集成异常检测

1.准备数据

获取内部数据集

from paddlets.datasets.repository import get_dataset

ts_data = get_dataset('NAB_TEMP') # label_col: 'label', feature_cols: 'value'

2.数据预处理

切分训练集，然后标准化数据

import paddle
import numpy as np
from paddlets.transform import StandardScaler

#set seed
seed = 2022
paddle.seed(seed)
np.random.seed(seed)
train_tsdata, test_tsdata = ts_data.split(0.15)

#standardize
scaler = StandardScaler('value')
scaler.fit(train_tsdata)
train_tsdata_scaled = scaler.transform(train_tsdata)
test_tsdata_scaled = scaler.transform(test_tsdata)

3.准备模型

为集成模型准备基础模型, in_chunk_len这个参数已经被提取到集成模型中，底层模型可以忽略这个参数。

from paddlets.models.anomaly import AutoEncoder
from paddlets.models.anomaly import VAE
ae_params = {"max_epochs":100}
vae_params = {"max_epochs":100}

4.组装和拟合模型

更多关于集成异常检测模型的信息，请参考 EnsembleAnomaly doc .

例子

from paddlets.ensemble import WeightingEnsembleAnomaly

model = WeightingEnsembleAnomaly(
in_chunk_len=2,
estimators=[(AutoEncoder, ae_params),(VAE, vae_params)],
mode = "voting")

model.fit(train_tsdata_scaled)

5. 模型预测和评估

用训练的模型进行预测和评估

from paddlets.metrics import F1,ACC,Precision,Recall

pred_label = model.predict(test_tsdata_scaled)
lable_name = pred_label.target.data.columns[0]
f1 = F1()(test_tsdata, pred_label)
precision = Precision()(test_tsdata, pred_label)
recall = Recall()(test_tsdata, pred_label)
print ('f1: ', f1[lable_name])
print ('precision: ', precision[lable_name])
print ('recall: ', recall[lable_name])