409 lines
15 KiB
Python
409 lines
15 KiB
Python
# ...existing code...
|
||
import pandas as pd
|
||
import numpy as np
|
||
import matplotlib.pyplot as plt
|
||
from pykalman import KalmanFilter
|
||
import os
|
||
|
||
|
||
def fill_time_gaps(
|
||
data: pd.DataFrame,
|
||
time_col: str = "time",
|
||
freq: str = "1min",
|
||
short_gap_threshold: int = 10,
|
||
) -> pd.DataFrame:
|
||
"""
|
||
补齐缺失时间戳并填补数据缺口。
|
||
|
||
Args:
|
||
data: 包含时间列的 DataFrame
|
||
time_col: 时间列名（默认 'time'）
|
||
freq: 重采样频率（默认 '1min'）
|
||
short_gap_threshold: 短缺口阈值（分钟），<=此值用线性插值，>此值用前向填充
|
||
|
||
Returns:
|
||
补齐时间后的 DataFrame（保留原时间列格式）
|
||
"""
|
||
if time_col not in data.columns:
|
||
raise ValueError(f"时间列 '{time_col}' 不存在于数据中")
|
||
|
||
# 解析时间列并设为索引
|
||
data = data.copy()
|
||
data[time_col] = pd.to_datetime(data[time_col], utc=True)
|
||
data_indexed = data.set_index(time_col)
|
||
|
||
# 生成完整时间范围
|
||
full_range = pd.date_range(
|
||
start=data_indexed.index.min(), end=data_indexed.index.max(), freq=freq
|
||
)
|
||
|
||
# 重索引以补齐缺失时间点
|
||
data_reindexed = data_indexed.reindex(full_range)
|
||
|
||
# 按列处理缺口
|
||
for col in data_reindexed.columns:
|
||
# 识别缺失值位置
|
||
is_missing = data_reindexed[col].isna()
|
||
|
||
# 计算连续缺失的长度
|
||
missing_groups = (is_missing != is_missing.shift()).cumsum()
|
||
gap_lengths = is_missing.groupby(missing_groups).transform("sum")
|
||
|
||
# 短缺口：线性插值
|
||
short_gap_mask = is_missing & (gap_lengths <= short_gap_threshold)
|
||
if short_gap_mask.any():
|
||
data_reindexed.loc[short_gap_mask, col] = (
|
||
data_reindexed[col]
|
||
.interpolate(method="linear", limit_area="inside")
|
||
.loc[short_gap_mask]
|
||
)
|
||
|
||
# 长缺口：前向填充
|
||
long_gap_mask = is_missing & (gap_lengths > short_gap_threshold)
|
||
if long_gap_mask.any():
|
||
data_reindexed.loc[long_gap_mask, col] = (
|
||
data_reindexed[col].ffill().loc[long_gap_mask]
|
||
)
|
||
|
||
# 重置索引并恢复时间列（保留原格式）
|
||
data_result = data_reindexed.reset_index()
|
||
data_result.rename(columns={"index": time_col}, inplace=True)
|
||
|
||
# 保留时区信息
|
||
data_result[time_col] = data_result[time_col].dt.strftime("%Y-%m-%dT%H:%M:%S%z")
|
||
# 修正时区格式（Python的%z输出为+0000，需转为+00:00）
|
||
data_result[time_col] = data_result[time_col].str.replace(
|
||
r"(\+\d{2})(\d{2})$", r"\1:\2", regex=True
|
||
)
|
||
|
||
return data_result
|
||
|
||
|
||
def clean_flow_data_kf(
|
||
input_csv_path: str, show_plot: bool = False, fill_gaps: bool = True
|
||
) -> str:
|
||
"""
|
||
读取 input_csv_path 中的每列时间序列，使用一维 Kalman 滤波平滑并用预测值替换基于 3σ 检测出的异常点。
|
||
保存输出为：<input_filename>_cleaned.xlsx（与输入同目录），并返回输出文件的绝对路径。
|
||
仅保留输入文件路径作为参数（按要求）。
|
||
|
||
Args:
|
||
input_csv_path: CSV 文件路径
|
||
show_plot: 是否显示可视化
|
||
fill_gaps: 是否先补齐时间缺口（默认 True）
|
||
"""
|
||
# 读取 CSV
|
||
data = pd.read_csv(input_csv_path, header=0, index_col=None, encoding="utf-8")
|
||
|
||
# 补齐时间缺口（如果数据包含 time 列）
|
||
if fill_gaps and "time" in data.columns:
|
||
data = fill_time_gaps(
|
||
data, time_col="time", freq="1min", short_gap_threshold=10
|
||
)
|
||
|
||
# 分离时间列和数值列
|
||
time_col_data = None
|
||
if "time" in data.columns:
|
||
