Acoustics-Based机器故障识别

downloadFolder = matlab.internal.examples.download金宝appSupportFile (“音频”,“AirCompressorDataset / AirCompressorDataset.zip”);dataFolder = tempdir;解压缩数据集(downloadFolder dataFolder) = fullfile (dataFolder,“AirCompressorDataset”);

广告= audioDatastore(数据集,IncludeSubfolders = true, LabelSource =“foldernames”);[adsTrain, adsValidation] = splitEachLabel(广告,0.9,0.1);countEachLabel (adsTrain)

countEachLabel (adsValidation)

adsTrain = shuffle (adsTrain);adsValidation = shuffle (adsValidation);

speedupExample =假;如果speedupExample adsTrain = splitEachLabel (adsTrain 20);结束

[sampleData, sampleDataInfo] =阅读(adsTrain);fs = sampleDataInfo.SampleRate;soundsc (sampleData fs)情节(sampleData)包含(“样本”)ylabel (“振幅”)标题(”状态:“+字符串(sampleDataInfo.Label)轴紧

trainFeatures =细胞(1,元素个数(adsTrain.Files));windowLength = 512;overlapLength = 0;aFE = audioFeatureExtractor (SampleRate = fs,…窗口=汉明(windowLength,“周期”),…OverlapLength = OverlapLength,…spectralCentroid = true,…spectralCrest = true,…spectralDecrease = true,…spectralEntropy = true,…spectralFlatness = true,…spectralFlux = false,…spectralKurtosis = true,…spectralRolloffPoint = true,…spectralSkewness = true,…spectralSlope = true,…spectralSpread = true);重置(adsTrain)抽搐为指数= 1:元素个数(adsTrain.Files)数据=阅读(adsTrain);trainFeatures{指数}=(提取(aFE、数据))';结束disp (“训练集的特征提取”+ toc +“秒”。);

trainLabels = adsTrain.Labels;numMixesPerInstance =2;(augData抽搐,augLabels] =混合物(trainFeatures, trainLabels numMixesPerInstance);trainLabels =猫(1 trainLabels augLabels);trainFeatures =猫(2 trainFeatures augData);disp (“功能增强的火车”+ toc +“秒”。);

validationFeatures =细胞(1,元素个数(adsValidation.Files));重置(adsValidation)抽搐为指数= 1:元素个数(adsValidation.Files)数据=阅读(adsValidation);validationFeatures{指数}=(提取(aFE、数据))';结束disp (“验证设置了“特征提取+ toc +“秒”。);

downloadPretrainedSystem =假;如果downloadPretrainedSystem downloadFolder = matlab.internal.examples.download金宝appSupportFile (“音频”,“AcousticsBasedMachineFaultRecognition / AcousticsBasedMachineFaultRecognition.zip”);dataFolder = tempdir;解压缩(downloadFolder dataFolder) netFolder = fullfile (dataFolder,“AcousticsBasedMachineFaultRecognition”);目录(netFolder)结束

numHiddenUnits =One hundred.;dropProb =0.2;层= […sequenceInputLayer (aFE.FeatureVectorLength正常化=“zscore”)lstmLayer (numHiddenUnits OutputMode =“序列”)dropoutLayer (dropProb) lstmLayer (numHiddenUnits OutputMode =“最后一次”)fullyConnectedLayer(元素个数(独特(adsTrain.Labels))) softmaxLayer classificationLayer];

miniBatchSize =128年;validationFrequency =地板(元素个数(trainFeatures) / miniBatchSize);选择= trainingOptions (“亚当”,…MiniBatchSize = MiniBatchSize,…MaxEpochs = 35,…情节=“训练进步”,…Verbose = false,…洗牌=“every-epoch”,…LearnRateSchedule =“分段”,…LearnRateDropPeriod = 30,…LearnRateDropFactor = 0.1,…ValidationData = {validationFeatures, adsValidation.Labels},…ValidationFrequency = ValidationFrequency);

airCompNet = trainNetwork (trainFeatures、trainLabels层,选择);

validationResults =分类(airCompNet validationFeatures);confusionchart (validationResults adsValidation.Labels,…Title =准确性:“+的意思是(validationResults = = adsValidation.Labels) * 100 +“(%)”);

文件名= fullfile (pwd,“extractAudioFeatures”);generateMATLABFunction (aFE,文件名,IsStreaming = true);

保存(“AirCompressorFaultRecognitionModel.mat”,“airCompNet”)

类型recognizeAirCompressorFault.m

重置(adsValidation) N = 10;标签=类别(ads.Labels);numLabels =元素个数(标签);%创建一个dsp。AsyncBuffer阅读音频流的方式audioSource = dsp.AsyncBuffer;%创建一个dsp。AsyncBuffer积累分数scoreBuffer = dsp.AsyncBuffer;%创建一个dsp。AsyncBuffer记录执行时间。timingBuffer = dsp.AsyncBuffer;%预先分配数组来存储结果streamingResults =分类(0 (N, 1));%遍历文件为fileIdx = 1: N%读一个音频文件,把它放在源缓冲区(数据、dataInfo) =阅读(adsValidation);写(audioSource、数据);%在帧内循环rs = true;而audioSource。NumUnreadSamples > = windowLength%得到音频数据的帧x =阅读(audioSource windowLength);%应用流分类器函数抽搐得分= recognizeAirCompressorFault (x, rs);写(timingBuffer toc);%存储分数进行分析写(scoreBuffer,得分);rs = false;结束重置(audioSource)%为该文件存储类的结果成绩=阅读(scoreBuffer);[~,结果]= max(分数(最终,:),[],2);streamingResults (fileIdx) =分类(标签(结果));随着时间的推移%情节分数比较图绘制(分数)% #好< * NASGU >传奇(字符串(airCompNet.Layers(结束). class),位置=“西北”)包含(“时间步”)ylabel (“分数”)标题(“已知的Label = "+字符串(dataInfo.Label),“预测标签= "+字符串(streamingResults (fileIdx))))结束

testError =意味着(validationResults (1: N) ~ = streamingResults);disp (“流分类器之间的误差和非:“+ testError * 100 +“(%)”)

executionTime =阅读(timingBuffer) * 1000;预算= (windowLength / aFE.SampleRate) * 1000;情节(executionTime“o”)标题(“执行时间”每帧)包含(“帧数”)ylabel (“时间(ms)”)yline(预算,”“,“预算”线宽= 2)

函数[augData, augLabels] =混合物(数据、标签、numMixesPerInstance) augData =细胞(1,元素个数(数据)* numMixesPerInstance);augLabels = repelem(标签,numMixesPerInstance);kk = 1;为2 = 1:元素个数(数据)为jj = 1: numMixesPerInstanceλ= max (min (randn. / 10) + 0.5 (1), 0);%找到所有可用的数据用不同的标签。availableData =找到(标签~ =标签(ii));%随机选择一个可用的数据用不同的标签。numAvailableData =元素个数(availableData);,numAvailableData idx =兰迪([1]);%的组合。augData {kk} =λ*{2}+(1λ)*数据{availableData (idx)};%指定标签作为随机设定的λ。如果λ<兰德augLabels (kk) =标签(availableData (idx));其他的augLabels (kk) =标签(ii);结束kk = kk + 1;结束结束结束