python爬虫爬取国家科技报告服务系统数据,共计30余万条
按学科分类【中图分类】
共计三十余万条科技报告数据
爬取的网址:https://www.nstrs.cn/kjbg/navigation
!!!
如果要完整地跑起来代码,需要先看一下我的这篇博客,完成IP代理池的相关配置:
https://www.cnblogs.com/rainbow-1/p/16725503.html
!!!
分析网站数据来源可以发现,是使用的post方式的请求,且参数列表如下:
那么我们需要做的就是模拟这个请求,同时需要带上我们自定义的参数,这里面需要的其实一个就是页码pageNo,另一个是分类,如下图:
parms = { "pageNo": i, "competentOrg": "", "jihuaId": "", "fieldCode": "", "classification": name, # 修改 "kjbgRegion": "", "kjbgType": "", "grade": "" }
简单说一下我都做了什么,首先是配置是IP代理池,存在redis数据库,每次【设置了随机延迟时间】随机取出一个进行访问。
其次使用了随机UserAgent请求头。
爬虫是直接使用post请求,携带参数抓获返回的json数据做解析并存入mysql数据库。
下面是代码:
爬虫方法report_crawler
也就是你需要直接运行的方法。
我这部分是从”社会科学总论”这个分类开始爬的,前面那些如果需要爬,就直接改pageList页码列表、nameList名称列表、tableList数据库表列表就可以【切记是一 一对应的!】
import jsonimport randomfrom time import sleepimport requestsfrom fake_useragent import UserAgentfrom report_data.into_mysql import insert_mysqlfrom report_data.ip_redis import my_redis"""post方法参数params:字典或字节序列,作为参数增加到链接中data:字典,字节序列或文件对象,作为请求的内容json:JSON格式的数据,作为Request的内容headers:字典,HTTP定制头(模拟浏览器进行访问)cookies:字典或CpplieJar,Request中的cookieauth:元祖,支持HTTP认证功能files:字典类型,传输文件timeout:设定超时时间,秒为单位proxies:字典类型,设定访问代理服务器,可以增加登陆认证allow_redirects:True//False,默认为True,重定向开关stream:True/False,默认为True,获取内容立即下载开关verify:True/False,默认为True,认证SSL证书开关cert:本地SSL证书路径"""# 页码pageList# 分类名称参数列表 nameList#def get_report(page,name,tableName): # ------------------------------ 修改页码 for i in range(1,page): print("---------------------------------") ua = UserAgent() print("【随机 UserAgent:】" + ua.random) # 随机产生headers temp_headers = ua.random # -------------------------------------- test_redis = my_redis() temp_proxy = test_redis.get_ip() print("【随机 IP:】" + temp_proxy) url="https://www.nstrs.cn/rest/kjbg/wfKjbg/list" # url2 = "https://www.nstrs.cn/rest/kjbg/wfKjbg/list?pageNo=2&competentOrg=&jihuaId=&fieldCode=&classification=医药、卫生&kjbgRegion=&kjbgType=&grade=" parms = { "pageNo": i, "competentOrg": "", "jihuaId": "", "fieldCode": "", "classification": name, # 修改 "kjbgRegion": "", "kjbgType": "", "grade": "" } other_parms={ 'User-Agent': temp_headers, 'https': 'http://'+temp_proxy, 'http': 'http://'+temp_proxy } sleeptime = random.uniform(0, 0.7) sleep(sleeptime) # print(url) response = requests.post(url, parms, other_parms) response.encoding='utf8' print(response.text+'\n') response_data = response.text # 返回数据 json_data = json.loads(response_data) # 封装字典 res_list_data = json_data['RESULT']['list'] # 一页 长度为10的list [{ },{ },{ } ... { }] len=10 """ 重新构建一个 list [{ }] """ for item in res_list_data: insert_mysql(item,name,tableName) returnif __name__ == '__main__': # 页码 pageList [] pageList = [788,779,656,584,573,510,440,361, 315,226,224,220,155,112,112, 87,53,50,39,33,18,12,5,4,2,2,2,2] nameList = [ "社会科学总论", "环境科学、安全科学", "建筑科学", "轻工业、手工业", "数理科学与化学", "能源与动力工程", "电工技术", "矿业工程", "经济", "文化、科学、教育、体育", "水利工程", "交通运输", "自然科学总论", "石油、天然气工业", "冶金工业", "武器工业", "航空、航天", "哲学、宗教", "原子能技术", "历史、地理", "政治、法律", "艺术", "语言、文字", "军事", "综合性图书", "文学", "语言、文学", "mks主义、ln主义、mzd思想、dxp理论" ] tableList = ["tech_c","tech_x","tech_tu","tech_ts","tech_o","tech_tk","tech_tm", "tech_td","tech_f","tech_g","tech_tv","tech_u", "tech_n","tech_te","tech_tf","tech_tj","tech_v","tech_b","tech_tl", "tech_k","tech_d","tech_j","tech_h","tech_e","tech_z","tech_i","tech_i","tech_a"] for i in range(0,len(tableList)): get_report(pageList[i],nameList[i],tableList[i])
目录方法category
返回一个中图分类号对应的名称
# 用以返回中图分类号def get_code(key): code_dict = { "医药、卫生":"R", "一般工业技术":"TB", "生物科学":"Q", "数理科学和化学":"O", "农业科学":"S", "工业技术":"T", "自动化技术、计算机技术":"TP", "天文学、地球科学":"P", "无线电电子学、电信技术":"TN", "金属学与金属工艺":"TG", "机械、仪表工业":"TH", "化学工业":"TQ", "社会科学总论":"C", "环境科学、安全科学":"X", "建筑科学":"TU", "轻工业、手工业":"TS", "数理科学与化学":"O", "能源与动力工程":"TK", "电工技术":"TM", "矿业工程":"TD", "经济":"F", "文化、科学、教育、体育":"G", "水利工程":"TV", "交通运输":"U", "自然科学总论":"N", "石油、天然气工业":"TE", "冶金工业":"TF", "武器工业":"TJ", "航空、航天":"V", "哲学、宗教":"B", "原子能技术":"TL", "历史、地理":"K", "政治、法律":"D", "艺术":"J", "语言、文字":"H", "军事":"E", "综合性图书":"Z", "文学":"I", "语言、文学":"I", "mks主义、ln主义、mzd思想、dxp理论":"A", } res = code_dict.get(key) return resif __name__ == '__main__': data = get_code("工业技术") print(data)
user_agent方法
