如何将字符串数组转换为浮点数数组以适配机器学习模型？

我正在处理热传感器数据，该数据以字符串数组形式存储，前三组数据示例如下：

[25.64177131652832, 25.456825256347656, 24.92892074584961, 24.316755294799805, 23.454368591308594, 22.40850067138672, 21.829116821289062, 21.764848709106445, 21.72584342956543, 24.04592514038086, 24.358749389648438, 24.438608169555664, 24.78814697265625, 24.301586151123047, 24.657527923583984, 24.677270889282227, 24.422500610351562, 24.059892654418945, 23.85132598876953]

[23.502470016479492, 23.628114700317383, 23.68063735961914, 23.141122817993164, 22.970962524414062, 22.815067291259766, 22.713998794555664, 22.732934951782227, 22.681840896606445, 23.653491973876953, 23.332386016845703, 23.260208129882812, 23.49359703063965, 23.29375648498535, 23.734455108642578, 23.88347053527832, 23.9250431060791, 23.524120330810547, 23.196813583374023]

[23.735689163208008, 23.837121963500977, 24.044170379638672, 23.415151596069336, 21.520160675048828, 20.2335205078125, 19.441017150878906, 19.177982330322266, 18.871313095092773, 23.82925796508789, 24.192703247070312, 23.958843231201172, 23.816261291503906, 23.37897300720215, 23.469127655029297, 23.600635528564453, 23.589786529541016, 23.144105911254883, 22.776491165161133]

我需要将这类数据转换为float类型以用于机器学习模型，尝试了以下代码：

df["thermal2"] = pd.to_numeric(df["thermal"],errors="coerce")

结果全部为NaN：

0    NaN
1    NaN
2    NaN
     ..
83   NaN
84   NaN

Name: thermal2, Length: 85, dtype: float64

请问该如何解决？

解决方案 #

问题原因 #

pd.to_numeric只能处理单个数值的字符串，无法直接解析包含多个数值的数组格式字符串，所以直接转换会返回NaN。

解决步骤 #

1. 解析数组字符串：用ast.literal_eval把字符串格式的数组转换成Python列表
2. 生成float特征：将列表展开为单独的float列（适合传统机器学习模型），或者保留数组格式（适合序列输入的模型）

代码示例 #

import ast
import pandas as pd

# 解析字符串数组为Python列表
df["thermal_parsed"] = df["thermal"].apply(ast.literal_eval)

# 将列表展开为多个float特征列
thermal_features = pd.DataFrame(df["thermal_parsed"].tolist(), index=df.index)
# 给特征列命名，比如temp_0、temp_1...
thermal_features.columns = [f"temp_{i}" for i in range(thermal_features.shape[1])]

# 合并回原DataFrame
df = pd.concat([df, thermal_features], axis=1)

验证结果 #

此时thermal_features中的每一列都是float类型，可以直接用于机器学习模型训练。

内容的提问来源于stack exchange，提问作者Delowar Hossain