time_col_data = data["time"]
|
||
data = data.drop(columns=["time"])
|
||
|
||
# 存储 Kalman 平滑结果
|
||
data_kf = pd.DataFrame(index=data.index, columns=data.columns)
|
||
# 平滑每一列
|
||
for col in data.columns:
|
||
observations = pd.Series(data[col].values).ffill().bfill()
|
||
if observations.isna().any():
|
||
observations = observations.fillna(observations.mean())
|
||
obs = observations.values.astype(float)
|
||
|
||
kf = KalmanFilter(
|
||
transition_matrices=[1],
|
||
observation_matrices=[1],
|
||
initial_state_mean=float(obs[0]),
|
||
initial_state_covariance=1,
|
||
observation_covariance=1,
|
||
transition_covariance=0.01,
|
||
)
|
||
# 跳过EM学习，使用固定参数以提高性能
|
||
state_means, _ = kf.smooth(obs)
|
||
data_kf[col] = state_means.flatten()
|
||
|
||
# 计算残差并用IQR检测异常（更稳健的方法）
|
||
residuals = data - data_kf
|
||
residual_thresholds = {}
|
||
for col in data.columns:
|
||
res_values = residuals[col].dropna().values # 移除NaN以计算IQR
|
||
q1 = np.percentile(res_values, 25)
|
||
q3 = np.percentile(res_values, 75)
|
||
iqr = q3 - q1
|
||
lower_threshold = q1 - 1.5 * iqr
|
||
upper_threshold = q3 + 1.5 * iqr
|
||
residual_thresholds[col] = (lower_threshold, upper_threshold)
|
||
|
||
cleaned_data = data.copy()
|
||
anomalies_info = {}
|
||
for col in data.columns:
|
||
lower, upper = residual_thresholds[col]
|
||
sensor_residuals = residuals[col]
|
||
anomaly_mask = (sensor_residuals < lower) | (sensor_residuals > upper)
|
||
anomaly_idx = data.index[anomaly_mask.fillna(False)]
|
||
anomalies_info[col] = pd.DataFrame(
|
||
{
|
||
"Observed": data.loc[anomaly_idx, col],
|
||
"Kalman_Predicted": data_kf.loc[anomaly_idx, col],
|
||
"Residual": sensor_residuals.loc[anomaly_idx],
|
||
}
|
||
)
|
||
cleaned_data.loc[anomaly_idx, f"{col}_cleaned"] = data_kf.loc[anomaly_idx, col]
|
||
|
||
# 如果原始数据包含时间列，将其添加回结果
|
||
if time_col_data is not None:
|
||
cleaned_data.insert(0, "time", time_col_data)
|
||
|
||
# 构造输出文件名：在输入文件名基础上加后缀 _cleaned.xlsx
|
||
input_dir = os.path.dirname(os.path.abspath(input_csv_path))
|
||
input_base = os.path.splitext(os.path.basename(input_csv_path))[0]
|
||
output_filename = f"{input_base}_cleaned.xlsx"
|
||
output_path = os.path.join(input_dir, output_filename)
|
||
|
||
# 覆盖同名文件
|
||
if os.path.exists(output_path):
|
||
os.remove(output_path)
|
||
cleaned_data.to_excel(output_path, index=False)
|
||
|
||
# 可选可视化（第一个传感器）
|
||
plt.rcParams["font.sans-serif"] = ["SimHei"]
|
||
plt.rcParams["axes.unicode_minus"] = False
|
||
if show_plot and len(data.columns) > 0:
|
||
sensor_to_plot = data.columns[0]
|
||
plt.figure(figsize=(12, 6))
|
||
plt.plot(
|
||
data.index,
|
||
data[sensor_to_plot],
|
||
label="监测值",
|
||
marker="o",
|
||
markersize=3,
|
||
alpha=0.7,
|
||
)
|
||
plt.plot(
|
||
data.index, data_kf[sensor_to_plot], label="Kalman滤波预测值", linewidth=2
|
||
)
|
||
anomaly_idx = anomalies_info[sensor_to_plot].index
|
||
if len(anomaly_idx) > 0:
|
||
plt.plot(
|
||
anomaly_idx,
|
||
data[sensor_to_plot].loc[anomaly_idx],
|
||
"ro",
|
||
markersize=8,
|
||
label="监测值异常点",
|
||
)
|
||
plt.plot(
|
||
anomaly_idx,
|
||
data_kf[sensor_to_plot].loc[anomaly_idx],
