INTRODUCTION

This repository implements SEMAS (Self-Enhanced Multi-Agent System), a multi-layer, multi-agent architecture for industrial anomaly detection in manufacturing environments. The system is designed following a Fog–Edge–Cloud paradigm, enabling scalable, distributed intelligence for data processing, anomaly detection, and system coordination.

The project is organized into modular components, each responsible for a specific layer or function in the SEMAS architecture.

.
├── agents # Organize specific agents according to three layers fog, edge, and cloud
│   ├── cloud_agents.py
│   ├── edge_agents.py
│   ├── fog_agents.py
│   ├── mqtt_agent.py
│   └── semas_agent.py
├── config # Config message broker for agents
│   └── config.py
├── data_processing # Data processing pipeline
│   └── processing.py
├── dataset # Data input storage.
│   ├── Boiler_emulator_dataset.csv
│   └── ieee-phm-2012-data-challenge-dataset
├── messagebroker # Message broker for transfer message.
│   └── broker.py
├── pipeline.py # Initialize full pipeline
├── README.md
├── requirements.txt

Folder Description:

• agents/ Contains the implementation of all agents in the SEMAS architecture, organized by deployment layer:

  • Fog agents handle intermediate aggregation and contextual reasoning.
  • Edge agents perform real-time anomaly detection close to data sources.
  • Cloud agents manage global coordination, system optimization, and long-term knowledge enhancement.

• config/ Stores configuration files for the message broker and system-level parameters used by agents.

• data_processing/ Implements the data processing pipeline, including preprocessing, feature extraction, and preparation for anomaly detection models.

• dataset/ Holds raw input datasets used for experimental evaluation, including boiler fault data and industrial challenge datasets.

• messagebroker/ Provides the messaging infrastructure that enables communication between distributed agents using a broker-based architecture.

INSTALLATION

Setup environment

uv venv
source .venv/bin/activate
uv pip install -r requirements.txt

TRAINING

To run the full pipeline, execute the main script:

python pipeline.py

It will do following steps:

• Load and preprocess data of boiler and wind turbin from dataset folder
• Run training and evaluation anormaly detection based on Multi-Agent System

DATASET

There are two datasets used in this project:

1. Boiler Dataset: a simulated industrial dataset that models the operation of a steam boiler system under both normal and faulty conditions.

• Source: https://ieee-dataport.org/open-access/simulated-boiler-fault-data

2. Turbin Dataset: this dataset contains SCADA (Supervisory Control and Data Acquisition) data collected from a real wind turbine.

• Source: https://www.kaggle.com/datasets/berkerisen/wind-turbine-scada-dataset

You can download and organize into folder dataset:

dataset
├── Boiler_emulator_dataset.csv
└── ieee-phm-2012-data-challenge-dataset-master
    ├── Full_Test_Set
    ├── Learning_set
    └── Test_set

RESULT

BOILER DATASET

Iteration	Accuracy	Precision	Recall	F1-score	ROC-AUC	Eval Time (s)	Predict Time (s)	RUL MAE	RUL RMSE
1	0.5306	0.3929	0.8352	0.5344	0.6695	0.0139	0.6512	35.2992	42.5011
2	0.5154	0.3866	0.8557	0.5325	0.6695	0.0079	0.3224	35.2992	42.5011
3	0.4980	0.3786	0.8676	0.5272	0.6695	0.0080	0.4006	35.2992	42.5011

Average F1: 0.5314, Precision: 0.3860, Recall: 0.8528

WIND_TURBINE DATASET

Iteration	Accuracy	Precision	Recall	F1-score	ROC-AUC	Eval Time (s)	Predict Time (s)	RUL MAE	RUL RMSE
1	0.5000	0.4898	1.0000	0.6575	0.5176	0.0017	0.3469	22.4011	28.3170
2	0.5000	0.4898	1.0000	0.6575	0.5180	0.0018	0.0385	22.4011	28.3170
3	0.5000	0.4898	1.0000	0.6575	0.5184	0.0018	0.0374	22.4011	28.3170

Average F1: 0.6575, Precision: 0.4898, Recall: 1.0000

CITATION

@misc{saleh2026selfevolvingmultiagentnetworkindustrial,
      title={Self-Evolving Multi-Agent Network for Industrial IoT Predictive Maintenance}, 
      author={Rebin Saleh and Khanh Pham Dinh and Balázs Villányi and Truong-Son Hy},
      year={2026},
      eprint={2602.16738},
      archivePrefix={arXiv},
      primaryClass={cs.MA},
      url={https://arxiv.org/abs/2602.16738}, 
}