1. Introduction
In the process of using household water heaters, residents will form different usage habits due to regional climate, different regions, and user age and gender differences. If home furnishing companies can deeply understand the usage habits of their products in different user groups, and develop functions that meet customer needs and habits, they will be able to open up new markets.
This article will build a bathing event recognition model based on the BP neural network algorithm, and then identify the bathing events of users in different regions, and then compare different customer groups according to the recognition results to provide the most suitable personalized products and improve the intelligent research and development of new products And formulate corresponding marketing strategies.
2. Data Exploration and Analysis
Explore and analyze the flow status of the water heater, among which the attributes of "whether there is water flow" and "water flow" can most intuitively reflect the water flow of the water heater, and explore and analyze these two attributes.
import pandas as pd
import matplotlib.pyplot as plt
inputfile ='./data/original_data.xls' #'./original_data.xls' # 输入的数据文件
data = pd.read_excel(inputfile) # 读取数据
# 查看有无水流的分布
# 数据提取
lv_non = pd.value_counts(data['有无水流'])['无']
lv_move = pd.value_counts(data['有无水流'])['有']
# 绘制条形图
fig = plt.figure(figsize = (6 ,5)) # 设置画布大小
plt.rcParams['font.sans-serif'] = 'SimHei' # 设置中文显示
plt.rcParams['axes.unicode_minus'] = False
plt.bar([0,1], height=[lv_non,lv_move], width=0.4, alpha=0.8, color='skyblue')
plt.xticks([index for index in range(2)], ['无','有'])
plt.xlabel('水流状态')
plt.ylabel('记录数')
plt.title('不同水流状态记录数(3001)')
plt.show()
plt.close()
# 查看水流量分布
water = data['水流量']
# 绘制水流量分布箱型图
fig = plt.figure(figsize = (5 ,8))
plt.boxplot(water,
patch_artist=True,
labels = ['水流量'], # 设置x轴标题
boxprops = {'facecolor':'lightblue'}) # 设置填充颜色
plt.title('水流量分布箱线图(3001)')
# 显示y坐标轴的底线
plt.grid(axis='y')
plt.show()Get the following bar and box plots
It can be seen that there are obviously more records in the state of no water flow than in the state of water flow. The box is close to 0, indicating that there are many records of no water flow, and the distribution of water flow is consistent with the distribution of water flow state.
3. Data preprocessing
3.1 Attribute reduction
Water heaters collect a lot of water consumption data, and the main object of analysis is water heater users, so redundant attributes will be deleted, for example: "water heater number", "whether there is water flow", "energy-saving mode".
import pandas as pd
import numpy as np
data = pd.read_excel('./data/original_data.xls')
print('初始状态的数据形状为：', data.shape)
# 删除热水器编号、有无水流、节能模式属性
data.drop(labels=["热水器编号","有无水流","节能模式"],axis=1,inplace=True)
print('删除冗余属性后的数据形状为：', data.shape)
data.to_csv('./data/water_heart.csv',index=False)The result is as follows
3.2 Classification of water use events
Water heater users’ water consumption data is stored in the database, which records various water consumption events, including bathing, washing hands, brushing teeth, washing face, laundry, washing vegetables, etc. What continuous data should be divided in a large number of state records in this case When a complete water use event. The time division of water use is mainly divided into two steps, that is, to determine the time interval of a single water use, and to calculate the time of two adjacent records.
# 读取数据
data = pd.read_csv('./data/water_heart.csv')
# 划分用水事件
threshold = pd.Timedelta('4 min') # 阈值为4分钟
data['发生时间'] = pd.to_datetime(data['发生时间'], format = '%Y%m%d%H%M%S') # 转换时间格式
data = data[data['水流量'] > 0] # 只要流量大于0的记录
sjKs = data['发生时间'].diff() > threshold # 相邻时间向前差分，比较是否大于阈值
sjKs.iloc[0] = True # 令第一个时间为第一个用水事件的开始事件
sjJs = sjKs.iloc[1:] # 向后差分的结果
sjJs = pd.concat([sjJs,pd.Series(True)]) # 令最后一个时间作为最后一个用水事件的结束时间
# 创建数据框，并定义用水事件序列
sj = pd.DataFrame(np.arange(1,sum(sjKs)+1),columns = ["事件序号"])
sj["事件起始编号"] = data.index[sjKs == 1]+1 # 定义用水事件的起始编号
sj["事件终止编号"] = data.index[sjJs == 1]+1 # 定义用水事件的终止编号
print('当阈值为4分钟的时候事件数目为：',sj.shape[0])
sj.to_csv('./data/sj.csv',index = False)
The result is as follows
3.3 Determine the duration threshold of a single water use event
For the data of a water heater user, the water use events are divided according to different thresholds to obtain the corresponding number of events, and the number of time is obtained by threshold changes and divisions. Count the number of water use events under each threshold, and then find the optimal threshold through threshold optimization.
