基于1D-CNN的齿轮故障诊断及TSNE可视化

数据来自kaggle。

#加载相关模块
import pandas as pd
import numpy as np
import os
import matplotlib.pyplot as plt
import seaborn as sns
#数据路径
Directory='Gear Data\BrokenTooth'
for root, dirs, files in os.walk(Directory):
for i in range (len(files)):
print(files[i])

path = os.path.join(root,files[0])
path
df_temp = pd.read_csv(path)
df_temp
#时域波形
plt.plot(df_temp.iloc[:,0])
#设置标签
load_col = [int(files[0][5:-4])/100 for j in range(len(df_temp))]
lab='F'
label_col = [lab for j in range(len(df_temp))]
label_col
df_temp['load']=load_col
df_temp['fault']=label_col
df_temp
#数据集处理
def MakeDataset(Directory,lab):
df=pd.DataFrame(columns=['a1','a2','a3','a4'])
for root, dirs, files in os.walk(Directory):
for i in range (len(files)):
path = os.path.join(root,files[i])
df_temp = pd.read_csv(path)
load_col = [int(files[i][5:-4])/100 for j in range(len(df_temp))]
label_col = [lab for j in range(len(df_temp))]
df_temp['load']=load_col
df_temp['fault']=label_col
df = pd.concat([df,df_temp],axis=0)
print(path)

return df
#故障数据
Directory='Gear Data\BrokenTooth'
df_F = MakeDataset(Directory,lab='F')
df_F
#健康数据
Directory='Gear Data\Healthy'
df_H = MakeDataset(Directory,lab='H')
df_H
#将故障数据和健康数据利用concat函数进行连接并输出，便于后续模型使用
df = pd.concat([df_F,df_H],axis=0)
df.to_csv('Gear_Fault_data.csv',index=False)
#数据归一化操作
df = pd.read_csv('Gear_Fault_data.csv')
from sklearn.preprocessing import StandardScaler
scaler=StandardScaler()
df.iloc[:,:-2]=scaler.fit_transform(df.iloc[:,:-2])
##为 CNN 创建数据集
from sklearn.preprocessing import LabelEncoder
from tensorflow.keras.utils import to_categorical
win_len=100 #窗口长度
stride=200 #移动步长
X=[]
Y=[]
for k in ['F','H']:

df_temp_1 = df[df['fault']==k]

for j in (np.arange(0,1,0.1)):
df_temp_2=df_temp_1[df_temp_1['load']==j]
for i in np.arange(0,len(df_temp_2)-(win_len),stride):
X.append(df_temp_2.iloc[i:i+win_len,:-1])
Y.append(df_temp_2.iloc[i+win_len,-1])
#训练数据
X=np.array(X)
X=X.reshape((X.shape[0],X.shape[1],X.shape[2],1))
#X = np.repeat(X, 3, axis=3) # To repeat into 3 chanel format
#标签
Y=np.array(Y)
encoder= LabelEncoder()
encoder.fit(Y)
encoded_Y = encoder.transform(Y)
OHE_Y = to_categorical(encoded_Y)
#训练集尺寸
X.shape
##T-sne可视化
X_pre_cnn = X.reshape(X.shape[0],X.shape[1]*X.shape[2])
from sklearn.manifold import TSNE
X_t_sne = TSNE(n_components=2, learning_rate='auto',verbose=1, perplexity=40, n_iter=300).fit_transform(X_pre_cnn)
tSNEdf = pd.DataFrame(data = X_t_sne, columns = ['t-SNE component 1', 't-SNE component 2'])
tSNEdf['Fault']=Y
#绘制2个主成分
fig, ax = plt.subplots(figsize=(7,7))
sns.scatterplot(x=tSNEdf['t-SNE component 1'],y=tSNEdf['t-SNE component 2'],hue='Fault',
data=tSNEdf,
legend="full",
alpha=0.3)
plt.show()
#训练集和测试集划分
from sklearn.model_selection import train_test_split
X_train,X_test,y_train,y_test = train_test_split(X,OHE_Y,test_size=0.3,shuffle=True)
#构建CNN模型
from tensorflow.keras.models import Sequential,Model
from tensorflow.keras.layers import Input,Dense, Dropout, Flatten
from tensorflow.keras.layers import Conv2D, MaxPooling2D
no_classes = 2 #2个类别
cnn_model = Sequential()
cnn_model.add(Conv2D(32, kernel_size=(20, 3),activation='relu',input_shape=(X.shape[1],X.shape[2],1),padding='same'))
cnn_model.add(MaxPooling2D((20, 2),strides=(5, 5),padding='same'))
cnn_model.add(Conv2D(64, (10, 3), activation='relu',padding='same'))
cnn_model.add(MaxPooling2D(pool_size=(10, 2),strides=(3, 3),padding='same'))
cnn_model.add(Flatten())
cnn_model.add(Dense(128, activation='relu'))