返回随机headers
from fake_useragent import UserAgent # 下载:pip install fake-useragentimport requestsua = UserAgent() # 实例化,需要联网但是网站不太稳定-可能耗时会长一些print(ua.random) # 随机产生headers = { 'User-Agent': ua.random # 伪装 }# 请求if __name__ == '__main__': url = 'https://www.baidu.com/' response = requests.get(url, headers=headers ,proxies={"http":"117.136.27.43"}) print(response.status_code)
ip_redis方法
从redis数据库取出一个ip并返回(前3000个随机一个,降序排列)
import randomimport redisclass my_redis: def get_ip(self): r = redis.Redis(host='127.0.0.1', port=6379, db=0,decode_responses=True) my_redis_data = r.zrange("proxies:universal",1,3000,True) return random.choice(my_redis_data) # print(len(my_redis_data))if __name__ == '__main__': test_redis=my_redis() data=test_redis.get_ip() print(data)
into_mysql方法
存入mysql数据库的方法
#连接数据库 获取游标import pymysqlfrom report_data.category import get_codedef get_conn(): """ :return: 连接,游标 """ # 创建连接 conn = pymysql.connect(host="127.0.0.1", user="root", password="reliable", db="tech", charset="utf8mb4") # 创建游标 cursor = conn.cursor() # 执行完毕返回的结果集默认以元组显示 if ((conn != None) & (cursor != None)): print("数据库连接成功 ...") else: print("数据库连接失败!") return conn, cursor#关闭数据库连接和游标def close_conn(conn, cursor): if cursor: cursor.close() if conn: conn.close() return 1# 数据表名# 中图分类名def insert_mysql(data,name,tableName): print(data['title']) id=data['id'] title=data['title'] alternativeTitle=data['alternativeTitle'] creator=data['creator'] abstractEn=data['abstractEn'] keywordsEn=data['keywordsEn'] abstractCn=data['abstractCn'] keywordsCn=data['keywordsCn'] creatOrorganization=data['creatOrorganization'] prepareOrganization=data['prepareOrganization'] publicDate=data['publicDate'] createTime=data['createTime'] projectName=data['projectName'] competentOrg=data['competentOrg'] projectSubjectName=data['projectSubjectName'] projectSubjectId=data['projectSubjectId'] #------------------------------ classification=name # 修改 #------------------------------ classificationCode=get_code(classification) # 需要调用get_code(name)获取 responsiblePerson = data['responsiblePerson'] supportChannel = data['supportChannel'] undertakeOrg = data['undertakeOrg'] kjbgSource = data['kjbgSource'] proposalDate = data['proposalDate'] submittedDate = data['submittedDate'] kjbgRegion = data['kjbgRegion'] collectionDate = data['collectionDate'] collectionNumber = data['collectionNumber'] fieldCode = data['fieldCode'] fieldId = data['fieldId'] kjbgQWAddress = data['kjbgQWAddress'] isNewRecord = data['isNewRecord'] sourceUrl = "https://www.nstrs.cn/kjbg/detail?id="+id # 需要自己拼 https://www.nstrs.cn/kjbg/detail?id= conn, cursor = get_conn() # ------------------------------ 修改表名 sql = "insert into `"+tableName+"` (id,title,alternativeTitle,creator,abstractEn," \ "keywordsEn,abstractCn,keywordsCn,creatOrorganization,prepareOrganization," \ "publicDate,createTime,projectName,competentOrg,projectSubjectName," \ "projectSubjectId,classification,classificationCode,responsiblePerson,supportChannel," \ "undertakeOrg,kjbgSource,proposalDate,submittedDate,kjbgRegion," \ "collectionDate,collectionNumber,fieldCode,fieldId,kjbgQWAddress," \ "isNewRecord,sourceUrl) values(%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s" \ ",%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s)" try: try: cursor.execute(sql, [id,title,alternativeTitle,creator,abstractEn, keywordsEn,abstractCn,keywordsCn,creatOrorganization,prepareOrganization, publicDate,createTime,projectName,competentOrg,projectSubjectName, projectSubjectId,classification,classificationCode,responsiblePerson,supportChannel, undertakeOrg,kjbgSource,proposalDate,submittedDate,kjbgRegion, collectionDate,collectionNumber,fieldCode,fieldId,kjbgQWAddress,isNewRecord,sourceUrl]) except pymysql.err.IntegrityError: print("主键冲突!") conn.commit() # 提交事务 update delete insert操作 except pymysql.err.IntegrityError: print("error!") finally: close_conn(conn, cursor) return 1if __name__ == '__main__': print()