# 确定单次用水事件时长阈值
n = 4 # 使用以后四个点的平均斜率
threshold = pd.Timedelta(minutes=5) # 专家阈值
data['发生时间'] = pd.to_datetime(data['发生时间'], format='%Y%m%d%H%M%S')
data = data[data['水流量'] > 0] # 只要流量大于0的记录
# 自定义函数：输入划分时间的时间阈值，得到划分的事件数
def event_num(ts):
d = data['发生时间'].diff() > ts # 相邻时间作差分，比较是否大于阈值
return d.sum() + 1 # 这样直接返回事件数
dt = [pd.Timedelta(minutes=i) for i in np.arange(1, 9, 0.25)]
h = pd.DataFrame(dt, columns=['阈值']) # 转换数据框，定义阈值列
h['事件数'] = h['阈值'].apply(event_num) # 计算每个阈值对应的事件数
h['斜率'] = h['事件数'].diff()/0.25 # 计算每两个相邻点对应的斜率
h['斜率指标']= h['斜率'].abs().rolling(4).mean() # 往前取n个斜率绝对值平均作为斜率指标
ts = h['阈值'][h['斜率指标'].idxmin() - n]
# 用idxmin返回最小值的Index，由于rolling_mean()计算的是前n个斜率的绝对值平均
# 所以结果要进行平移（-n）
if ts > threshold:
ts = pd.Timedelta(minutes=4)
print('计算出的单次用水时长的阈值为：',ts)
The result is as follows
3.4 Attribute construction
1) Construct water use duration and frequency attributes
The duration of water use for different water use events is one of the basic attributes. The attributes related to the duration of water use can only distinguish some water use events, and the water use pauses and frequencies of different water use events are also different. Build water usage duration and water usage frequency attributes.
data = pd.read_csv('./data/water_heart.csv',encoding='utf-8') # 读取热水器使用数据记录
sj = pd.read_csv('./data/sj.csv') # 读取用水事件记录
# 转换时间格式
data["发生时间"] = pd.to_datetime(data["发生时间"],format="%Y%m%d%H%M%S")
# 构造特征：总用水时长
timeDel = pd.Timedelta("0.5 sec")
sj["事件开始时间"] = data.iloc[sj["事件起始编号"]-1,0].values- timeDel
sj["事件结束时间"] = data.iloc[sj["事件终止编号"]-1,0].values + timeDel
sj['洗浴时间点'] = [i.hour for i in sj["事件开始时间"]]
sj["总用水时长"] = np.int64(sj["事件结束时间"] - sj["事件开始时间"])/1000000000 + 1
# 构造用水停顿事件
# 构造特征“停顿开始时间”、“停顿结束时间”
# 停顿开始时间指从有水流到无水流，停顿结束时间指从无水流到有水流
for i in range(len(data)-1):
if (data.loc[i,"水流量"] != 0) & (data.loc[i + 1,"水流量"] == 0) :
data.loc[i + 1,"停顿开始时间"] = data.loc[i +1, "发生时间"] - timeDel
if (data.loc[i,"水流量"] == 0) & (data.loc[i + 1,"水流量"] != 0) :
data.loc[i,"停顿结束时间"] = data.loc[i , "发生时间"] + timeDel
# 提取停顿开始时间与结束时间所对应行号，放在数据框Stop中
indStopStart = data.index[data["停顿开始时间"].notnull()]+1
indStopEnd = data.index[data["停顿结束时间"].notnull()]+1
Stop = pd.DataFrame(data={"停顿开始编号":indStopStart[:-1],
"停顿结束编号":indStopEnd[1:]})
# 计算停顿时长，并放在数据框stop中，停顿时长=停顿结束时间-停顿结束时间
Stop["停顿时长"] = np.int64(data.loc[indStopEnd[1:]-1,"停顿结束时间"].values-
data.loc[indStopStart[:-1]-1,"停顿开始时间"].values)/1000000000
# 将每次停顿与事件匹配,停顿的开始时间要大于事件的开始时间，
# 且停顿的结束时间要小于事件的结束时间
for i in range(len(sj)):
Stop.loc[(Stop["停顿开始编号"] > sj.loc[i,"事件起始编号"]) &
(Stop["停顿结束编号"] < sj.loc[i,"事件终止编号"]),"停顿归属事件"]=i+1
# 删除停顿次数为0的事件
Stop = Stop[Stop["停顿归属事件"].notnull()]
# 构造特征 用水事件停顿总时长、停顿次数、停顿平均时长、
# 用水时长，用水/总时长
stopAgg = Stop.groupby("停顿归属事件").agg({"停顿时长":sum,"停顿开始编号":len})
