基于自编码器的滚动轴承失效异常检测（Python）

摘要：# importing packagesimport pandas as pdimport numpy as npfrom sklearn.preprocessing import StandardScalerimport seaborn as snsimpo

# importing packagesimport pandas as pdimport numpy as npfrom sklearn.preprocessing import StandardScalerimport seaborn as snsimport matplotlib.pyplot as pltimport matplotlib.image as mpimg%matplotlib inlinesns.set(color_codes=True)import globimport osimport pickleimport joblibfrom keras.layers import Densefrom keras.models import Model, Sequential, load_modelfrom keras import regularizersdata_dir = '2nd_test'merged_data = pd.DataFrame# Looping over all files from 12th Feb to 19th Febfor filename in os.listdir(data_dir):dataset=pd.read_csv(os.path.join(data_dir, filename), sep='\t')dataset_mean_abs = np.array(dataset.abs.mean)dataset_mean_abs = pd.DataFrame(dataset_mean_abs.reshape(1,4))dataset_mean_abs.index = [filename]merged_data = pd.concat([merged_data, dataset_mean_abs], ignore_index=False)# Renaming columnsmerged_data.columns = ['Bearing 1','Bearing 2','Bearing 3','Bearing 4']# Identifying index as dateTime formatmerged_data.index = pd.to_datetime(merged_data.index, format='%Y.%m.%d.%H.%M.%S')merged_data = merged_data.sort_indexmerged_data.to_csv('merged_dataset_BearingTest_2.csv')df_filename = "merged_dataset_BearingTest_2.csv"df = pd.read_csv(df_filename, index_col="Unnamed: 0")df

# total missing value each columnsfor col in df.columns.tolist:print('{} column missing values: {}'.format(col, df[col].isnull.sum))Bearing 1 column missing values: 0Bearing 2 column missing values: 0Bearing 3 column missing values: 0Bearing 4 column missing values: 0# Visualisation Dataax = df.plot(figsize = (14,5), title="Vibration Activity" , legend = True)ax.set(xlabel="Time", ylabel="Vibration")plt.axvline(x=df.index[-1], linewidth=4, color='b')

df['2004-02-15 14:52:39':'2004-02-16 14:52:39'].plot(figsize = (10,5), title="Zoom Vibration Activity" , legend = True)ax.set(xlabel="Time", ylabel="Vibration")plt.xticks(rotation=25);

# split datasettrain_df = df[:'2004-02-15 20:52:39'] # represent normal operating conditionstest_df = df['2004-02-15 20:52:39':] # leading up to the bearing failuretrain_df.plot(figsize = (10,5), title="Train dataset" , legend = True)plt.xticks(rotation=25);

# Standardize the data.scaler = StandardScalerscaler = scaler.fit(train_df)x_train = pd.DataFrame(scaler.transform(train_df),columns=train_df.columns, index=train_df.index)x_test = pd.DataFrame(scaler.transform(test_df),columns=test_df.columns, index=test_df.index)## modellingfor i, col_name in enumerate(df.columns):print('================= TRAINING %s col_name)curr_xtrain = np.array(x_train[[col_name]])autoencoder = Sequentialautoencoder.add(Dense(9,activation='elu',kernel_initializer='glorot_uniform',kernel_regularizer=regularizers.l2(0.0),input_shape=(curr_xtrain.shape[1],)))autoencoder.add(Dense(3,activation='elu',kernel_initializer='glorot_uniform'))autoencoder.add(Dense(9,activation='elu',kernel_initializer='glorot_uniform'))autoencoder.add(Dense(curr_xtrain.shape[1],kernel_initializer='glorot_uniform'))autoencoder.compile(optimizer='adam', loss="mse")# fitting the modelhist = autoencoder.fit(curr_xtrain, curr_xtrain, batch_size=10,epochs=30,validation_split=0.05,verbose = 1)# save model to model pathautoencoder.save("model/model_%s.h5"%col_name)# visualize training-validation lossplt.plot(hist.history['loss'],'r',label='Training loss')plt.plot(hist.history['val_loss'],'b',label='Validation loss')plt.title("Training-Validation Loss %s"%col_name)plt.legend(loc='upper right')plt.xlabel('Epochs')plt.ylabel('Loss [mse]')plt.ylim([0,.1])plt.savefig('/content/drive/MyDrive/Machine life detection/Image_Auto_Encoder_Anamoly_Detection/plot_training-validation loss_%s.png'%col_name)plt.closeprint('\n================== DONE Epoch 23/3047/47 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0159 - val_loss: 0.0212Epoch 24/3047/47 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0110 - val_loss: 0.0223Epoch 25/3047/47 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0121 - val_loss: 0.0199Epoch 26/3047/47 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0107 - val_loss: 0.0183Epoch 27/3047/47 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0109 - val_loss: 0.0177Epoch 28/3047/47 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0089 - val_loss: 0.0161Epoch 29/3047/47 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0081 - val_loss: 0.0155Epoch 30/3047/47 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0091 - val_loss: 0.0144from keras.models import load_modelfrom keras.losses import MeanSquaredErrormodels = {}for col_name in df.columns:models[col_name] = load_model(f"/content/model/model_{col_name}.h5", custom_objects={'mse': MeanSquaredError})# visualization x_test prediction# leading up to the bearing failurefor col in df.columns:x_test_array = np.array(x_test[col])x_test_pred = models[col].predict(x_test_array)df_test_pred = pd.DataFrame(x_test_pred)df_test_pred["Predicted"] = x_test_preddf_test_pred["Actual"] = x_test_arraydf_test_pred.index = x_test[col].indexdf_test_pred.plot(figsize = (10,5), title=f"Actual vs Predicted {col}" , legend = True)plt.xticks(rotation=15)plt.show

# Get train MAE loss.list_threshold = for col in df.columns:x_train_pred = models[col].predict(np.array(x_train[col]))train_mae_loss = np.mean(np.abs(x_train_pred - np.array(x_train[col])),axis=1)plt.hist(train_mae_loss, bins=50)plt.xlabel("Train MAE loss")plt.ylabel("No of samples")plt.title(col)plt.show# Get reconstruction loss thresholdthreshold = np.max(train_mae_loss)list_threshold.append(threshold)print(f"Reconstruction error threshold {col}: ", threshold)

for col in df.columns:window_percentage = 20k = int(len(x_test[col]) * (window_percentage/100))N = len(x_test[col])get_bands = lambda data : (np.mean(x_train[col]) + 3*np.std(x_train[col]),np.mean(x_train[col]) - 3*np.std(x_train[col]))bands = [get_bands(x_test[col][range(0 if i - k upper) | (x_test[col]