MCP 服务器

Industrial AI Assistant MCP Server

Enables AI assistants to monitor and analyze industrial process data from DCS/SCADA systems via MCP protocol, with domain expert knowledge for anomaly detection and recommendations.

README

Industrial AI Assistant Framework

An open-source framework for integrating LLM-based AI assistants with Industrial Control Systems (ICS/DCS) via MCP protocol and Composio, featuring domain expert knowledge embedding.

English | 中文

<a name="english"></a>English

Overview

This project provides a reference architecture and implementation framework for building AI-powered assistant systems that can:

Monitor real-time industrial process data from DCS/SCADA systems
Apply domain expert knowledge for intelligent anomaly detection
Generate actionable recommendations for control room operators
Provide field operation guidance through mobile interfaces

Key Innovation: The framework demonstrates how to embed domain expert thinking patterns (like pressure balance analysis, causal chain reasoning) into LLM-based AI systems for industrial applications.

⚠️ Important Disclaimer

This is a reference implementation and research framework only:

✅ Suitable for: Learning, research, proof-of-concept development
✅ Contains: Generic architecture patterns, integration methodologies
❌ Does NOT contain: Production-ready code, specific device configurations, proprietary industrial data
⚠️ WARNING: Direct connection to production DCS systems requires thorough security assessment and should only be done by qualified industrial automation professionals

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                    AI Assistant Layer                            │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐          │
│  │ Domain Expert │  │  Knowledge   │  │  Decision    │          │
│  │   Persona    │  │   Retrieval  │  │   Engine     │          │
│  └──────┬───────┘  └──────┬───────┘  └──────┬───────┘          │
│         └─────────────────┼─────────────────┘                   │
│                           ▼                                     │
│                ┌─────────────────────┐                         │
│                │     MCP Server      │                         │
│                │  (Model Context     │                         │
│                │   Protocol)         │                         │
│                └──────────┬──────────┘                         │
└───────────────────────────┼─────────────────────────────────────┘
                            │
┌───────────────────────────┼─────────────────────────────────────┐
│                    Integration Layer (Composio)                  │
│                           ▼                                      │
│                ┌─────────────────────┐                          │
│                │   Protocol Adapters │                          │
│                │  (OPC UA/Modbus/API)│                          │
│                └──────────┬──────────┘                          │
└───────────────────────────┼─────────────────────────────────────┘
                            │
┌───────────────────────────┼─────────────────────────────────────┐
│              Industrial Control System (Read-Only)               │
│                           ▼                                      │
│              ┌─────────────────────────┐                        │
│              │  DCS/SCADA/PLC Systems  │                        │
│              │    (Via OPC UA/etc)     │                        │
│              └─────────────────────────┘                        │
└─────────────────────────────────────────────────────────────────┘

Core Concepts

1. Domain Expert Knowledge Embedding

The framework demonstrates how to structure and embed industrial domain knowledge:

# Example: Expert reasoning pattern
class ExpertAnalyzer:
    """
    Implements expert thinking patterns for industrial process analysis:
    - Pressure balance analysis
    - Causal chain derivation
    - Design intent understanding
    """

    def analyze_pressure_system(self, upstream_pressure, downstream_pressure):
        """
        Expert heuristic: All fluid systems are pressure balance problems
        """
        pressure_diff = upstream_pressure - downstream_pressure

        if pressure_diff < 0:
            return {
                "issue": "Pressure mismatch",
                "reason": "Low pressure cannot feed high pressure system",
                "recommendation": "Check pump status or upstream blockages"
            }

        return {"status": "normal", "pressure_diff": pressure_diff}

2. Safety-First Architecture

Security Zones:
┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│   Level 3   │ ←→  │   Level 2   │ ←→  │   Level 1   │
│  (AI Layer) │     │ (Data Layer)│     │ (DCS Layer) │
│             │     │             │     │  READ-ONLY  │
└─────────────┘     └─────────────┘     └─────────────┘
       ↑                                    ↑
   Unidirectional                      Control Network
   Data Diode                         (Isolated)

Key Safety Principles:

✅ Read-only by default - AI never writes to DCS directly
✅ Human-in-the-loop - All recommendations require operator confirmation
✅ Fail-safe design - AI failure does not affect DCS operation
✅ Physical isolation - Network segmentation with data diodes

Project Structure

industrial-ai-assistant/
├── docs/
│   ├── architecture.md           # System architecture documentation
│   ├── security.md               # Security guidelines
│   ├── mcp-integration.md        # MCP protocol integration guide
│   └── expert-knowledge.md       # How to embed domain expertise
├── src/
│   ├── mcp_server/               # MCP server implementation
│   │   ├── server.py
│   │   └── tools/
│   ├── knowledge_base/           # Knowledge base structures
│   │   ├── equipment.py
│   │   ├── process.py
│   │   └── expert_patterns.py
│   ├── analyzers/                # Analysis engines
│   │   ├── pressure_analyzer.py
│   │   ├── trend_analyzer.py
│   │   └── causal_analyzer.py
│   └── interfaces/               # User interfaces
│       ├── control_room.py
│       └── field_assistant.py
├── examples/
│   ├── chemical_process/         # Chemical industry example
│   ├── power_plant/              # Power generation example
│   └── oil_gas/                  # Oil & gas example
├── tests/
└── README.md

