前言：

(资料图片仅供参考)

本文基于预训练的卷积神经网络（Convolutional Neural Network，CNN）进行模型创建,进行调整后应用到已有的数据集电熔镁砂（Fused magnesium furnace, FMF）工况识别（实验室数据集，不公开）。

正文：

这里不对数据集背景展开介绍，为了方便说明，视FMF数据集为一个故障识别问题，故障工况的标签为1，非故障工况的标签为0。FMF数据集样本来自于实际电熔镁炉同一批炉次工况运行的片段，后期经过图像增强处理，图像大小为32X32X3的RGB图像。其中使用训练集traindata2.mat包含3010张图像，故障数与非故障数比例为1923：1087；使用测试集testdata4.mat包含626张图像，故障数与非故障数比例为416：210。

实验环境使用MATLAB 2022b的深度学习工具箱以及附加功能资源。

建议先安装好附加资源管理器的AlexNet

修改全连接层

非常失败的一次尝试，为了适应Alexnet网络输入，调整图片输入为227*227*3，训练结果只有91%的准确率，在测试集上效果不理想

%% Alexnet迁移学习

%% 清空环境和图窗

clear

close all

%% 导入模型

net = alexnet;

layers = net.Layers

layers(end-2) = fullyConnectedLayer(2);  %故障识别为二分类

layers(end) = classificationLayer;

%% 准备数据集

load("traindata2.mat")

len = length(label);

img(:,:,:,len) = zeros(227,227,3); %预分配内存

for i  = 1:len

img(:,:,:,i) = imresize(video(:,:,:,i),[227 227]); %调整输入大小为227*227*3

end

rng(10)   % 设置seed，便于复现

randIndex = randperm(len);

len1 = round(0.7*len);

trainIndex = randIndex(1:len1);

testIndex = randIndex(len1+1:end);

trainlabels = label(trainIndex);

trainImages = img(:,:,:,trainIndex);

vallabels = label(testIndex);

valImages = img(:,:,:,testIndex);

trainlabels = categorical(trainlabels);

vallabels = categorical(vallabels);

%% 模型训练

options = trainingOptions('adam',...

'ExecutionEnvironment','gpu',...

'InitialLearnRate',0.01,...

'MiniBatchSize',128,...

'MaxEpochs',100,...

'Shuffle','every-epoch',...

'ValidationData',{valImages,vallabels},...

'ValidationFrequency',50,...

'Verbose',false,...

'Plots','training-progress');

net = trainNetwork(trainImages,trainlabels,layers,options);

%% 测试集验证

load('testdata4_label.mat')

testlabels = categorical(label);

len1 = length(testlabels);

img1(:,:,:,len1) = zeros(227,227,3);

for i = 1:len1

img1(:,:,:,i) = imresize(video(:,:,:,i),[227 227]);

end

pred_labels = classify(net,img1);

test_Acc = sum(pred_labels == testlabels')/len1

cm = confusionchart(testlabels',pred_labels); %推荐使用

cm.ColumnSummary = 'column-normalized';

cm.RowSummary = 'row-normalized';

cm.Title = 'SAPI-FMF Confusion Matrix';

对比实验，使用自定义的CNN，网络结构更简单，不对原始输入进行调整

%% 清空变量和图窗

clear;

close all;

%% 导入数据集

load("traindata2.mat")

len = length(label);  

rng(10)   % 设置seed，便于复现

randIndex = randperm(len);

len1 = round(0.7*len);

trainIndex = randIndex(1:len1);

testIndex = randIndex(len1+1:end);

trainlabels = label(trainIndex);

trainImages = video(:,:,:,trainIndex);

vallabels = label(testIndex);

valImages = video(:,:,:,testIndex);

trainlabels = categorical(trainlabels);

vallabels = categorical(vallabels);

%% 定义神经网络

layers = [

imageInputLayer([32,32,3])

convolution2dLayer(5,32,'Padding','same')

maxPooling2dLayer(2,"Stride",2,"Padding","same")

dropoutLayer(0.3)

convolution2dLayer(3,32)

batchNormalizationLayer

reluLayer

maxPooling2dLayer(2,'Stride',2)

fullyConnectedLayer(2)

softmaxLayer

classificationLayer

];

options = trainingOptions('adam',...

'ExecutionEnvironment','multi-gpu',...

'InitialLearnRate',0.01,...

'MiniBatchSize',128,...

'MaxEpochs',100,...

'Shuffle','every-epoch',...

'ValidationData',{valImages,vallabels},...

'ValidationFrequency',50,...

'Verbose',false,...

'Plots','training-progress');

net = trainNetwork(trainImages,trainlabels,layers,options);

%% 测试集验证

load('testdata4_label.mat')

testlabels = categorical(label);

testImages = video;

len2 = length(testlabels);

pred_labels = classify(net,testImages);

test_Acc = sum(pred_labels == testlabels')/len2

cm = confusionchart(testlabels',pred_labels); %推荐使用

cm.ColumnSummary = 'column-normalized';

cm.RowSummary = 'row-normalized';

cm.Title = 'SAPI-FMF Confusion Matrix';

cm.RowSummary = 'row-normalized';

cm.Title = 'SAPI-FMF Confusion Matrix';

针对二维卷积神经网络的模型，训练效果均不太理想，同时由于所选数据集分布不合理，训练的模型倾向于无故障的分类（实际的数据集比所选数据集要多），模型准确性和泛化能力更差；另一方面，模型参数和优化参数没有充分调整。本文借助于MATLAB进行初步的验证和尝试，后面会使用基于python的深度学习框架pytorch或者Tensorflow进行模型的搭建和训练。