AI-Driven Edge Computing: An Approach to Architect Sustainable AI Training Platforms

Problem Statement

With Increasing AI workloads and IOT data streams, current centralized AI training platforms face scalability and cost issues. Centralized AI and Data cause massive costs and less scalable. High load on central servers, network congestion are the main issues.

Our Solution

We introduce an AI training and analytics platform which utilises edge resources for running AI and data workloads. Our platform provides a local Streamlit application for users to collaborate and train AI models locally with zero coding while utilising their resources. This does not involve moving the dataset through the network, which lowers the load on central servers and the network. The platform also provides storage for IOT data and IoT-related AI predictions on edge servers. Pre-trained Public Models are also moved from central servers to edge servers for faster delivery and lower network congestion.

User Model Training is done as Federated Learning though Flower framework.

Key Features:

• Users can train and use AI models through a user-friendly UI without any programming knowledge.
• Multiple users can train models collaboratively while keeping sensitive data locally, which ensures data privacy and reduces network traffic.
• Users can train a variety of custom AI models related to classification, regression, forecasting, anomaly detection, etc.
• The system is capable of processing live sensor/IoT data at the edge servers to collect data for model training, inference, and providing live insights.
• Collection of pre-trained models is cached and delivered through edge servers to users with minimal latency.
• Users can view real-time analytics and predictions related to sensor/IoT data in a centralized dashboard.

Useful Info

grafana

username: admin
pword: fyp12345

influxdb

username: navod
pword: fyp12345
organization: fyp
token: 3wvWUxmtdBM03hm9YgTEa91s6ofQ73G4gQ54uNR0Ek59zpJNMGOagj1UR1GKw3D1f5Elw-zS78rEwY7akZGmOw==

Access MLflow ui

url:http://localhost:5001/

Run The Simulation

1. build airflow image with packages

cd edge_server/airflow_env
docker build . --tag extended_airflow:latest

2. initialize airflow

in project root

docker compose up airflow-init

3. start influxdb, grafana, MLflow, MySQL services

in project root

docker-compose up -d

4. ingest models to MLflow

run temp python files in mlflow folder

5. start central server

cd central_server
uvicorn api:app --host 0.0.0.0 --port 8000 --reload

6. ingest events

cd central_server/event_ingester
python events_ingest.py

7. start edge server

cd edge_server
uvicorn api:app --host 0.0.0.0 --port 8001 --reload

8. start common_cameras

cd sensors/common_cameras
python camera.py

9. start power plants

cd sensors/power_plants
python plants.py

10. start traffic cameras

cd sensors/traffic_cameras
python camera.py

11. start weather data stream

cd sensors/weather_sensors
python weather_stream.py

12. start realtime solar power prediction

cd edge_server/power_gen_pred
python realtime_pred.py

13. start realtime weather prediction

cd edge_server/weather_pred
python realtime_pred.py

14. start realtime traffic prediction

cd edge_server/traffic_pred
python traffic_pred.py

15. start realtime crowd prediction

cd edge_server/crowd_pred
python crowd_pred.py

16. run local web app:

cd desktop_app
streamlit run app.py