滚动轴承故障检测基于双路神经网络滚动轴承故障诊断系统系统融合原始振动信号和二维时频图像进行多输入（多通道）故障诊断。PyTorch框架实现模型，可单双路-北京尧图网络

基于双路神经网络的滚动轴承故障诊断文章目录基于双路神经网络的滚动轴承故障诊断文字及内容代码仅供参考1. 数据准备2. 完整代码实现2.1 数据加载和预处理2.2 模型构建2.3 模型训练2.4 结果可视化运行步骤实际运行情况如图融合了原始振动信号和二维信号时频图像的多输入(多通道)故障诊断方法单路和双路都可时频图像算法可选小波变换短时傅里叶变换马尔可夫变迁场格拉姆角场S变换递归图等预处理不提供数据集可选凯斯西储大学东南大学江南大学等等使用 Pytorch和Tensorflow框架构建一个基于双路神经网络的滚动轴承故障诊断系统。该系统将融合原始振动信号和二维时频图像进行多输入多通道故障诊断。使用PyTorch框架来实现这个模型单路和双路两种模式都可文字及内容代码仅供参考1. 数据准备假设我们有一个滚动轴承振动信号数据集其中包含正常和故障样本。数据集格式如下datasets/ └── case_western_reserve_university/ ├── train/ │ ├── normal/ │ ├── outer_race_fault/ │ ├── inner_race_fault/ │ └── ball_fault/ ├── val/ │ ├── normal/ │ ├── outer_race_fault/ │ ├── inner_race_fault/ │ └── ball_fault/ └── test/ ├── normal/ ├── outer_race_fault/ ├── inner_race_fault/ └── ball_fault/每个文件夹中包含多个.npy文件每个文件是一个形状为(timesteps, features)的数组。同时我们还需要生成相应的二维时频图像。2. 完整代码实现以下是完整的Python代码包括数据加载、预处理、模型构建、训练和评估。您可以直接运行此代码。2.1 数据加载和预处理[titleData Loading and Preprocessing]importosimportnumpyasnpfromsklearn.preprocessingimportStandardScalerfromkeras.utilsimportto_categoricalimporttorchfromtorch.utils.dataimportDataset,DataLoaderimporttorchvision.transformsastransformsfromscipy.signalimportstftimportmatplotlib.pyplotasplt# Load data from .npy filesdefload_data(data_dir,classes):X[]y[]forlabel,clsinenumerate(classes):class_diros.path.join(data_dir,cls)filesos.listdir(class_dir)forfileinfiles:filepathos.path.join(class_dir,file)datanp.load(filepath)X.append(data)y.append(label)Xnp.array(X)ynp.array(y)returnX,y# Preprocess time-domain signalsdefpreprocess_signals(X_train,X_val,X_test):scalerStandardScaler()n_timesteps,n_featuresX_train.shape[1],X_train.shape[2]X_trainscaler.fit_transform(X_train.reshape(-1,n_features)).reshape(-1,n_timesteps,n_features)X_valscaler.transform(X_val.reshape(-1,n_features)).reshape(-1,n_timesteps,n_features)X_testscaler.transform(X_test.reshape(-1,n_features)).reshape(-1,n_timesteps,n_features)returnX_train,X_val,X_test# Generate STFT images from time-domain signalsdefgenerate_stft_images(X):stft_images[]forsignalinX:f,t,Zxxstft(signal[:,0],fs1000,nperseg256,noverlap128)magnitude_spectrogramnp.abs(Zxx)normalized_image(magnitude_spectrogram-np.min(magnitude_spectrogram))/(np.max(magnitude_spectrogram)-np.min(magnitude_spectrogram))stft_images.append(normalized_image)returnnp.array(stft_images)# Custom dataset classclassBearingDataset(Dataset):def__init__(self,X_signal,X_image,y,transformNone):self.X_signalX_signal self.X_imageX_image self.yy self.transformtransformdef__len__(self):returnlen(self.y)def__getitem__(self,idx):signalself.X_signal[idx]imageself.X_image[idx]labelself.y[idx]ifself.transform:imageself.transform(image)returnsignal,image,label# Load and preprocess classification dataclasses[normal,outer_race_fault,inner_race_fault,ball_fault]data_dirdatasets/case_western_reserve_universityX_train_class,y_train_classload_data(os.path.join(data_dir,train),classes)X_val_class,y_val_classload_data(os.path.join(data_dir,val),classes)X_test_class,y_test_classload_data(os.path.join(data_dir,test),classes)y_train_classto_categorical(y_train_class,num_classeslen(classes))y_val_classto_categorical(y_val_class,num_classeslen(classes))y_test_classto_categorical(y_test_class,num_classeslen(classes))X_train_class,X_val_class,X_test_classpreprocess_signals(X_train_class,X_val_class,X_test_class)# Generate STFT imagesX_train_imagegenerate_stft_images(X_train_class)X_val_imagegenerate_stft_images(X_val_class)X_test_imagegenerate_stft_images(X_test_class)# Create datasets and dataloaderstransformtransforms.Compose([transforms.ToTensor(),])train_datasetBearingDataset(X_train_class,X_train_image,y_train_class,transformtransform)val_datasetBearingDataset(X_val_class,X_val_image,y_val_class,transformtransform)test_datasetBearingDataset(X_test_class,X_test_image,y_test_class,transformtransform)train_loaderDataLoader(train_dataset,batch_size32,shuffleTrue)val_loaderDataLoader(val_dataset,batch_size32,shuffleFalse)test_loaderDataLoader(test_dataset,batch_size32,shuffleFalse)2.2 模型构建[titleModel Construction]importtorch.nnasnnimporttorch.nn.functionalasF# Single-channel CNN modelclassSingleChannelCNN(nn.Module):def__init__(self,input_shape,num_classes):super(SingleChannelCNN,self).__init__()self.conv1nn.Conv1d(in_channelsinput_shape[1],out_channels64,kernel_size5,padding2)self.poolnn.MaxPool1d(kernel_size2)self.dropoutnn.Dropout(p0.2)self.fc1nn.Linear(64*(input_shape[0]//2),128)self.fc2nn.Linear(128,num_classes)defforward(self,x):xself.pool(F.relu(self.conv1(x)))xself.dropout(x)xx.view(x.size(0),-1)xF.relu(self.fc1(x))xself.dropout(x)xself.fc2(x)returnx# Two-channel CNN BiLSTM Attention modelclassDualChannelCNNBiLSTMAtt(nn.Module):def__init__(self,input_shape_signal,input_shape_image,num_classes):super(DualChannelCNNBiLSTMAtt,self).