层及输出层的节点数 14

3.3.3 隐含层的节点数 14

3.3.4 学习参数的选择 14

3.3.5 模型的建立 15

3.3.6 BP神经网络的优化 16

3.4 遗传算法优化BP神经网络 16

3.4.1 BP神经网络的优化 16

3.4.2 编码方案的选择 19

3.4.3 初始种群的选择 19

3.4.4 适应度函数 19

3.4.5 遗传操作 19

3.4.6 遗传算法的控制参数 20

3.5 基于遗传神经网络故障模型的建立 21

3.6 软件实现的方案 24

3.7 本章小结 24

4 软件应用程序设计 25

4.1 数据传输子系统的设计 25

4.1.1 数据库访问技术 25

4.1.2 数据库的建立 25

4.1.3 软件实现方案 25

4.2 其他功能模块设计 26

4.2.1 数据实时显示 26

4.2.2 曲线动态显示 26

4.2.3 数据查询及打印 27

4.2.4 故障预测模块 27

4.2.5 帮助系统 27

4.2.6 软件系统应用程序发布 27

4.3 本章小结 27

5系统的测试结果及应用 29

5.1 故障预测的测试 29