WMS智能调度实战：构建机器学习特征表的完整指南

从零搭建特征工程流水线，让仓库托盘任务实现动态出库决策

摘要

本文详细介绍如何为WMS（仓库管理系统）构建机器学习特征表，以支持托盘任务的智能调度（一层整托出库 vs 二层拣选出库）。涵盖：环境部署（CentOS 7/麒麟V10）、Oracle数据库（通过DBLINK访问备库）、Python特征工程脚本、库存分区设计（快速区、立库区、散件区、预先成件区）、销售滚动窗口计算、促销特征提取等。提供完整可运行的代码模板，帮助读者快速落地一套自动化特征工程流水线。

---

一、业务背景与特征表设计目标

在出版行业/电商仓库中，托盘任务下发常因重复调出同一托盘而效率低下。核心思路 ：通过机器学习预测SKU的月均销量，并结合库存分布动态决定出库口（一层整托 vs 二层拣选）。为此需要构建一张特征表 WMS_ML_FEATURES，存储每个SKU每月的历史销售、库存、促销等指标。

特征表主要字段：

• 销售特征：最近7/30/90天托数、件数、趋势斜率、环比增长率
• 库存特征：按库位类型拆分为快速区（直发区+越库区）、立库区、散件区、预先成件区的托数和件数
• 商品属性：出版社、分类、上架天数
• 促销特征：是否有促销、平均折扣强度
• 标签：目标月份的实际月均销售托数/件数

---

二、环境准备（以CentOS 7 / 麒麟V10为例）

2.1 安装Python 3.9

sudo yum install -y gcc gcc-c++ make openssl-devel bzip2-devel libffi-devel sqlite-devel
cd /usr/src
sudo wget https://www.python.org/ftp/python/3.9.18/Python-3.9.18.tgz
sudo tar xzf Python-3.9.18.tgz
cd Python-3.9.18
sudo ./configure --enable-optimizations
sudo make -j$(nproc)
sudo make altinstall
# 验证
/usr/local/bin/python3.9 --version

2.2 安装Oracle Instant Client（厚模式，支持Oracle 11g）

# 下载两个RPM包（需Oracle账号）
# oracle-instantclient-basic-19.30.0.0.0-1.x86_64.rpm
# oracle-instantclient-devel-19.30.0.0.0-1.x86_64.rpm
sudo yum localinstall -y oracle-instantclient-*.rpm
echo "/usr/lib/oracle/19.30/client64/lib" | sudo tee /etc/ld.so.conf.d/oracle.conf
sudo ldconfig

2.3 创建项目虚拟环境并安装依赖

mkdir -p /opt/wms_ml
cd /opt/wms_ml
/usr/local/bin/python3.9 -m venv venv
source venv/bin/activate
pip install --upgrade pip

requirements.txt:

pandas==1.5.3
numpy==1.23.5
scikit-learn==1.0.2
lightgbm==3.3.5
sqlalchemy>=1.4.0
oracledb>=1.4.0
python-dateutil>=2.8.0

安装：

pip install -r requirements.txt --only-binary=:all:

---

三、数据库准备（在报表库中执行）

3.1 创建DBLINK指向源库备库

CREATE DATABASE LINK SOURCE_DB_LINK
CONNECT TO readonly_user IDENTIFIED BY "password"
USING '(DESCRIPTION=(ADDRESS=(PROTOCOL=TCP)(HOST=10.0.0.1)(PORT=1521))(CONNECT_DATA=(SERVICE_NAME=STANDBY)))';

3.2 创建特征表 WMS_ML_FEATURES

完整建表语句（含注释）：

CREATE TABLE WMS_ML_FEATURES (
    item_id                 NUMBER(20)   NOT NULL,
    begin_month             DATE         NOT NULL,
    sales_last_7d           NUMBER,
    sales_last_30d          NUMBER,
    sales_last_90d          NUMBER,
    avg_sales_last_30d      NUMBER,
    trend_slope_30d         NUMBER,
    mom_growth              NUMBER,
    monthly_avg_tuo         NUMBER,
    sales_last_7d_qty       NUMBER,
    sales_last_30d_qty      NUMBER,
    sales_last_90d_qty      NUMBER,
    avg_sales_last_30d_qty  NUMBER,
    monthly_avg_qty         NUMBER,
    piece_sales_last_30d    NUMBER,
    avg_piece_sales_per_day NUMBER,
    carton_sales_last_30d   NUMBER,
    avg_carton_sales_per_day NUMBER,
    inv_fast_zone_tuo       NUMBER,
    inv_fast_zone_qty       NUMBER,
    inv_rack_tuo            NUMBER,
    inv_rack_qty            NUMBER,
    inv_piece_zone_tuo      NUMBER,
    inv_piece_zone_qty      NUMBER,
    inv_pre_packed_tuo      NUMBER,
    inv_pre_packed_qty      NUMBER,
    age_days                NUMBER,
    publisher_code          VARCHAR2(50),
    category_code           VARCHAR2(50),
    has_promotion           NUMBER(1),
    promotion_intensity     NUMBER
);
ALTER TABLE WMS_ML_FEATURES ADD CONSTRAINT PK_WMS_ML_FEATURES PRIMARY KEY (item_id, begin_month);
-- 索引
CREATE INDEX idx_ml_features_month ON WMS_ML_FEATURES(begin_month);
CREATE INDEX idx_ml_features_item ON WMS_ML_FEATURES(item_id);
-- 列注释（略）

