我是靠谱客的博主 开放音响,最近开发中收集的这篇文章主要介绍python实现xlsx文件分析详解,觉得挺不错的,现在分享给大家,希望可以做个参考。

概述

python脚本实现xlsx文件解析,供大家参考,具体内容如下

环境配置:

1.系统环境:Windows 7 64bit
2.编译环境:Python3.4.3
3.依赖库: os sys xlrd re
4.其他工具:none
5.前置条件:待处理的xlsx文件

脚本由来

最近的工作是做测试,而有一项任务呢,就是分析每天机器人巡检时采集的数据,包括各种传感器,CO2、O2、噪声等等,每天的数据也有上千条,通过站控的导出数据功能,会把数据库里面导出成xlsx文件,而这项任务要分析一下当天采集的数据是否在正常范围,要计算摄像头的识别率和识别准确率,自己傻呵呵的每天都在手动操作,突然觉得很浪费时间,索性写个python脚本吧,这样每天一条命令,就能得到自己想看的数据结果。每天至少节省10分钟!
这是要解析的xlsx文件: 

 

一般手动就得筛选、排序、打开计算器计算 - - 繁琐枯燥乏味
还是python大法好

代码浅析

流程图

脚本demo

#-*- coding:utf-8 -*-
import xlrd
import os
import sys
import logging
import re
#logging.basicConfig(level=logging.DEBUG)

xfile = sys.argv[1]

dateList = []
InspectionType = []
InspectionRresult = []

def load_data():

  CO2Type = []
  O2Type = []
  NoiseType = []
  SupwareType = []
  TowareType = []
  TemperatureType = []
  HumidityType = []
  InfraredType = []

  CO2Result = []
  O2Result = []
  NoiseResult = []
  SupwareResult = []
  TowareResult = []
  TemperatureResult = []
  HumidityResult = []
  InfraredResult = []

  logging.debug(InspectionType)
  logging.debug(InspectionRresult)


  for index, value in enumerate(InspectionType):
    if value == "二氧化碳":                   #CO2Type
      CO2Type.extend(value)
      logging.debug(index)
      logging.debug("CO2 RESULT:  "+InspectionRresult[index])
      CO2Result.append(InspectionRresult[index])

    if value == "氧气传感器":                  #O2Type
      O2Type.extend(value)
      O2Result.append(InspectionRresult[index])

    if value == "噪声传感器":                  #NoiseType
      NoiseType.extend(value)
      NoiseResult.append(InspectionRresult[index])


    if value == "局放(超声波测量)":               #SupwareType
      SupwareType.extend(value)
      SupwareResult.append(InspectionRresult[index])

    if value == "局放(地电波测量)":               #SupwareType
      TowareType.extend(value)
      TowareResult.append(InspectionRresult[index])

    if value == "温度传感器":                  #TemperatureType
      TemperatureType.extend(value)
      TemperatureResult.append(InspectionRresult[index])      

    if value == "湿度传感器":                  #TemperatureType
      HumidityType.extend(value)
      HumidityResult.append(InspectionRresult[index])

    if value == "温度(红外测量)":                  #TemperatureType
      InfraredType.extend(value)
      InfraredResult.append(InspectionRresult[index])      
  logging.debug(CO2Result)
  logging.debug(O2Result)
  logging.debug(NoiseResult)
  logging.debug(SupwareResult)
  logging.debug(TowareResult)
  logging.debug(TemperatureResult)
  logging.debug(HumidityResult)    
  logging.debug(InfraredResult)   
  return CO2Result,O2Result,NoiseResult,SupwareResult,TowareResult,TemperatureResult,HumidityResult,InfraredResult

def get_data_print(co2,o2,noise,supware,toware,temperature,humidity,infrared):
  co2 = list(map(eval,co2))
  o2 = list(map(eval,o2))
  noise = list(map(eval,noise))
  supware = list(map(eval,supware))
  toware = list(map(eval,toware))
  temperature = list(map(eval,temperature))
  humidity = list(map(eval,humidity))
  infrared = list(map(eval,infrared))

  co2Min = min(co2)
  co2Max = max(co2)
  logging.debug("CO2 min value :~~"+str(co2Min))
  logging.debug("CO2 max value :~~"+str(co2Max))

  o2Min = min(o2)
  o2Max = max(o2)
  noiseMin = min(noise)
  noiseMax = max(noise)

  supwareMin = min(supware)
  supwareMax = max(supware)

  towareMin = min(toware)
  towareMax = max(toware)

  temperatureMin = min(temperature)
  temperatureMax = max(temperature)

  humidityMin = min(humidity)
  humidityMax = max(humidity)

  infraredMin = min(infrared)
  infraredMax = max(infrared)

