Pytorch-Transformer轴承故障一维信号分类(三)

import numpy as np
import pandas as pd
from scipy.io import loadmat

file_names = ['0_0.mat','7_1.mat','7_2.mat','7_3.mat','14_1.mat','14_2.mat','14_3.mat','21_1.mat','21_2.mat','21_3.mat']

for file in file_names:
    # 读取MAT文件
    data = loadmat(f'matfiles\\{file}')
    print(list(data.keys()))

第二步，读取MAT文件驱动端数据

python 复制代码

# 采用驱动端数据
data_columns = ['X097_DE_time', 'X105_DE_time', 'X118_DE_time', 'X130_DE_time', 'X169_DE_time',
                'X185_DE_time','X197_DE_time','X209_DE_time','X222_DE_time','X234_DE_time']
columns_name = ['de_normal','de_7_inner','de_7_ball','de_7_outer','de_14_inner','de_14_ball','de_14_outer','de_21_inner','de_21_ball','de_21_outer']
data_12k_10c = pd.DataFrame()
for index in range(10):
    # 读取MAT文件
    data = loadmat(f'matfiles\\{file_names[index]}')
    dataList = data[data_columns[index]].reshape(-1)
    data_12k_10c[columns_name[index]] = dataList[:119808]  # 121048  min: 121265
print(data_12k_10c.shape)
data_12k_10c

第三步，制作数据集

train_set、val_set、test_set 均为按照7：2：1划分训练集、验证集、测试集，最后保存数据

第四步，制作训练集和标签

python 复制代码

# 制作数据集和标签
import torch

# 这些转换是为了将数据和标签从Pandas数据结构转换为PyTorch可以处理的张量，
# 以便在神经网络中进行训练和预测。

def make_data_labels(dataframe):
    '''
        参数 dataframe: 数据框
        返回 x_data: 数据集     torch.tensor
            y_label: 对应标签值  torch.tensor
    '''
    # 信号值
    x_data = dataframe.iloc[:,0:-1]
    # 标签值
    y_label = dataframe.iloc[:,-1]
    x_data = torch.tensor(x_data.values).float()
    y_label = torch.tensor(y_label.values.astype('int64')) # 指定了这些张量的数据类型为64位整数，通常用于分类任务的类别标签
    return x_data, y_label

# 加载数据
train_set = load('train_set')
val_set = load('val_set')
test_set = load('test_set')

# 制作标签
train_xdata, train_ylabel = make_data_labels(train_set)
val_xdata, val_ylabel = make_data_labels(val_set)
test_xdata, test_ylabel = make_data_labels(test_set)
# 保存数据
dump(train_xdata, 'trainX_1024_10c')
dump(val_xdata, 'valX_1024_10c')
dump(test_xdata, 'testX_1024_10c')
dump(train_ylabel, 'trainY_1024_10c')
dump(val_ylabel, 'valY_1024_10c')
dump(test_ylabel, 'testY_1024_10c')

1.2 数据加载，训练数据、测试数据分组，数据分batch

python 复制代码

import torch
from joblib import dump, load
import torch.utils.data as Data
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
# 参数与配置
torch.manual_seed(100)  # 设置随机种子，以使实验结果具有可重复性
device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 有GPU先用GPU训练

# 加载数据集
def dataloader(batch_size, workers=2):
    # 训练集
    train_xdata = load('trainX_1024_10c')
    train_ylabel = load('trainY_1024_10c')
    # 验证集
    val_xdata = load('valX_1024_10c')
    val_ylabel = load('valY_1024_10c')
    # 测试集
    test_xdata = load('testX_1024_10c')
    test_ylabel = load('testY_1024_10c')

    # 加载数据
    train_loader = Data.DataLoader(dataset=Data.TensorDataset(train_xdata, train_ylabel),
                                   batch_size=batch_size, shuffle=True, num_workers=workers, drop_last=True)
    val_loader = Data.DataLoader(dataset=Data.TensorDataset(val_xdata, val_ylabel),
                                 batch_size=batch_size, shuffle=True, num_workers=workers, drop_last=True)
    test_loader = Data.DataLoader(dataset=Data.TensorDataset(test_xdata, test_ylabel),
                                  batch_size=batch_size, shuffle=True, num_workers=workers, drop_last=True)
    return train_loader, val_loader, test_loader

batch_size = 32
# 加载数据
train_loader, val_loader, test_loader = dataloader(batch_size)

2 Transformer分类模型和超参数选取

2.1 定义Transformer分类模型，采用Transformer架构中的编码器：

注意：输入数据进行了堆叠，把一个1*1024 的序列进行划分堆叠成形状为 32 * 32，就使输入序列的长度降下来了

2.2 定义模型参数

python 复制代码

# 模型参数
input_dim = 32 # 输入维度
hidden_dim = 512  # 注意力维度
output_dim  = 10  # 输出维度
num_layers = 4   # 编码器层数
num_heads = 8    # 多头注意力头数
batch_size = 32
# 模型
model = TransformerModel(input_dim, output_dim, hidden_dim, num_layers, num_heads, batch_size)  
model = model.to(device)
loss_function = nn.CrossEntropyLoss(reduction='sum')  # loss
learn_rate = 0.0003
optimizer = torch.optim.Adam(model.parameters(), lr=learn_rate)  # 优化器

2.3 模型结构

3 Transformer模型训练与评估

3.1 模型训练

训练结果

100个epoch，准确率将近90%，Transformer模型分类效果良好，参数过拟合了，适当调整模型参数，降低模型复杂度，还可以进一步提高分类准确率。

注意调整参数：

可以适当增加 Transforme编码器层数和隐藏层的维度，微调学习率；
调整多头注意力的头数，增加更多的 epoch （注意防止过拟合）
可以改变一维信号堆叠的形状（设置合适的长度和维度）

3.2 模型评估

python 复制代码

# 模型 测试集 验证  
import torch.nn.functional as F

# 加载模型
model =torch.load('best_model_transformer.pt')
# model = torch.load('best_model_cnn2d.pt', map_location=torch.device('cpu'))

# 将模型设置为评估模式
model.eval()
# 使用测试集数据进行推断
with torch.no_grad():
    correct_test = 0
    test_loss = 0
    for test_data, test_label in test_loader:
        test_data, test_label = test_data.to(device), test_label.to(device)
        test_output = model(test_data)
        probabilities = F.softmax(test_output, dim=1)
        predicted_labels = torch.argmax(probabilities, dim=1)
        correct_test += (predicted_labels == test_label).sum().item()
        loss = loss_function(test_output, test_label)
        test_loss += loss.item()

test_accuracy = correct_test / len(test_loader.dataset)
test_loss = test_loss / len(test_loader.dataset)
print(f'Test Accuracy: {test_accuracy:4.4f}  Test Loss: {test_loss:10.8f}')

Test Accuracy: 0.9570  Test Loss: 0.12100271