---

四、特征工程核心代码

4.1 配置文件 config.py

# 报表库连接（目标库）
TARGET_DB = {
    'user': 'report_user',
    'password': '***',
    'host': '10.0.0.2',
    'port': 1521,
    'service_name': 'REPORT'
}
TARGET_DATABASE_URL = f"oracle+oracledb://{TARGET_DB['user']}:{TARGET_DB['password']}@" \
                      f"{TARGET_DB['host']}:{TARGET_DB['port']}/?service_name={TARGET_DB['service_name']}"

# 源库DBLINK名称
SOURCE_DB_LINK = "SOURCE_DB_LINK"

# 表名
FEATURE_TABLE = "WMS_ML_FEATURES"

# 特征列（数值型）
NUMERIC_FEATURES = [
    'sales_last_7d', 'sales_last_30d', 'sales_last_90d',
    'avg_sales_last_30d', 'trend_slope_30d', 'mom_growth',
    'sales_last_7d_qty', 'sales_last_30d_qty', 'sales_last_90d_qty',
    'avg_sales_last_30d_qty',
    'piece_sales_last_30d', 'avg_piece_sales_per_day',
    'carton_sales_last_30d', 'avg_carton_sales_per_day',
    'inv_fast_zone_tuo', 'inv_fast_zone_qty',
    'inv_rack_tuo', 'inv_rack_qty',
    'inv_piece_zone_tuo', 'inv_piece_zone_qty',
    'inv_pre_packed_tuo', 'inv_pre_packed_qty',
    'age_days', 'has_promotion', 'promotion_intensity'
]
CATEGORICAL_FEATURES = ['publisher_code', 'category_code']
ALL_FEATURES = NUMERIC_FEATURES + CATEGORICAL_FEATURES

4.2 数据库工具 db_utils.py

import oracledb
from sqlalchemy import create_engine, text
import pandas as pd
from config import TARGET_DATABASE_URL

# 启用厚模式
oracledb.init_oracle_client(lib_dir="/usr/lib/oracle/19.30/client64/lib")

class OracleDB:
    def __init__(self, url):
        self.engine = create_engine(url, pool_pre_ping=True, pool_recycle=3600)
    def query(self, sql, params=None):
        with self.engine.connect() as conn:
            return pd.read_sql(text(sql), conn, params=params)
    def execute(self, sql, params=None):
        with self.engine.connect() as conn:
            conn.execute(text(sql), params)
            conn.commit()
    def to_sql(self, df, table, if_exists='append'):
        # 简化版写入
        df.to_sql(table.upper(), self.engine, if_exists=if_exists, index=False)

target_db = OracleDB(TARGET_DATABASE_URL)

4.3 工具函数 utils.py

import logging
from datetime import datetime, timedelta
import numpy as np

def setup_logger(name, log_file):
    logger = logging.getLogger(name)
    logger.setLevel(logging.INFO)
    handler = logging.FileHandler(log_file)
    formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
    handler.setFormatter(formatter)
    logger.addHandler(handler)
    return logger

def add_months(dt, months):
    # 月份加减
    new_month = dt.month + months
    new_year = dt.year + (new_month - 1) // 12
    new_month = (new_month - 1) % 12 + 1
    return datetime(new_year, new_month, min(dt.day, 28))  # 简化

def calculate_trend_slope(series):
    x = np.arange(len(series))
    slope = np.polyfit(x, series, 1)[0] if len(series) >= 3 else 0.0
    return slope

