【python】高精度的多元多项式分体拟合

import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
from sklearn.linear_model import LinearRegression
from sklearn.preprocessing import PolynomialFeatures
from sklearn.metrics import mean_squared_error

data=pd.read_csv('')
cells=data.columns.tolist()[3:]

Preg=PolynomialFeatures(degree=5,include_bias=False)

#### 三维分体拟合，可以提高拟合精度
sigma_bins=[-3,-1,1,3]
temp_bins=[-50,0,65,130]
voltage_bins=[0,0.7,1,1.5]

for cell_index,cell in enumerate(cells):
	df=data[['sigma','temp','voltage',cell]].dropna(axis=0)
	df=df[df[cell]!='failed']  ## 对 dataframe稍微处理下去除dirty data

	x_inputs=df[['sigma','temp','voltage']]
	max_error_per_cell=0
	max_error_segment_per_cell=""
	max_mse_per_cell=0
	max_mse_segment_per_cell=""

	segment_index=0
	for sigma_index,sigma_bin in enumerate(['sbin1','sbin2','sbin3']):
		filter_sigma_df=df[pd.cut(df['sigma'],sigma_bins,labels=['sbin1','sbin2','sbin3'],include_lowest=True)==sigma_bin] ## 根据预先定义的分段条件对第一维度分段
		for temp_index,temp_bin in enumerate(['tbin1','tbin2','tbin3']):
			filter_temp_df=filter_sigma_df[pd.cut(filter_sigma_df['temp'],temp_bins,labels=['tbin1','tbin2','tbin3'],include_lowest=True)==temp_bin] ## 根据预先定义的分段条件对第二维度分段
			for voltage_index,voltage in enumerate(['vbin1','vbin2','vbin3']):
				filter_voltage_df=filter_temp_df[pd.cut(filter_temp_df['voltage'],voltage_bins,labels=['vbin1','vbin2','vbin3'],include_lowest=True)==voltage_bin] ## 根据预先定义的分段条件对第三维度分段
				filter_df=filter_voltage_df
				segment_index+=1
				if filter_df.empty is True:
					continue
				x=filter_df[['sigma','temp','voltage']]  ## 三个特征维度的输入
				y=filter_df[cell].astype(float)*1e10  ## 待拟合的值，值如果太小，稍微放大下乘个常数
				xx=Preg.fit_transform(np.array(x))  ## 将sigma temp voltage三个维度的输入按照PolynomialFeature定义的degree构造出多项式
				xx_names=Preg.fit(np.array(x)).get_feature_names(['sigma','temp','voltage'])  ## 将sigma temp voltage三个维度的输入按照PolynomialFeature定义的degree构造出多项式名称
				regr=LinearRegression()
				regr.fit(xx,y)  ## 进行拟合
				weights=regr.coef_  ## 有需求可以记录下权重
				bias=regr.intercept_  ## 有需求可以记录下截距
				predict=regr.predict(xx)
				max_error=((y-predict)/y)*100.abs().max()
				mse=mean_squared_error(y,predict)
				rmse=np.sqrt(mse)
				print("====================")
				print("degree= ",index)
				print("coefficients= ",regr.coef_)
				print("intercept= " ,regr.interept_)
				print("R^2=  ",regr.score(xx,y))
				print("====================")

				if max_error>max_error_per_cell:
					max_error_per_cell=max_error   ## 记录分体拟合后的最差segment的最大%误差
					max_error_segment_per_cell=segment_index ## 记录分体拟合后的最差segment的index，方便debug
				if mse>max_mse_per_cell:
					max_mse_per_cell=mse
					max_mse_segment_per_cell=segment_index

	print("### max error:",max_error_per_cell,"at segment:",max_error_segment_per_cell)
	print("### max MSE:",max_mse_per_cell,"at segment",max_mse_segment_per_cell)

实现三个特征的多项式分体（一维分段，二维分面，三位分体）拟合，最高阶为5，精度极高

参考：
https://blog.csdn.net/zhaohongfei_358/article/details/133901980
https://blog.csdn.net/l645317186/article/details/124963010
https://www.cnblogs.com/qiu-hua/p/14965659.html