cnn_model.add(Dense(no_classes, activation='softmax'))
cnn_model.summary()
cnn_model.compile(loss='categorical_crossentropy', optimizer='adam',metrics=['accuracy'])
#设置训练参数并训练CNN
batch_size = 128
epochs = 5
history = cnn_model.fit(X_train, y_train, batch_size=batch_size,epochs=epochs,verbose=1,validation_data=(X_test,y_test),shuffle=True)
#保存模型
cnn_model.save('CNN_model_gear.h5')
##模型性能计算
def inv_Transform_result(y_pred):
y_pred = y_pred.argmax(axis=1)
y_pred = encoder.inverse_transform(y_pred)
return y_pred
#预测
y_pred=cnn_model.predict(X_test)
Y_pred=inv_Transform_result(y_pred)
Y_test = inv_Transform_result(y_test)
from sklearn.metrics import confusion_matrix
#混淆矩阵
plt.figure(figsize=(5,5))
cm = confusion_matrix(Y_test, Y_pred)
f = sns.heatmap(cm, annot=True, fmt='d',xticklabels=encoder.classes_,yticklabels=encoder.classes_)
plt.show()
#输出可视化
dummy_cnn = Model(inputs=cnn_model.input,outputs=cnn_model.layers[5].output)
y_viz = dummy_cnn.predict(X_train)
y_viz.shape
from sklearn.manifold import TSNE
X_t_sne = TSNE(n_components=2, learning_rate='auto',verbose=1, perplexity=40, n_iter=300).fit_transform(y_viz)
tSNEdf = pd.DataFrame(data = X_t_sne, columns = ['principal component 1', 'principal component 2'])
tSNEdf['Fault']=inv_Transform_result(y_train)
# 绘制两个主成分分量
fig, ax = plt.subplots(figsize=(10,10))
sns.scatterplot(x=tSNEdf['principal component 1'],y=tSNEdf['principal component 2'],hue='Fault',
data=tSNEdf,
legend="full",
alpha=0.3)
plt.show()
#Flatten层可视化
dummy_cnn = Model(inputs=cnn_model.input,outputs=cnn_model.layers[4].output)
y_viz = dummy_cnn.predict(X_train)
from sklearn.manifold import TSNE
X_t_sne = TSNE(n_components=2, learning_rate='auto',verbose=1, perplexity=40, n_iter=300).fit_transform(y_viz)
tSNEdf = pd.DataFrame(data = X_t_sne, columns = ['t-SNE component 1', 't-SNE component 2'])
tSNEdf['Fault']=inv_Transform_result(y_train)
# 绘制两个主成分
fig, ax = plt.subplots(figsize=(7,7))
sns.scatterplot(x=tSNEdf['t-SNE component 1'],y=tSNEdf['t-SNE component 2'],hue='Fault',
data=tSNEdf,
legend="full",
alpha=0.3)
plt.show()