def safe_divide(a, b):
    return a / b if b != 0 else 0

4.4 特征工程主脚本 feature_engineering.py（核心片段）

由于代码较长，此处展示关键函数，完整代码请见文末附录。

库存快照函数（分区逻辑）：

def fetch_inventory_snapshot(snapshot_date):
    sql = f"""
    SELECT 
        ext.item_id,
        CASE 
            WHEN loc.loc_type IN ('直发区',  '越库区') THEN '快速区'
            WHEN loc.loc_type IN ( '异形品区', '托盘立库区', '托盘立库大储位区') THEN '立库'
            WHEN loc.loc_type IN ('箱式立库区', '阁楼区',) THEN '散件区'
            WHEN loc.loc_type = '预先成件区' THEN '预先成件区'
            ELSE '其他'
        END AS zone_type,
        SUM(FLOOR(ext.quantity_bu / pu.convert_figure)) AS total_qty,
        SUM(FLOOR(ext.quantity_bu / (pu.convert_figure * i.supportcode_quantity))) AS total_tuo
    FROM WMS_INVENTORY_EXTEND_GOOD_BAK@{SOURCE_DB_LINK} ext
    JOIN WMS_ITEM@{SOURCE_DB_LINK} i ON ext.item_id = i.id
    JOIN WMS_LOCATION@{SOURCE_DB_LINK} loc ON ext.location_id = loc.id
    JOIN WMS_PACKAGE_UNIT@{SOURCE_DB_LINK} pu ON pu.item_id = ext.item_id AND pu.line_no = 3
    WHERE ext.data_date = TRUNC(:snapshot_date)
      AND ext.status = 'ENABLED'
      AND i.supportcode_quantity > 0
      AND loc.loc_type NOT IN ('不良品区','待检区','退货区','报废区','盘点区','待修复区')
    GROUP BY ext.item_id, 
        CASE WHEN loc.loc_type IN ('直发区','越库区') THEN '快速区'
             WHEN loc.loc_type IN ('异形品区','托盘立库区','托盘立库大储位区') THEN '立库'
             WHEN loc.loc_type IN ('箱式立库区','阁楼区') THEN '散件区'
             WHEN loc.loc_type = '预先成件区' THEN '预先成件区'
             ELSE '其他' END
    )
    SELECT item_id, zone_type, SUM(total_qty) AS qty, SUM(total_tuo) AS tuo
    FROM inv_detail WHERE zone_type != '其他' GROUP BY item_id, zone_type
    """
    df = target_db.query(sql, {'snapshot_date': snapshot_date})
    # 透视成宽表...
    return result_df

单月特征构建：

def build_features_for_month(target_month):
    month_start = target_month
    history_end = month_start - timedelta(days=1)
    history_start = history_end - timedelta(days=365)
    
    sales_df = fetch_sales_data(history_start, history_end)
    sku_df = fetch_sku_master()
    inv_df = fetch_inventory_snapshot(month_start - timedelta(days=1))
    promo_df = fetch_promotion_data(month_start, add_months(month_start,1))
    
    rows = []
    for item_id in sales_df['item_id'].unique():
        # 时间序列补齐，滚动求和，斜率计算等
        # ... 详细计算省略
        rows.append({...})
    return pd.DataFrame(rows)

---

五、执行特征工程

5.1 设置时间范围并运行

if __name__ == "__main__":
    start = datetime(2025, 1, 1)
    end = datetime(2026, 3, 1)
    build_full_feature_table(start, end)

5.2 后台执行

cd /opt/wms_ml
source venv/bin/activate
nohup python feature_engineering.py > logs/feature_engineering.log 2>&1 &
tail -f logs/feature_engineering.log

5.3 验证结果

SELECT COUNT(*) FROM WMS_ML_FEATURES;
SELECT * FROM WMS_ML_FEATURES WHERE begin_month = DATE '2026-04-01' AND ROWNUM <= 10;

---

六、常见问题与优化

问题	解决方案
库存查询卡死	创建索引，或者将聚合结果物化到报表库本地表
内存不足	缩小历史窗口（365→180天）或增加swap
终端断开进程终止	使用screen或nohup
缺少_sqlite3	编译Python前安装sqlite-devel
pandas编译失败	使用--only-binary=:all:安装预编译wheel

---

七、总结

通过本文提供的完整流程，你可以在自己的WMS环境中构建出一张高质量的特征表，为托盘任务智能调度提供数据基础。核心要点：

• 环境准备：Python 3.9 + Oracle Instant Client（厚模式）
• 特征设计：销售滚动窗口、库存分区、促销、商品属性
• 性能优化：索引、物化视图、分批处理
• 运维：每日增量更新，保留历史库存备份

下一步：利用特征表训练LightGBM模型，并将预测值集成到WMS出库口决策规则中，实现动态一层/二层出库。

---

附录：完整feature_engineering.py代码

由于篇幅，评论区留言交流获取。