Orange
BotiqueAI
📡 Orange🔬 Machine LearningClustering · Unsupervised · 5G Networks
Case Study

ML for 5G Beamforming

Using unsupervised machine learning to group users by behaviour and signal profile, so antenna beams serve more people, more efficiently.

ML
Driven approach
No hard-coded rules
Unsup.
Learning
No labelled training data
Clustering
Users
By behaviour & signal profile
5G
Optimisation target
Beamforming efficiency
The Challenge

Serve more users.
With the same antenna.

5G antennas can steer radio beams with high precision, but deciding which beam shape to use, for which group of users, at which moment, is an enormously complex optimisation problem. Traditional approaches rely on fixed configurations that leave significant capacity on the table.

Orange needed a data-driven, scalable approach that could automatically learn user patterns, without requiring labelled training data or manual rule engineering.

Static beam configurations
Traditional beamforming uses fixed patterns that don't adapt to where users actually are or how they move.
Per-user optimisation doesn't scale
Computing an individual beam configuration for every user in real-time is computationally prohibitive at 5G scale.
Heterogeneous usage patterns
Users vary wildly: in location, device, mobility, and data demand. No single beam strategy works for all of them.
The Solution

Cluster the users,
optimise the beams.

Instead of trying to optimise a beam for every individual user in real time, we learned to group users into behaviorally coherent clusters using unsupervised ML. Each cluster gets one optimised beam configuration, a far more scalable approach that maintains high throughput without exponential compute cost.

01
Feature engineering
Extract signal-level features per user: RSRP, RSRQ, SINR, spatial position, throughput history, mobility patterns.
02
Unsupervised clustering
K-Means and DBSCAN group users with similar profiles, no labels needed. Each cluster represents a distinct usage archetype.
03
Per-cluster beam optimisation
Instead of one beam per user, one optimised beam configuration per cluster. Dramatic reduction in computation with minimal throughput loss.
04
Evaluation & iteration
Silhouette score, Davies-Bouldin index and simulated throughput metrics validate cluster quality. Models retrain as user distributions shift.
K-Means
Primary algorithm
Spatial & signal-based grouping
DBSCAN
Outlier detection
Removes noise before clustering
PCA
Dimensionality
Reduces feature space first
Zero
Labelled data
Fully unsupervised pipeline
ML Techniques

The unsupervised toolkit.

📍
K-Means: spatial clustering

Users grouped by position and signal profile. K-Means finds natural groupings without supervision, each centroid representing an ideal beam target.

Euclidean distanceElbow method for kCentroid-based
🌀
DBSCAN: noise filtering

Identifies outlier users who don't fit any cluster (isolated users or edge-of-cell cases) and removes them before main clustering to improve quality.

Density-basedNo cluster count neededRobust to outliers
📉
PCA: dimensionality reduction

Compresses the high-dimensional signal feature space before clustering. Reduces noise, speeds up training, and reveals the dominant variance directions.

Principal componentsVariance explainedPre-clustering step
📊
Evaluation metrics

Silhouette score measures how well-separated clusters are. Davies-Bouldin index measures compactness. Simulated throughput validates real-world impact.

Silhouette scoreDavies-BouldinThroughput simulation
Tech Stack

Pure ML, no black boxes.

🐍
Python
Core language for data processing, feature engineering and model training
🔢
Scikit-learn
K-Means, DBSCAN, PCA: the full unsupervised toolkit
📐
NumPy / Pandas
Signal data ingestion, reshaping, and feature matrix preparation
📊
Matplotlib / Seaborn
Cluster visualisation and evaluation plot generation
📡
Radio signal data
RSRP, RSRQ, SINR, spatial coordinates and mobility traces per user
🔄
Iterative retraining
Models retrain as user density patterns and network topology evolve
BotiqueAI
Applied ML for industrial use cases
← Back to portfolio