DASHing Factory is a modular research platform designed to study, simulate, emulate, and optimize video streaming and network performance in 5G factory automation environments. It provides an end-to-end workflow including:
- NS-3-based 5G/NR simulation : 5G-LENA, realistic channel models.
- Mininet-based link emulation : TC/NetEm.
- Real DASH streaming : Caddy + goDash.
- Dataset generation : from logs, trace files, pcaps, and QoE metrics.
- Machine Learning / Deep Learning analytics : for QoE prediction and anomaly detection.
The project is aligned with specifications from 3GPP, 5G-ACIA, NIST, and ISA, and targets realistic Industry 4.0 factory scenarios involving mobile robots, cameras, and delay-sensitive video streaming applications.
Outline of the DASHing Factory platform architecture : (1) Scenario Specification (3GPP-aligned), (2) Network Simulation (NS-3), (3) Network Emulation (Mininet, Linux TC, NetEm), (4) Data Collection and Processing (Pandas), and (5) Insights/Predictions (e.g., QoE prediction / anomaly detection - tensorFlow, scikit-learn).
The dataset comprises +20 features, organized as follows :
| Player-Features | Algorithm : six variantes tested : Conventional, Elastic, Arbiter +, BBA, Logistic, Exponential. Seg_Dur : segment duration in ms, Width & Height: in pixels, Play_Pos : current Playback position, Stall_Dur : stall or freeze duration in ms, Buff_Level : buffer level in ms |
| Application Features | RTT : determined using HTTP head request (ms), persegment_RTT : uplink RTT at the network level, Throughput : downlink throughput, Packets : downlink number of packets |
| Per-Segment Statistics | Segment_no : segment number (60 segments streamed, of 2s each), Arr_time : arrival time, Del_Time : delivery of the segment, Rep_Level : representation selected for the segment, Del_Rate : delivery rate of network, Act_Rate : actual rate in Kbps, Byte_Size : byte size of this segment, tag : target variable in case of Anomaly Detection : 10 classes availables : Normal, or Anomaly injected (duplication, reordering or corruption each with three level of gravity : low 5%, medium 30% or high 60%) |
| QoE related features | five quality of experience models : P1203, Clae, Duanmu, Yin, Yu. |
The dataset is provided for research and educational purposes only. The final dataset is available under ./dataset/dashing_factory_v03.csv.
The repository contains subfolders for each major version of the Dashing Factory platform. Each version addresses different layers of the pipeline:
- Version 01 : Foundational NS-3 Simulation Layer : foundation for understanding protocol-level behavior of DASH streaming in indoor 5G factories.
- Version 02 : Updated Models + High-Quality Datasets : refined the simulation environment using updated international guidelines:
- Version 03 : ML-Driven Analytics Layer (current directory) : introducing a complete ML/DL-based anomaly detection framework.
- Extends the pipeline with a Mininet-based emulation layer, where tc/netem is used to inject controlled network anomalies (packet reordering, duplication, and corruption), each applied at multiple severity levels, c.f.
./testbed. - Combined with real DASH traffic and NS-3–derived link conditions, this version :
- Generates a unified dataset (~22k samples)
- Applies ML/DL models: SVM, MLP, TabNet, XGBoost for Multi-perspective anomaly detection : Binary classification, Multi-class (normal, reorder, duplicate, corrupt), High-gravity classification (severity +50%), Fine-grained classification (type × severity), c.f.
./notebooks. - Extensive evaluation and results (k-fold CV, accuracy/F1 metrics).
- Extends the pipeline with a Mininet-based emulation layer, where tc/netem is used to inject controlled network anomalies (packet reordering, duplication, and corruption), each applied at multiple severity levels, c.f.
For a complete and in-depth description of the emulation pipeline, NS-3 configuration, dataset generation process and ML/DL models please refer to the full research documents:
📄 A. Kabou, M. Khanfouci (2025), DASHing Factory: A Data-Driven Simulation and Emulation Platform for Optimizing 5G Industrial Applications, Mitsubishi Electric R&D Centre Europe, Rennes, France.
📄 Kabou, A., & Khanfouci, M. (2025). Feature Selection for Data-Driven Optimization: A Case Study on Adaptive Video Streaming in 5G-Enabled Factories. In 2025 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit).
The platform was developed as part of research activities at Mitsubishi Electric R&D Centre Europe (MERCE), Rennes, France, within research activities on QoE-aware networking and 5G industrial systems. The opinions and results presented in this repository are those of the authors and do not necessarily reflect official product developments.
This project is released under the MIT License - see the LICENSE file for details.