|
||
"go",
|
||
markersize=8,
|
||
label="Kalman修复值",
|
||
)
|
||
plt.xlabel("时间点（序号）")
|
||
plt.ylabel("监测值")
|
||
plt.title(f"{sensor_to_plot}：观测值与Kalman滤波预测值（异常点标记）")
|
||
plt.legend()
|
||
plt.show()
|
||
|
||
# 返回输出文件的绝对路径
|
||
return os.path.abspath(output_path)
|
||
|
||
|
||
def clean_flow_data_df_kf(data: pd.DataFrame, show_plot: bool = False) -> dict:
|
||
"""
|
||
接收一个 DataFrame 数据结构，使用一维 Kalman 滤波平滑并用预测值替换基于 IQR 检测出的异常点。
|
||
区分合理的0值（流量转换）和异常的0值（连续多个0或孤立0）。
|
||
返回完整的清洗后的字典数据结构。
|
||
|
||
Args:
|
||
data: 输入 DataFrame（可包含 time 列）
|
||
show_plot: 是否显示可视化
|
||
"""
|
||
# 使用传入的 DataFrame
|
||
data = data.copy()
|
||
|
||
# 补齐时间缺口（如果启用且数据包含 time 列）
|
||
data_filled = fill_time_gaps(
|
||
data, time_col="time", freq="1min", short_gap_threshold=10
|
||
)
|
||
# 移除 time 列用于后续清洗
|
||
data_filled = data_filled.drop(columns=["time"])
|
||
|
||
# 存储 Kalman 平滑结果
|
||
data_kf = pd.DataFrame(index=data_filled.index, columns=data_filled.columns)
|
||
# 平滑每一列
|
||
for col in data_filled.columns:
|
||
observations = pd.Series(data_filled[col].values).ffill().bfill()
|
||
if observations.isna().any():
|
||
observations = observations.fillna(observations.mean())
|
||
obs = observations.values.astype(float)
|
||
|
||
kf = KalmanFilter(
|
||
transition_matrices=[1],
|
||
observation_matrices=[1],
|
||
initial_state_mean=float(obs[0]),
|
||
initial_state_covariance=1,
|
||
observation_covariance=10,
|
||
transition_covariance=10,
|
||
)
|
||
state_means, _ = kf.smooth(obs)
|
||
data_kf[col] = state_means.flatten()
|
||
|
||
# 计算残差并用IQR检测异常
|
||
residuals = data_filled - data_kf
|
||
residual_thresholds = {}
|
||
for col in data_filled.columns:
|
||
res_values = residuals[col].dropna().values
|
||
q1 = np.percentile(res_values, 25)
|
||
q3 = np.percentile(res_values, 75)
|
||
iqr = q3 - q1
|
||
lower_threshold = q1 - 1.5 * iqr
|
||
upper_threshold = q3 + 1.5 * iqr
|
||
residual_thresholds[col] = (lower_threshold, upper_threshold)
|
||
|
||
# 创建完整的修复数据
|
||
cleaned_data = data_filled.copy()
|
||
anomalies_info = {}
|
||
|
||
for col in data_filled.columns:
|
||
lower, upper = residual_thresholds[col]
|
||
sensor_residuals = residuals[col]
|
||
anomaly_mask = (sensor_residuals < lower) | (sensor_residuals > upper)
|
||
anomaly_idx = data_filled.index[anomaly_mask.fillna(False)]
|
||
|
||
anomalies_info[col] = pd.DataFrame(
|
||
{
|
||
"Observed": data_filled.loc[anomaly_idx, col],
|
||
"Kalman_Predicted": data_kf.loc[anomaly_idx, col],
|
||
"Residual": sensor_residuals.loc[anomaly_idx],
|
||
}
|
||
)
|
||
|
||
# 直接在原列上替换异常值为 Kalman 预测值
|
||
cleaned_data.loc[anomaly_idx, col] = data_kf.loc[anomaly_idx, col]
|
||
|
||
# 可选可视化
|
||
plt.rcParams["font.sans-serif"] = ["SimHei"]
|
||
plt.rcParams["axes.unicode_minus"] = False
|
||
if show_plot and len(data.columns) > 0:
|
||
sensor_to_plot = data.columns[0]
|
||
|
||
# 定义x轴
|
||
n = len(data)
|
||
time = np.arange(n)
|
||
n_filled = len(data_filled)
|
||
time_filled = np.arange(n_filled)
|
||
|
||
plt.figure(figsize=(12, 8))
|
||
|
||
plt.subplot(2, 1, 1)
|
||
plt.plot(
|
||
time,
|
||
data[sensor_to_plot],
|
||