__init__()# Signal branchself.signal_conv1nn.Conv1d(in_channelsinput_shape_signal[1],out_channels64,kernel_size5,padding2)self.signal_poolnn.MaxPool1d(kernel_size2)self.signal_dropoutnn.Dropout(p0.2)self.signal_lstmnn.LSTM(input_size64,hidden_size64,bidirectionalTrue,batch_firstTrue)self.signal_attentionself.attention_3d_block# Image branchself.image_conv1nn.Conv2d(in_channels1,out_channels64,kernel_size(5,5),padding2)self.image_poolnn.MaxPool2d(kernel_size(2,2))self.image_dropoutnn.Dropout(p0.2)self.image_flattennn.Flatten()self.image_fc1nn.Linear(64*(input_shape_image[0]//2)*(input_shape_image[1]//2),128)# Combined branchself.fc_combined1nn.Linear(128*2,256)self.fc_combined2nn.Linear(256,num_classes)defattention_3d_block(self,inputs):# inputs.shape (batch_size, seq_len, feature_dim)single_attention_vectorFalseattention_weightsF.tanh(self.dense(inputs))attention_weightsattention_weights.permute(0,2,1)attention_weightsF.softmax(attention_weights,dim-1)context_vectortorch.bmm(attention_weights,inputs)ifsingle_attention_vector:context_vectorcontext_vector.squeeze(1)else:context_vectorcontext_vector.permute(0,2,1)returncontext_vectordefdense(self,inputs):returnnn.Linear(inputs.shape[-1],1)(inputs)defforward(self,signal,image):# Signal branchsignalself.signal_pool(F.relu(self.signal_conv1(signal)))signalself.signal_dropout(signal)signal,_self.signal_lstm(signal)signal_attself.signal_attention(signal)signal_flatsignal_att.view(signal_att.size(0),-1)# Image branchimageself.image_pool(F.relu(self.image_conv1(image.unsqueeze(1))))imageself.image_dropout(image)image_flatself.image_flatten(image)image_outF.relu(self.image_fc1(image_flat))# Combined branchcombinedtorch.cat((signal_flat,image_out),dim1)combinedF.relu(self.fc_combined1(combined))combinedself.fc_combined2(combined)returncombined# Define modelsinput_shape_signalX_train_class.shape[1:]input_shape_imageX_train_image.shape[1:]single_channel_cnnSingleChannelCNN(input_shape_signal,num_classes)dual_channel_cnn_bilstm_attDualChannelCNNBiLSTMAtt(input_shape_signal,input_shape_image,num_classes)print(single_channel_cnn)print(dual_channel_cnn_bilstm_att)2.3 模型训练[titleModel Training]importtorch.optimasoptim# Function to train a modeldeftrain_model(model,train_loader,val_loader,criterion,optimizer,num_epochs50):devicetorch.device(cudaiftorch.cuda.is_available()elsecpu)model.to(device)best_acc0.0best_model_wtscopy.deepcopy(model.state_dict())forepochinrange(num_epochs):print(fEpoch{epoch}/{num_epochs-1})print(-*10)# Each epoch has a training and validation phaseforphasein[train,val]:ifphasetrain:model.train()# Set model to training modeelse:model.eval()# Set model to evaluate moderunning_loss0.0running_corrects0# Iterate over data.forsignals,images,labelsintrain_loaderifphasetrainelseval_loader:signalssignals.float().to(device)imagesimages.float().to(device)labelstorch.tensor(labels.argmax(axis1)).long().to(device)# Zero the parameter gradientsoptimizer.zero_grad()# Forwardwithtorch.set_grad_enabled(phasetrain):outputsmodel(signals,images)_,predstorch.max(outputs,1)losscriterion(outputs,labels)# Backward optimize only if in training phaseifphasetrain:loss.backward()optimizer.step()# Statisticsrunning_lossloss.item()*signals.size(0)running_correctstorch.sum(predslabels.data)epoch_lossrunning_loss/len(train_loader.datasetifphasetrainelseval_loader.dataset)epoch_accrunning_corrects.double()/len(train_loader.datasetifphasetrainelseval_loader.dataset)print(f{phase}Loss:{epoch_loss:.4f}Acc:{epoch_acc:.4f})# Deep copy the modelifphasevalandepoch_accbest_acc:best_accepoch_acc best_model_wtscopy.deepcopy(model.state_dict())print()print(fBest val Acc:{best_acc:4f})# Load best model weightsmodel.load_state_dict(best_model_wts)returnmodel# Train single-channel CNNcriterionnn.CrossEntropyLoss()optimizer_singleoptim.Adam(single_channel_cnn.parameters(),lr0.001)print(Training Single-Channel CNN...)single_channel_cnntrain_model(single_channel_cnn,train_loader,val_loader,criterion,optimizer_single)# Train dual-channel CNN BiLSTM Attentionoptimizer_dualoptim.Adam(dual_channel_cnn_bilstm_att.parameters(),lr0.001)print(\nTraining Dual-Channel CNN BiLSTM Attention...)dual_channel_cnn_bilstm_atttrain_model(dual_channel_cnn_bilstm_att,train_loader,val_loader,criterion,optimizer_dual)2.4 结果可视化[titleResult Visualization]importcopy# Evaluate models on test setdefevaluate_model(model,test_loader):devicetorch.device(cudaiftorch.cuda.is_available()elsecpu)model.to(device)model.eval()running_corrects0withtorch.no_grad():forsignals,images,labelsintest_loader:signalssignals.float().to(device)imagesimages.float().to(device)labelstorch.tensor(labels.argmax(axis1)).long().to(device)outputsmodel(signals,images)_,predstorch.max(outputs,1)running_correctstorch.sum(predslabels.data)accrunning_corrects.double()/len(test_loader.dataset)print(fTest Accuracy:{acc:.4f})# Evaluate single-channel CNNprint(Evaluating Single-Channel CNN...)evaluate_model(single_channel_cnn,test_loader)# Evaluate dual-channel CNN BiLSTM Attentionprint(\nEvaluating Dual-Channel CNN BiLSTM Attention...)evaluate_model(dual_channel_cnn_bilstm_att,test_loader)# Plot example STFT imagesplt.figure(figsize(12,4))foriinrange(4):plt.subplot(1,4,i1)plt.imshow(X_train_image[i],cmapviridis)plt.title(classes[y_train_class[i].argmax()])plt.axis(off)plt.show()运行步骤确保数据集路径正确将您的数据集放在datasets/case_western_reserve_university目录下。确保.npy文件存在并且格式正确。安装必要的库确保您已经安装了所需的库如numpy,pandas,torch,matplotlib,scipy等。您可以使用以下命令安装这些库pipinstallnumpy pandas torch matplotlib scipy scikit-learn运行代码直接运行上述完整的代码即可完成数据加载、预处理、模型构建、训练和评估。