Quick Start

Prerequisites

Python 3.9+
Node.js 18+ (for MCP components)

Installation

# Clone the repository
git clone https://github.com/yourusername/industrial-ai-assistant.git
cd industrial-ai-assistant

# Install dependencies
pip install -r requirements.txt

# Install MCP SDK
npm install @modelcontextprotocol/sdk

Basic Usage

from industrial_ai import MCPClient, ExpertAnalyzer

# Initialize MCP client
client = MCPClient(
    server_url="opc.tcp://your-dcs-server:4840",
    read_only=True  # Safety: enforce read-only mode
)

# Connect to knowledge base
knowledge_base = ExpertAnalyzer.load_knowledge_base("examples/chemical_process")

# Analyze real-time data
while True:
    data = client.read_process_data(tags=["TEMP_001", "PRESS_001", "FLOW_001"])

    # Apply expert analysis
    result = knowledge_base.analyze(data)

    if result.has_anomaly:
        print(f"⚠️ Anomaly detected: {result.description}")
        print(f"💡 Recommendation: {result.recommendation}")

    time.sleep(5)

Integration Examples

OPC UA Integration

from industrial_ai.adapters import OPCUAAdapter

adapter = OPCUAAdapter(
    endpoint="opc.tcp://dcs-server:4840",
    security_policy="Basic256Sha256"
)

# Read process variables
values = adapter.read_values([
    "ns=2;s=Reactor.Temperature",
    "ns=2;s=Reactor.Pressure"
])

MCP Server Configuration

{
  "mcpServers": {
    "industrial-dcs": {
      "command": "python",
      "args": ["-m", "industrial_ai.mcp_server"],
      "env": {
        "DCS_ENDPOINT": "opc.tcp://localhost:4840",
        "READ_ONLY_MODE": "true"
      }
    }
  }
}

Knowledge Base Template

Create your own domain knowledge base:

# knowledge_base/equipment_template.yaml
equipment:
  - id: REACTOR_001
    name: Gasification Reactor
    type: reactor
    parameters:
      - name: temperature
        tag: TI_001
        normal_range: [1400, 1700]
        unit: celsius
        alarms:
          high: 1750
          high_high: 1800

      - name: pressure
        tag: PI_001
        normal_range: [4.5, 5.5]
        unit: mpa

    expert_rules:
      - name: pressure_balance_check
        description: "Verify pressure differential across reactor"
        condition: "inlet_pressure - outlet_pressure < 0.1"
        action: "alert_operator"

      - name: temperature_trend
        description: "Monitor temperature change rate"
        condition: "temperature_slope > 50"
        action: "predictive_alert"

<a name="中文"></a>中文

项目概述

本项目提供了一个参考架构和实现框架，用于构建能够执行以下功能的AI辅助工业控制系统：

从DCS/SCADA系统监控实时工业过程数据
应用领域专家知识进行智能异常检测
为控制室操作员生成可执行的建议
通过移动界面提供现场操作指导

核心创新：该框架展示了如何将领域专家思维模式（如压力平衡分析、因果链推导）嵌入到基于LLM的工业AI系统中。

⚠️ 重要声明

这只是一个参考实现和研究框架：

✅ 适用于：学习、研究、概念验证开发
✅ 包含：通用架构模式、集成方法论
❌ 不包含：生产就绪代码、特定设备配置、专有工业数据
⚠️ 警告：直接连接生产DCS系统需要全面的安全评估，且只能由合格的工业自动化专业人员执行

核心概念

1. 领域专家知识嵌入

框架展示了如何结构化和嵌入工业领域知识：

# 示例：专家推理模式
class ExpertAnalyzer:
    """
    实现工业过程分析的专家思维模式：
    - 压力平衡分析
    - 因果链推导
    - 设计意图理解
    """

    def analyze_pressure_system(self, upstream_pressure, downstream_pressure):
        """
        专家启发式：所有流体系统都是压力平衡问题
        """
        pressure_diff = upstream_pressure - downstream_pressure

        if pressure_diff < 0:
            return {
                "issue": "压力不匹配",
                "reason": "低压无法向高压系统输送",
                "recommendation": "检查泵状态或上游堵塞"
            }

        return {"status": "正常", "pressure_diff": pressure_diff}

2. 安全第一架构

关键安全原则：

✅ 默认只读 - AI绝不直接向DCS写入
✅ 人机协同 - 所有建议需要操作员确认
✅ 故障安全设计 - AI故障不影响DCS运行
✅ 物理隔离 - 通过网络分段和数据隔离实现

快速开始

# 克隆仓库
git clone https://github.com/yourusername/industrial-ai-assistant.git
cd industrial-ai-assistant

# 安装依赖
pip install -r requirements.txt

# 运行示例
python examples/chemical_process/demo.py

文档

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

License

This project is licensed under the MIT License - see LICENSE file for details.

Acknowledgments

Model Context Protocol (MCP) - Anthropic's open protocol
Composio - Integration framework
Industrial automation community for domain knowledge patterns

Contact

Issues: GitHub Issues
Discussions: GitHub Discussions

Disclaimer: This is an open-source research project. Use in production environments at your own risk and only after thorough security assessment.