pei-nwdaf-ml

ML service for training, inference, and anomaly detection on 5G network data.

How it works

• Forecast models: ANN/LSTM models trained on windowed ClickHouse data, predict future metric values
• Anomaly models: Autoencoder-based models detect anomalous windows via reconstruction error
• Training jobs run as Kubernetes Jobs; model weights and scalers stored in MLflow/MinIO
• A background monitoring loop re-evaluates and retrains forecast models when performance degrades

Running

cp .env.example .env
docker compose up -d

Port: 8060

API

Base path: /v1

Forecast

Method	Endpoint	Description
GET	/fields	List available metric fields
POST	/models	Create model config
GET	/models	List models
GET	/models/{id}	Model detail
DELETE	/models/{id}	Delete model
POST	/training/train	Queue training job
GET	/training/jobs	List jobs
GET	/training/jobs/{id}	Job detail
DELETE	/training/jobs/{id}	Cancel job
POST	/inference	Run inference
POST	/performance/{field}/evaluate	Score and elect best model
GET	/performance/{field}/best	Current best model
GET	/performance/{field}/status	Monitoring state machine status

Anomaly

Method	Endpoint	Description
POST	/anomaly/models	Create anomaly model config
GET	/anomaly/models	List anomaly models
GET	/anomaly/models/{id}	Model detail
DELETE	/anomaly/models/{id}	Delete model
POST	/anomaly/training/train	Queue training job
GET	/anomaly/training/jobs	List jobs
POST	/anomaly/detect	Run anomaly detection (explain=true for SHAP attributions)
POST	/anomaly/models/{id}/importance	Compute permutation feature importance
GET	/anomaly/models/{id}/importance	Get cached importance

Environment

Variable	Description
MLFLOW_TRACKING_URI	MLflow server URL
DATABASE_URL	PostgreSQL connection string
DATA_STORAGE_API_URL	Data Storage base URL
TRAIN_USE_KUBE	Run training jobs on Kubernetes (true/false)
TRAIN_KUBE_HOST	Kubernetes API server URL
TRAIN_KUBE_TOKEN	Service account token
TRAIN_KUBE_NAMESPACE	Kubernetes namespace
TRAIN_KUBE_IMAGE	Training worker container image
MONITORING_ENABLED	Enable background monitoring loop (default: true)
MONITORING_INTERVAL_SECONDS	Monitor check interval (default: 300)
API_PORT	Port (default: 8060)