最终爬取三十多万条科技报告,按中图分类建立了mysql数据表,分表存储不同分类的数据。
【其中的数理科学和化学,数理科学与化学这两个分类做了合并,合并为数理科学和化学类,属O】
【语言、文学和文学做了合并,同属 I 文学类】
附几张结果图:
最后说一下数据表结构:
/*Navicat MySQL Data TransferSource Server : reliableSource Server Version : 80013Source Host : localhost:3306Source Database : techTarget Server Type : MYSQLTarget Server Version : 80013File Encoding : 65001Date: 2022-09-24 13:54:05*/SET FOREIGN_KEY_CHECKS=0;-- ------------------------------ Table structure for tech_o-- ----------------------------DROP TABLE IF EXISTS `tech_o`;CREATE TABLE `tech_o` ( `id` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci NOT NULL COMMENT 'ID', `title` text CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci COMMENT '中文标题', `alternativeTitle` text CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci COMMENT '英文标题', `creator` text CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci COMMENT '作者', `abstractEn` text CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci COMMENT '英文摘要', `keywordsEn` text CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci COMMENT '英文关键字', `abstractCn` text CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci COMMENT '中文摘要', `keywordsCn` text CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci COMMENT '中文关键字', `creatOrorganization` text CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci COMMENT '创建者组织', `prepareOrganization` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '预备组织', `publicDate` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '公布时间', `createTime` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '编制时间', `projectName` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '项目名称', `competentOrg` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '项目地址', `projectSubjectName` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '项目主题名称', `projectSubjectId` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '项目主题ID', `classification` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '中图分类名称', `classificationCode` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '中图分类号', `responsiblePerson` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '负责人', `supportChannel` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '主办方', `undertakeOrg` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '承办方', `kjbgSource` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '科技报告来源单位', `proposalDate` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '提议时间', `submittedDate` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '提交时间', `kjbgRegion` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '科技报告所属行政区划', `collectionDate` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '收集时间', `collectionNumber` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '收集编号', `fieldCode` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '领域代码', `fieldId` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '领域ID', `kjbgQWAddress` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '报告链接', `isNewRecord` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '是否新记录', `sourceUrl` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_general_ci DEFAULT NULL COMMENT '国家科技报告服务系统收录链接', PRIMARY KEY (`id`)) ENGINE=InnoDB DEFAULT CHARSET=utf8;
如果需要获取这部分数据,可关注我的微信公众号【小杨的挨踢IT生活】,回复 “科技报告” 获取下载链接。
好看请赞,养成习惯:) 本文来自博客园,作者:靠谱杨, 转载请注明原文链接:https://www.cnblogs.com/rainbow-1/p/16725576.html
关于笔者: 我的主页
文章同步51CTO,可以帮忙踩一踩 ~ 我的51CTO博客
更多日常分享尽在我的VX公众号:小杨的挨踢IT生活