基于城市维度的电动汽车充电桩充电异常识别数据

本数据在充电桩智能运维领域具备精准故障定位能力，具体应用场景如下：
1. 对平台的价值（区域运维效率）：可通过多订单特征分析识别机械部件故障，自动拦截天气、用户行为等临时干扰，推荐维修优先级。当同一城市设备连续多次判定为故障时，自动触发维修工单，减少人工排查成本。
2. 对桩企的价值（区域产品适配）：可用于支撑发现设备与城市特征的关联规律，指导区域定制化产品迭代。
3. 对场站商家的价值（城市服务优化）：监测城市内充电设施健康状态，针对性提升高需求区域的服务稳定性。
4. 对城市管理的价值（设施安全监管与优化）：有助于识别高频故障桩型分布，辅助制定充电桩强制检修政策，规避公共安全隐患。1.数据采集：原始数据经授权合法获取并使用，采集字段包括城市、桩企名称、订单创建时间、充电用户id、异常标签、湿度、订单id、降水量、温度、订单充电时长。
2.特征加工：基于采集数据进行近30天滚动窗口计算。加工字段包含：用户历史充电订单数（统计用户30天内总订单量）、用户历史充电枪数（累计充电枪使用次数）、用户历史充电失败订单数（过滤异常标签≥1的订单计数）、用户历史充电失败枪数（统计失败订单对应枪数）。
3.样本构建：按城市维度抽取异常订单（异常标签≥1）与正常订单（异常标签=0）按1:3比例采样。将订单创建时间前推30天内的特征加工字段与当前订单的采集字段拼接，形成训练样本，每个样本对应唯一订单id。
4.模型训练：采用XGBoost算法构建二分类模型，以异常结论（0/1）为输出目标，使用训练样本进行模型训练，并通过城市维度交叉验证，优化模型对区域共性故障模式的敏感性，利用SHAP值剔除环境因素的干扰。
5.异常诊断：部署时实时采集近20笔异常订单数据，输入模型后输出诊断结论（0-正常或1-异常）。

This dataset possesses precise fault localization capabilities in the intelligent operation and maintenance domain of charging piles, with specific application scenarios as follows:
1. Value for the platform (regional operation and maintenance efficiency): It can identify mechanical component faults via multi-order feature analysis, automatically filter out temporary interferences such as weather and user behaviors, and recommend maintenance priorities. When devices in the same city are consecutively diagnosed as faulty multiple times, it will automatically trigger maintenance work orders, reducing manual troubleshooting costs.
2. Value for charging pile enterprises (regional product adaptation): It can support the discovery of association rules between equipment and urban characteristics, guiding the iteration of regional customized products.
3. Value for charging station operators (urban service optimization): It monitors the health status of charging facilities within the city, and targeted improves service stability in high-demand areas.
4. Value for urban management (facility safety supervision and optimization): It helps identify the distribution of high-frequency faulty pile types, assists in formulating mandatory maintenance policies for charging piles, and avoids public safety hazards.

1. Data Collection: The original data is legally obtained and used with authorization. The collected fields include: city, charging pile enterprise name, order creation time, charging user ID, anomaly label, humidity, order ID, precipitation, temperature, and order charging duration.
2. Feature Engineering: Rolling window calculations over the past 30 days are conducted based on the collected data. The processed fields include: user's historical charging order count (total orders placed by the user within 30 days), user's historical charging gun usage count (cumulative number of charging gun uses), user's historical failed charging order count (count of orders with anomaly label ≥1 after filtering), and user's historical failed charging gun count (number of charging guns corresponding to failed orders).
3. Sample Construction: Abnormal orders (anomaly label ≥1) and normal orders (anomaly label = 0) are sampled at a ratio of 1:3 from each city dimension. The feature engineering fields within 30 days prior to the order creation time are spliced with the collected fields of the current order to form training samples, where each sample corresponds to a unique order ID.
4. Model Training: The XGBoost algorithm is employed to construct a binary classification model, with the anomaly conclusion (0 for normal, 1 for abnormal) as the output target. The model is trained using the training samples, and city-dimension cross-validation is performed to optimize the model's sensitivity to regional common fault patterns. SHAP values are utilized to eliminate interference from environmental factors.
5. Anomaly Diagnosis: During deployment, the latest 20 abnormal order datasets are collected in real time, input into the model, and the diagnosis conclusion (0 for normal or 1 for abnormal) is output.