sj.loc[stopAgg.index - 1,"总停顿时长"] = stopAgg.loc[:,"停顿时长"].values
sj.loc[stopAgg.index-1,"停顿次数"] = stopAgg.loc[:,"停顿开始编号"].values
sj.fillna(0,inplace=True) # 对缺失值用0插补
stopNo0 = sj["停顿次数"] != 0 # 判断用水事件是否存在停顿
sj.loc[stopNo0,"平均停顿时长"] = sj.loc[stopNo0,"总停顿时长"]/sj.loc[stopNo0,"停顿次数"]
sj.fillna(0,inplace=True) # 对缺失值用0插补
sj["用水时长"] = sj["总用水时长"] - sj["总停顿时长"] # 定义特征用水时长
sj["用水/总时长"] = sj["用水时长"] / sj["总用水时长"] # 定义特征 用水/总时长
print('用水事件用水时长与频率特征构造完成后数据的特征为：\n',sj.columns)
print('用水事件用水时长与频率特征构造完成后数据的前5行5列特征为：\n',sj.iloc[:5,:5])
The result is as follows
2) Construct water consumption and fluctuation attributes
In addition to the duration, pause, and frequency of water use, water consumption is also an important attribute to identify whether the event is a bathing event. At the same time, fluctuations in water use are also the key to distinguish different water use events. Construct water consumption and water fluctuation attributes on the basis of water use duration and frequency attributes.
data["水流量"] = data["水流量"] / 60 # 原单位L/min，现转换为L/sec
sj["总用水量"] = 0 # 给总用水量赋一个初始值0
for i in range(len(sj)):
Start = sj.loc[i,"事件起始编号"]-1
End = sj.loc[i,"事件终止编号"]-1
if Start != End:
for j in range(Start,End):
if data.loc[j,"水流量"] != 0:
sj.loc[i,"总用水量"] = (data.loc[j + 1,"发生时间"] -
data.loc[j,"发生时间"]).seconds* \
data.loc[j,"水流量"] + sj.loc[i,"总用水量"]
sj.loc[i,"总用水量"] = sj.loc[i,"总用水量"] + data.loc[End,"水流量"] * 2
else:
sj.loc[i,"总用水量"] = data.loc[Start,"水流量"] * 2
sj["平均水流量"] = sj["总用水量"] / sj["用水时长"] # 定义特征 平均水流量
# 构造特征：水流量波动
# 水流量波动=∑(((单次水流的值-平均水流量)^2)*持续时间)/用水时长
sj["水流量波动"] = 0 # 给水流量波动赋一个初始值0
for i in range(len(sj)):
Start = sj.loc[i,"事件起始编号"] - 1
End = sj.loc[i,"事件终止编号"] - 1
for j in range(Start,End + 1):
if data.loc[j,"水流量"] != 0:
slbd = (data.loc[j,"水流量"] - sj.loc[i,"平均水流量"])**2
slsj = (data.loc[j + 1,"发生时间"] - data.loc[j,"发生时间"]).seconds
sj.loc[i,"水流量波动"] = slbd * slsj + sj.loc[i,"水流量波动"]
sj.loc[i,"水流量波动"] = sj.loc[i,"水流量波动"] / sj.loc[i,"用水时长"]
# 构造特征：停顿时长波动
# 停顿时长波动=∑(((单次停顿时长-平均停顿时长)^2)*持续时间)/总停顿时长
sj["停顿时长波动"] = 0 # 给停顿时长波动赋一个初始值0
for i in range(len(sj)):
if sj.loc[i,"停顿次数"] > 1: # 当停顿次数为0或1时，停顿时长波动值为0，故排除
for j in Stop.loc[Stop["停顿归属事件"] == (i+1),"停顿时长"].values:
sj.loc[i,"停顿时长波动"] = ((j - sj.loc[i,"平均停顿时长"])**2) * j + \
sj.loc[i,"停顿时长波动"]
sj.loc[i,"停顿时长波动"] = sj.loc[i,"停顿时长波动"] / sj.loc[i,"总停顿时长"]
print('用水量和波动特征构造完成后数据的特征为：\n',sj.columns)
print('用水量和波动特征构造完成后数据的前5行5列特征为：\n',sj.iloc[:5,:5])The result is as follows
3.5 Screening Candidate Bathing Events
The identification of a bathing event is based on the identification of a water use event, that is, identifying which water use events are bathing events from the divided water use events. Candidate bathing events are screened based on the constructed water duration and water consumption attributes.
