keras-anomaly-detection 代码分析——本质上就是SAELSTM时间序列预测

Posted 2021-01-15 bonelee

keras-anomaly-detection

Anomaly detection implemented in Keras

The source codes of the recurrent, convolutional and feedforward networks auto-encoders for anomaly detection can be found in keras_anomaly_detection/library/convolutional.py and keras_anomaly_detection/library/recurrent.py and keras_anomaly_detection/library/feedforward.py

The the anomaly detection is implemented using auto-encoder with convolutional, feedforward, and recurrent networks and can be applied to:

• timeseries data to detect timeseries time windows that have anomaly pattern
  • LstmAutoEncoder in keras_anomaly_detection/library/recurrent.py
  • Conv1DAutoEncoder in keras_anomaly_detection/library/convolutional.py
  • CnnLstmAutoEncoder in keras_anomaly_detection/library/recurrent.py
  • BidirectionalLstmAutoEncoder in keras_anomaly_detection/library/recurrent.py
• structured data (i.e., tabular data) to detect anomaly in data records
  • Conv1DAutoEncoder in keras_anomaly_detection/library/convolutional.py
  • FeedforwardAutoEncoder in keras_anomaly_detection/library/feedforward.py
     
    看LSTM的模型吧：

        def create_model(time_window_size, metric):
            model = Sequential()
            model.add(LSTM(units=128, input_shape=(time_window_size, 1), return_sequences=False))
    
            model.add(Dense(units=time_window_size, activation=‘linear‘))
    
            model.compile(optimizer=‘adam‘, loss=‘mean_squared_error‘, metrics=[metric])
            print(model.summary())
    return model
    

    再看feedforward的模型：

        def create_model(self, input_dim):
            encoding_dim = 14
            input_layer = Input(shape=(input_dim,))
    
            encoder = Dense(encoding_dim, activation="tanh",
                            activity_regularizer=regularizers.l1(10e-5))(input_layer)
            encoder = Dense(encoding_dim // 2, activation="relu")(encoder)
    
            decoder = Dense(encoding_dim // 2, activation=‘tanh‘)(encoder)
            decoder = Dense(input_dim, activation=‘relu‘)(decoder)
    
            model = Model(inputs=input_layer, outputs=decoder)
            model.compile(optimizer=‘adam‘,
                          loss=‘mean_squared_error‘,
    metrics=[‘accuracy‘])
    

     CNN的：

        def create_model(time_window_size, metric):
            model = Sequential()
            model.add(Conv1D(filters=256, kernel_size=5, padding=‘same‘, activation=‘relu‘,
                             input_shape=(time_window_size, 1)))
            model.add(GlobalMaxPool1D())
    
            model.add(Dense(units=time_window_size, activation=‘linear‘))
    
            model.compile(optimizer=‘adam‘, loss=‘mean_squared_error‘, metrics=[metric])
            print(model.summary())
            return model
    

     都是将输出设置成自己，异常点就是查看偏离那90%的预测error较大的点。