hz-application-ml/LIB/MIDDLE/Anomaly_Detection/V1_0_0/anomalyPCA.py

#热失控预警：PCA异常指数

import pandas as pd
import numpy as np
from scipy.signal import savgol_filter
from sklearn.preprocessing import RobustScaler
from sklearn.decomposition import PCA

#筛选特征
def makedataset1(df_data):
    df_data=df_data.drop(['Unnamed: 0','总电流[A]','GSM信号','外电压','SOH[%]','开关状态','充电状态','故障等级','故障代码','绝缘电阻','上锁状态','加热状态','单体均衡状态','总输出状态'],axis=1,errors='ignore')
    df_data=df_data.drop(["单体温度"+str(i) for i in range(1,5)],axis=1,errors='ignore')
    df_data=df_data.drop(["其他温度"+str(i) for i in range(1,7)],axis=1,errors='ignore')
    listV=[s for s in list(df_data) if '单体电压' in s]
    for i in range(1,len(listV)+1):
        df_data=df_data[(df_data['单体电压'+str(i)]>2200) & (df_data['单体电压'+str(i)]<4800)]
    df_data=df_data[df_data['SOC[%]']>20]
    df_data['时间']=[df_data.loc[i,'时间戳'][0:15] for i in df_data.index]
    df_data=df_data.drop('时间戳',axis=1)
    data_set=df_data.groupby('时间').mean(False)
    for k in data_set.columns:
        data_set[k]=savgol_filter(data_set[k],3,2)
    return data_set

#新建统计特征
def makedataset2(df_data):
    data_set=makedataset1(df_data)
    listV=[s for s in list(df_data) if '单体电压' in s]
    data_set["最低单体电压"]=data_set[["单体电压"+str(i) for i in range(1,len(listV)+1)]].min(axis=1)
    data_set["最高单体电压"]=data_set[["单体电压"+str(i) for i in range(1,len(listV)+1)]].max(axis=1)
    data_set["平均单体电压"]=data_set[["单体电压"+str(i) for i in range(1,len(listV)+1)]].mean(axis=1)
    data_set["最大单体压差"]=[data_set.loc[k,"最高单体电压"]-data_set.loc[k,"最低单体电压"] for k in data_set.index]
    data_set["低压差"]=[data_set.loc[k,"平均单体电压"]-data_set.loc[k,"最低单体电压"] for k in data_set.index]
    data_set=data_set.drop(["单体电压"+str(i) for i in range(1,len(listV)+1)],axis=1)
    return data_set

#标准化
def process(data_set):
    features=data_set.columns
    sX=RobustScaler(copy=True)
    data_set2=data_set.copy()
    data_set2.loc[:,features]=sX.fit_transform(data_set2[features])
    return data_set2

#异常指数函数
def anomalyScores(originalDF,reducedDF):
    loss=np.sum((np.array(originalDF)-np.array(reducedDF))**2,axis=1)
    loss=pd.Series(data=loss,index=originalDF.index)
    loss=(loss-np.min(loss))/(np.max(loss)-np.min(loss))
    return loss

#建立PCA模型
def anomalyPCA(x_train_pro):
    n_components=4
    whiten=True
    random_state=2
    pca=PCA(n_components=n_components,whiten=whiten,random_state=random_state)
    pca.fit(x_train_pro)
    return pca

#判断PCA异常指数
def transform(df_data_pro,model,df_data):
    #降维
    X_train=model.transform(df_data_pro)
    X_train=pd.DataFrame(data=X_train,index=df_data_pro.index)
    #还原
    X_train_inverse=model.inverse_transform(X_train)
    X_train_inverse=pd.DataFrame(data=X_train_inverse,index=df_data_pro.index)
    #异常指数
    anomalyScoresModel=anomalyScores(df_data_pro,X_train_inverse)
    df_data2=df_data.copy()
    if len(anomalyScoresModel)>15:
        anomalyScoresModel=savgol_filter(anomalyScoresModel,15,3)
        df_data2['anomalyScores_'+str(model)]=anomalyScoresModel
    else:
        df_data2['anomalyScores_'+str(model)]=0
    return df_data2

#判断离群
def detect_outliers(data,pred,threshold=3):
    anomaly=data['anomalyScores_PCA(n_components=4, random_state=2, whiten=True)']
    anomalypred=pred['anomalyScores_PCA(n_components=4, random_state=2, whiten=True)']
    mean_d=np.mean(anomaly.values)
    std_d=np.std(anomaly.values)
    max_score=np.max(anomaly.values)
    outliers2=pd.DataFrame()
    for k in anomalypred.index:
        z_score= (anomalypred[k]-mean_d)/std_d
        if (np.abs(z_score) >threshold) & (anomalypred[k]>max_score):
            outliers2=outliers2.append(pred[anomalypred.values==anomalypred[k]])
    return outliers2

#训练模型
def train_model(data_train):
    x_train1=makedataset1(data_train) 
    x_train2=makedataset2(data_train)  
    x_train_pro1=process(x_train1) 
    x_train_pro2=process(x_train2) 
    pca1=anomalyPCA(x_train_pro1) 
    pca2=anomalyPCA(x_train_pro2) 
    res1=transform(x_train_pro1,pca1,x_train1)
    res2=transform(x_train_pro2,pca2,x_train2)
    return pca1,pca2,res1,res2

#预测
def prediction(data_test,pca1,pca2):
    x_test1=makedataset1(data_test) 
    x_test2=makedataset2(data_test) 
    x_test_pro1=process(x_test1) 
    x_test_pro2=process(x_test2) 
    pred1=transform(x_test_pro1,pca1,x_test1)
    pred2=transform(x_test_pro2,pca2,x_test2)
    return pred1,pred2

def boxplot_fill(res2):
    col=res2['低压差']
    # 计算iqr：数据四分之三分位值与四分之一分位值的差
    iqr=col.quantile(0.75)-col.quantile(0.25)
    # 根据iqr计算异常值判断阈值
    u_th=col.quantile(0.75) + 2*iqr # 上界
    return u_th

#判定异常
def check_anomaly(outliers1,outliers2,res2):
    outliers=pd.merge(outliers1,outliers2,on='时间')
    outliers=outliers[outliers['SOC[%]_x']>50]
    outliers=outliers.drop(['总电压[V]_y','单体压差_y','SOC[%]_y'],axis=1)
    u_th=boxplot_fill(res2)
    outliers=outliers[outliers['低压差']>u_th]
    return outliers