sj_bool = (sj['用水时长'] >100) & (sj['总用水时长'] > 120) & (sj['总用水量'] > 5)
sj_final = sj.loc[sj_bool,:]
sj_final.to_excel('./data/sj_final.xlsx',index=False)
print('筛选出候选洗浴事件前的数据形状为：',sj.shape)
print('筛选出候选洗浴事件后的数据形状为：',sj_final.shape)The result is as follows
Before screening, the total number of water use events was 172, and after screening, there were 75 water use events remaining.
4. Model Construction
Based on the modeling sample data, a BP neural network model is established to identify bathing events. The construction of BP neural network model needs to pay attention to the magnitude difference between the attributes of the data itself, so it needs to be standardized to eliminate the magnitude difference. In order to facilitate the subsequent application of the model, the model can be saved with the joblib.dump function.
import pandas as pd
from sklearn.preprocessing import StandardScaler
from sklearn.neural_network import MLPClassifier
from sklearn.externals import joblib
# 读取数据
Xtrain = pd.read_excel('./data/sj_final.xlsx')
ytrain = pd.read_excel('./data/water_heater_log.xlsx')
test = pd.read_excel('./data/test_data.xlsx')
# 训练集测试集区分。
x_train, x_test, y_train, y_test = Xtrain.iloc[:,5:],test.iloc[:,4:-1],\
ytrain.iloc[:,-1],test.iloc[:,-1]
# 标准化
stdScaler = StandardScaler().fit(x_train)
x_stdtrain = stdScaler.transform(x_train)
x_stdtest = stdScaler.transform(x_test)
# 建立模型
bpnn = MLPClassifier(hidden_layer_sizes = (17,10), max_iter = 200, solver = 'lbfgs',random_state=50)
bpnn.fit(x_stdtrain, y_train)
# 保存模型
joblib.dump(bpnn,'./data/water_heater_nnet.m')
print('构建的模型为：\n',bpnn)The result is as follows
When training the BP neural network, the parameters of the neural network were optimized, and it was found that the training effect of the neural network with two hidden layers was better, and the training effect was better when the hidden nodes of the two hidden layers were 17 and 10 respectively. good.
5. Model checking
Combined with the relevant knowledge of model evaluation, it is more objective and accurate to use precision rate (precision), recall rate (recall) and F1 value to measure the effect of model evaluation. At the same time, combined with the ROC curve, the effect of the model can be evaluated more intuitively.
# 模型评价
from sklearn.metrics import classification_report
from sklearn.metrics import roc_curve
from sklearn.externals import joblib
import matplotlib.pyplot as plt
bpnn = joblib.load('./data/water_heater_nnet.m') # 加载模型
y_pred = bpnn.predict(x_stdtest) # 返回预测结果
print('神经网络预测结果评价报告：\n',classification_report(y_test,y_pred))
# 绘制roc曲线图
plt.rcParams['font.sans-serif'] = 'SimHei' # 显示中文
plt.rcParams['axes.unicode_minus'] = False # 显示负号
fpr, tpr, thresholds = roc_curve(y_pred,y_test) # 求出TPR和FPR
plt.figure(figsize=(6,4)) # 创建画布
plt.plot(fpr,tpr) # 绘制曲线
plt.title('用户用水事件识别ROC曲线(3001)') # 标题
plt.xlabel('FPR') # x轴标签
plt.ylabel('TPR') # y轴标签
plt.savefig('./data/用户用水事件识别ROC曲线.png') # 保存图片
plt.show() # 显示图形The result is as follows
It can be obtained that the precision rate (precision) in bathing event recognition is very high, as high as 90%, and the recall rate (recall) also reaches more than 70%. Therefore, it can be determined that the model created this time is effective and effective, and can be used in the identification of actual bathing events.