滚动轴承故障检测基于双路神经网络滚动轴承故障诊断系统系统融合原始振动信号和二维时频图像进行多输入（多通道）故障诊断。PyTorch框架实现模型，可单双路

相关推荐

从点应用走向意图驱动，Joule Work 正在重写 SAP 企业软件的人机交互方式

STM32F405驱动24位DAC1220：高精度SPI接口实战与波形生成

ProofCouncil：基于LLM多智能体协作的数学问题求解框架部署指南

更多资讯

C++游戏开发中enum class实战：从传统enum坑到类型安全迁移

Android纵向滑动页面实战：三容器选型与避坑指南

低附着路面半挂汽车列车折叠失稳仿真与差动制动控制

OMA-VMD自动寻优：告别手动调参的变分模态分解实战

包管理器三剑客：npm、Yarn、pnpm底层原理与工程实践对比

MySQL JDBC驱动版本与连接参数配置实战指南

今日更新

声子晶体仿真：三角晶格与六角晶格的能带计算实战

度假村安保定岗定编：岗位梳理、编制测算与排班设计

COMSOL等离子体模块实战：氩气双层DBD仿真建模全解析

本周精选

交通拥堵预测毕设：基于Python+LSTM的车辆流量预测系统全流程解析

MATLAB/Simulink搭建10机39节点电力系统暂态稳定仿真实战

无人机航拍三维重建全流程：从SfM到网格生成的避坑指南

本月热门

我发现了一个新思路：用 Remotion + Claude Code 像写代码一样自动化生成短视频

Windows下 Codex 中 Chrome 和 Computer Use 插件不可用问题排查及解决参考方式：TaoToken 统一 Key 配置与验证

2026 大模型集体涨价：用 Python 做企业 Token 成本测算与选型避坑（附配置）

读完文章，想聊聊您的网站？