AD-STGN for RCA in CMMS

This repository provides the code and data implementation for the AD-STGN model, designed for Root Cause Analysis (RCA) in complex, nonlinear, and highly coupled continuous manufacturing processes. Datasets Overview: To validate the effectiveness of the framework, experiments are conducted on two widely-recognized cyber-physical benchmarks and one real-world industrial case study. 1. Tennessee Eastman Process (TEP) The TEP dataset is a widely adopted benchmark for RCA, simulating a complex, nonlinear continuous process. System Network: Spatial graph comprising N=51 topological nodes (40 continuous sensor measurements, 11 manipulated control actions). Training Set: 230,500 steady-state samples (used exclusively to prevent data leakage and for causal graph extraction). Testing Set: Mixed sequences constructed by stitching 1,081 steady-state testing samples and 800 faulty testing samples. 2. Secure Water Treatment (SWaT) Operational data collected from a raw water purification plant across six physical stages (raw water supply, chemical dosing, ultrafiltration, dechlorination, reverse osmosis, and backwash). System Network: N=50 input nodes (24 continuous sensors, 26 discrete actuators) and 1 prediction target. Training Set: 496,800 steady-state samples representing normal behavior (collected over 7 days). Testing Set: 449,919 samples containing anomalous events (collected over 4 days of cyber-physical attacks). 3. Real-world Case Study: Injection Molding Production Line Deployed on a real-world injection molding production line of a leading global Automotive Electronics manufacturer in Tianjin. Data was collected from Manufacturing Execution Systems and IoT sensors between February 2, 2025, and March 14, 2025. Process Stages: Clamping, injection, holding, cooling, ejection, and robot picking/placing. System Network: 66 continuous sensor measurements (granular process dynamics) and 7 discrete control actions (machine setpoints and equipment configurations). Terminal State Indicator: X66 (part temperature when placed by the robotic arm). It integrates the cumulative thermal-mechanical history. Deviations correlate with terminal Injection Molding Process Defect (PMT) issues (e.g., internal bubbles, warpage, incomplete cooling). Data Split: Training Set: 88,000 normal steady-state samples. Validation Set: 22,614 normal samples. Testing Set: Chronological sequence where anomalies start from the 75th sample, representing a transition from normal operation to a PMT outbreak.本仓库提供了针对复杂非线性强耦合连续制造过程根因分析（Root Cause Analysis, RCA）的AD-STGN模型的代码与数据实现方案。 数据集概览： 为验证所提框架的有效性，本研究在两项广泛认可的信息物理系统基准数据集与一项真实工业案例研究上开展了实验。 1. 田纳西东蒙过程（Tennessee Eastman Process, TEP） TEP数据集是根因分析领域广泛使用的基准数据集，模拟了复杂非线性连续制造过程。 系统网络：由51个拓扑节点构成的空间图（包含40个连续传感器测量值与11个操控控制动作）。 训练集：230,500个稳态样本（仅用于防止数据泄露与因果图提取）。 测试集：由1,081个稳态测试样本与800个故障测试样本拼接得到的混合序列。 2. 安全水处理系统（Secure Water Treatment, SWaT） 该数据集采集自一座原水净化厂的六个物理处理阶段（原水供应、药剂投加、超滤、脱氯、反渗透与反冲洗）的运行数据。 系统网络：包含50个输入节点（24个连续传感器、26个离散执行器）与1个预测目标。 训练集：496,800个代表正常运行状态的稳态样本（采集时长7天）。 测试集：449,919个包含异常事件的样本（采集自4天的信息物理攻击实验）。 3. 真实工业案例研究：注塑生产线 该案例部署于天津某全球领先汽车电子制造商的真实注塑生产线。数据于2025年2月2日至2025年3月14日期间，通过制造执行系统（Manufacturing Execution Systems, MES）与物联网（Internet of Things, IoT）传感器采集得到。 工艺流程阶段：合模、注塑、保压、冷却、顶出与机器人取放料。 系统网络：66个连续传感器测量值（用于刻画精细的过程动态）与7个离散控制动作（包含机器设定值与设备配置参数）。 终端状态指标：X66（机械臂取件时的工件温度），该指标整合了累积热机械历程，其偏差与注塑成型终端缺陷（Injection Molding Process Defect, PMT）问题相关，例如内部气泡、翘曲变形与冷却不充分。 数据划分：训练集：88,000个正常稳态样本。 验证集：22,614个正常样本。 测试集：按时间顺序排列的序列，异常事件从第75个样本开始，代表从正常运行到PMT缺陷爆发的过渡过程。