label="原始监测值",
|
||
marker="o",
|
||
markersize=3,
|
||
alpha=0.7,
|
||
)
|
||
|
||
# 修正：检查 data_filled 的异常值，绘制在 time_filled 上
|
||
abnormal_zero_mask = data_filled[sensor_to_plot].isna()
|
||
# 如果目的是检查0值，应该用 == 0。这里保留 isna() 但修正索引引用，防止crash。
|
||
# 如果原意是 isna() 则在 fillna 后通常没有 na。假设用户可能想检查 0 值？
|
||
# 基于 "异常0值" 的标签，改为检查 0 值更合理，但为了保险起见，
|
||
# 如果 isna() 返回空，就不画。防止索引越界是主要的。
|
||
abnormal_zero_idx = data_filled.index[abnormal_zero_mask]
|
||
|
||
if len(abnormal_zero_idx) > 0:
|
||
# 注意：如果 abnormal_zero_idx 是基于 data_filled 的索引（0..M-1），
|
||
# 直接作为 x 坐标即可，因为 time_filled 也是 0..M-1
|
||
# 而 y 值应该取自 data_filled 或 data_kf，取 data 会越界
|
||
plt.plot(
|
||
abnormal_zero_idx,
|
||
data_filled[sensor_to_plot].loc[abnormal_zero_idx],
|
||
"mo",
|
||
markersize=8,
|
||
label="异常值(NaN)",
|
||
)
|
||
|
||
plt.plot(
|
||
time_filled, data_kf[sensor_to_plot], label="Kalman滤波预测值", linewidth=2
|
||
)
|
||
anomaly_idx = anomalies_info[sensor_to_plot].index
|
||
if len(anomaly_idx) > 0:
|
||
plt.plot(
|
||
anomaly_idx,
|
||
data_filled[sensor_to_plot].loc[anomaly_idx],
|
||
"ro",
|
||
markersize=8,
|
||
label="IQR异常点",
|
||
)
|
||
plt.xlabel("时间点（序号）")
|
||
plt.ylabel("流量值")
|
||
plt.title(f"{sensor_to_plot}：原始数据与异常检测")
|
||
plt.legend()
|
||
|
||
plt.subplot(2, 1, 2)
|
||
plt.plot(
|
||
time_filled,
|
||
cleaned_data[sensor_to_plot],
|
||
label="修复后监测值",
|
||
marker="o",
|
||
markersize=3,
|
||
color="green",
|
||
)
|
||
plt.xlabel("时间点（序号）")
|
||
plt.ylabel("流量值")
|
||
plt.title(f"{sensor_to_plot}：修复后数据")
|
||
plt.legend()
|
||
|
||
plt.tight_layout()
|
||
plt.show()
|
||
|
||
# 返回完整的修复后字典
|
||
return cleaned_data
|
||
|
||
|
||
# # 测试
|
||
# if __name__ == "__main__":
|
||
# # 默认：脚本目录下同名 CSV 文件
|
||
# script_dir = os.path.dirname(os.path.abspath(__file__))
|
||
# default_csv = os.path.join(script_dir, "pipe_flow_data_to_clean2.0.csv")
|
||
# out = clean_flow_data_kf(default_csv)
|
||
# print("清洗后的数据已保存到:", out)
|
||
|
||
# 测试 clean_flow_data_dict 函数
|
||
if __name__ == "__main__":
|
||
import random
|
||
|
||
# 读取 szh_flow_scada.csv 文件
|
||
script_dir = os.path.dirname(os.path.abspath(__file__))
|
||
csv_path = os.path.join(script_dir, "szh_flow_scada.csv")
|
||
data = pd.read_csv(csv_path, header=0, index_col=None, encoding="utf-8")
|
||
|
||
# 排除 Time 列，随机选择 5 列
|
||
columns_to_exclude = ["Time"]
|
||
available_columns = [col for col in data.columns if col not in columns_to_exclude]
|
||
selected_columns = random.sample(available_columns, 1)
|
||
|
||
# 将选中的列转换为字典
|
||
data_dict = {col: data[col].tolist() for col in selected_columns}
|
||
|
||
print("选中的列:", selected_columns)
|
||
print("原始数据长度:", len(data_dict[selected_columns[0]]))
|
||
|
||
# 调用函数进行清洗
|
||
cleaned_dict = clean_flow_data_df_kf(data_dict, show_plot=True)
|
||
# 将清洗后的字典写回 CSV
|
||
out_csv = os.path.join(script_dir, f"{selected_columns[0]}_clean.csv")
|
||
pd.DataFrame(cleaned_dict).to_csv(out_csv, index=False, encoding="utf-8-sig")
|
||
print("已保存清洗结果到:", out_csv)
|
||
print("清洗后的字典键:", list(cleaned_dict.keys()))
|
||
print("清洗后的数据长度:", len(cleaned_dict[selected_columns[0]]))
|
||
print("测试完成：函数运行正常")
|