如何用自定义最小趋势分解信号?替换seasonal_decompose默认均值趋势
用最小值作为趋势项的信号分解方案
seasonal_decompose的趋势项默认通过移动平均计算,无法直接指定为最小值逻辑。要实现以最小值为趋势项、保证tempo恒正的需求,可通过以下两种方式处理:
方法一:替换分解后的趋势项
场景1:全局最小值作为趋势(固定下界)
如果信号的下界是固定值,直接用全局最小值生成趋势线,再重新推导季节项和残差:
import numpy as np from statsmodels.tsa.seasonal import seasonal_decompose myPeriod = 60 # 先执行常规分解(可选,主要是参考原季节项逻辑) decompose_result_mult = seasonal_decompose(mySignal, model="additive", period=myPeriod) # 生成全局最小值趋势线 global_min = np.nanmin(mySignal) trend = np.full_like(mySignal, global_min) # 计算去趋势序列 detrended = mySignal - trend # 提取季节项:按周期位置取均值 seasonal = np.zeros_like(mySignal) for i in range(myPeriod): seasonal[i::myPeriod] = np.nanmean(detrended[i::myPeriod]) # 计算残差 residual = detrended - seasonal # 生成恒正的tempo tempo = residual + seasonal # 等价于 mySignal - trend
场景2:滚动最小值作为趋势(动态下界)
如果信号的下界随时间变化,用滚动窗口计算最小值作为趋势线:
import numpy as np import pandas as pd myPeriod = 60 window_size = 2 * myPeriod # 窗口大小可根据信号特性调整 # 将numpy数组转为Series,方便计算滚动最小值 signal_series = pd.Series(mySignal) # 生成滚动最小值趋势线,min_periods=1保证开头数据有效 trend = signal_series.rolling(window=window_size, min_periods=1).min().values # 后续步骤同全局最小值场景 detrended = mySignal - trend seasonal = np.zeros_like(mySignal) for i in range(myPeriod): seasonal[i::myPeriod] = np.nanmean(detrended[i::myPeriod]) residual = detrended - seasonal tempo = mySignal - trend
方法二:完全自定义分解逻辑
直接跳过seasonal_decompose,按需求实现分解流程,核心逻辑是先确定趋势下界,再提取季节成分:
import numpy as np import pandas as pd myPeriod = 60 window_size = 2 * myPeriod # 1. 生成趋势项(这里用滚动最小值,可替换为全局最小值) signal_series = pd.Series(mySignal) trend = signal_series.rolling(window=window_size, min_periods=1).min().values # 2. 计算去趋势序列 detrended = mySignal - trend # 3. 提取季节项:对每个周期位置的去趋势值取均值 seasonal = np.zeros_like(mySignal) for idx in range(myPeriod): # 取所有周期中第idx位置的数值,计算均值(忽略NaN) period_vals = detrended[idx::myPeriod] seasonal[idx::myPeriod] = np.nanmean(period_vals) # 4. 计算残差 residual = detrended - seasonal # 5. 生成恒正的tempo tempo = residual + seasonal
关键说明
- 无论哪种方法,
tempo = mySignal - trend,只要trend始终小于等于对应位置的mySignal,tempo就会恒正。 - 滚动窗口的大小可根据信号的周期特性调整,建议设为周期的整数倍,避免季节成分干扰趋势计算。
内容的提问来源于stack exchange,提问作者chang thenoob
相关产品推荐
相关产品推荐