  print("CO2 values :",co2Min,'~~~~~~~',co2Max)
  print("o2 values :",o2Min,'~~~~~~~',o2Max)
  print("noise values :",noiseMin,'~~~~~~~',noiseMax)
  print("supware values :",supwareMin,'~~~~~~~',supwareMax)
  print("toware values :",towareMin,'~~~~~~~',towareMax)
  print("temperature values :",temperatureMin,'~~~~~~~',temperatureMax)
  print("humidity values :",humidityMin,'~~~~~~~',humidityMax)
  print("infrared values :",infraredMin,'~~~~~~~',infraredMax)

def cal_picture():
  result7to19List = []
  result19to7List = []
  count7to19List = []
  count19to7List = []
  count7to19Dict = {}
  count19to7Dict = {}

  failfind7to19cnt = 0
  failfind19to7cnt = 0
  photoType = []
  photoDateList = []
  allPhotoResult = []

  for index,value in enumerate(InspectionType):            #按照巡检类型筛选出视觉类,通过索引值同步时间、巡检结果
    if value == "开关(视觉识别)" or value == "旋钮(视觉识别)" or \
      value == "电流表(视觉识别)" or value == "电压表(视觉识别)":
      photoType.extend(value)
      photoDateList.append(dateList[index])
      allPhotoResult.append(InspectionRresult[index])
  for index,value in enumerate(photoDateList):
    if value[-8:] > '07:00:00' and value[-8:] < '19:00:00':
      result7to19List.append(allPhotoResult[index])
    if value[-8:] > '19:00:00' or value[-8:] < '7:00:00':
      result19to7List.append(allPhotoResult[index])

  logging.debug(result7to19List[-20:])
  logging.debug(result19to7List[:20])

  noduplicate7to19Set=set(result7to19List)              #里面无重复项
  for item in noduplicate7to19Set:
    count7to19List.append(result7to19List.count(item))
  logging.debug(count7to19List)
  count7to19Dict= dict(zip(list(noduplicate7to19Set),count7to19List))

  noduplicate19to7Set=set(result19to7List)              
  for item in noduplicate19to7Set:
    count19to7List.append(result19to7List.count(item))
  count19to7Dict= dict(zip(list(noduplicate19to7Set),count19to7List))

  logging.debug(count7to19Dict)

  None7to19cnt = count7to19Dict['']
  all7to19cnt = len(result7to19List)
  None19to7cnt = count19to7Dict['']
  all19to7cnt = len(result19to7List)

  logging.debug(None7to19cnt)

  for key in count7to19Dict:
    if count7to19Dict[key] == 1 :
      failfind7to19cnt = failfind7to19cnt+1
    if re.match('识别失败:*',key):
      failfind7to19cnt = failfind7to19cnt+ count7to19Dict[key]

  for key in count19to7Dict:
    if count19to7Dict[key] == 1 :
      failfind19to7cnt = failfind19to7cnt+1 
    if re.match('识别失败:*',key):
      failfind19to7cnt = failfind19to7cnt+count19to7Dict[key]
  logging.debug(all19to7cnt)

  print("7:00 ~~~ 19:00 识别率:",(all7to19cnt-None7to19cnt)/all7to19cnt)
  print("7:00 ~~~ 19:00 识别准确率:",(all7to19cnt-None7to19cnt-failfind7to19cnt)/(all7to19cnt-None7to19cnt))
  print("19:00 ~~~ 7:00 识别率:",(all19to7cnt-None19to7cnt)/all19to7cnt)
  print("19:00 ~~~ 7:00 识别准确率:",(all19to7cnt-None19to7cnt-failfind19to7cnt)/(all19to7cnt-None19to7cnt))
#读取xlsx文件
xlsxdata=xlrd.open_workbook(xfile)
tablepage=xlsxdata.sheets()[0]
dateList.extend(tablepage.col_values(5))
InspectionType.extend(tablepage.col_values(3))
InspectionRresult.extend(tablepage.col_values(6))

cal_picture()
co2,o2,noise,supware,toware,temperature,humidity,infrared=load_data()
get_data_print(co2,o2,noise,supware,toware,temperature,humidity,infrared)

结果图

回顾与总结

渐渐体会到python脚本的优势所在。
python在代码保密上可能是解释性语言共有的小小缺陷,做项目还是C/C++,当然是指传统项目
写python很开心啊

以上就是本文的全部内容,希望对大家的学习有所帮助,也希望大家多多支持脚本之家。

最后

以上就是开放音响为你收集整理的python实现xlsx文件分析详解的全部内容,希望文章能够帮你解决python实现xlsx文件分析详解所遇到的程序开发问题。

如果觉得靠谱客网站的内容还不错,欢迎将靠谱客网站推荐给程序员好友。

本图文内容来源于网友提供,作为学习参考使用,或来自网络收集整理,版权属于原作者所有。
点赞(130)

评论列表共有 0 条评论

立即
投稿
返回
顶部