無監督異常檢測之LSTM組成的AE

我本來就是處理時間序列異常檢測的，之前用了全連接層以及CNN層組成的AE去擬合原始時間序列，發現效果不佳。當利用LSTM組成AE去擬合時間序列時發現，擬合的效果很好。但是，利用重構誤差去做異常檢測這條路依舊不通，因為發現異常曲線的擬合效果也很好……算了，這次先不打算做時間序列異常檢測了。在這里把“基於LSTM的auto-encoder”的代碼分享出來。

代碼參考了Jason Brownlee大佬修改的：具體鏈接我找不到了，當他的博客我還能找到，感興趣自己翻一翻，記得在LSTM網絡那一章

https://machinelearningmastery.com/multivariate-time-series-forecasting-lstms-keras/

from keras.layers import Input, Dense, LSTM
from keras.models import Model
from keras import backend as K
import numpy as np
from pandas import read_csv 
from matplotlib import pyplot
import numpy

from numpy import array
from keras.models import Sequential
from keras.layers import RepeatVector
from keras.layers import TimeDistributed
from keras.utils import plot_model

#導入數據，前8000個正常樣本，剩下的樣本包括正常和異常時間序列，每個樣本是1行48列
dataset = read_csv('randperm_zerone_Dataset.csv')
values = dataset.values
XY= values
n_train_hours1 =7000
n_train_hours3 =8000
trainX=XY[:n_train_hours1,:]
validX =XY[n_train_hours1:n_train_hours3, :]
testX =XY[n_train_hours3:, :]
train3DX = trainX.reshape((trainX.shape[0], trainX.shape[1],1))
valid3DX =validX.reshape((validX.shape[0], validX.shape[1],1))
test3DX = testX.reshape((testX.shape[0],testX.shape[1],1))
# 編碼器
sequence = train3DX
# reshape input into [samples, timesteps, features]
n_in = 48
# define model
model = Sequential()
model.add(LSTM(100, activation='relu', input_shape=(n_in,1)))
model.add(RepeatVector(n_in))
model.add(LSTM(100, activation='relu', return_sequences=True))
model.add(TimeDistributed(Dense(1)))
model.compile(optimizer='adam', loss='mse')
model.summary()
# fit model
history=model.fit(train3DX, train3DX, shuffle=True,epochs=300,validation_data=(valid3DX, valid3DX))
pyplot.plot(history.history['loss'], label='train')
pyplot.plot(history.history['val_loss'], label='valid')
pyplot.legend()
pyplot.show()
# demonstrate recreation
yhat = model.predict(sequence)
ReconstructedData=yhat.reshape((yhat.shape[0], -1))
numpy.savetxt("ReconstructedData.csv", ReconstructedData, delimiter=',')