AI-Driven SON Tool for Frequency Optimization in Unlicensed Spectrum

Technologies Used

Python (Pandas, NumPy, Glob): Automated data ingestion, transformation, and merging
The SON Tool leverages Python as the foundation for data handling.

• Pandas provides structured dataframes for reading and combining thousands of network logs from multi-vendor devices.

• NumPy supports large-scale numerical computations and feature vector manipulations needed for KPI normalization and correlation analysis.

• Glob automates the retrieval and aggregation of multiple Excel or CSV files from network folders, enabling efficient monthly consolidation of performance data.
  Together, these libraries streamline the ingestion-to-processing pipeline, converting fragmented vendor logs into a unified, analysis-ready dataset in a consistent format across all Base Stations and CPEs.

Machine Learning (PCA, Mutual Information Gain): Feature selection and interference modeling
The interference detection engine is powered by two core ML techniques:

• Principal Component Analysis (PCA): Reduces dimensionality from 30+ raw KPIs into 3–5 principal components, capturing 90–95% of the signal variance. This highlights the most impactful features influencing network interference (like RSSI, CINR, Channel Utilization).

• Mutual Information Gain (MI): Quantifies non-linear dependencies between KPIs and interference events. It identifies which variables carry the most predictive value for link degradation.
  The output of these models forms a weighted KPI matrix, which classifies each frequency channel as “clean” or “interfered,” forming the analytical foundation for AI-driven frequency recommendations.

AI/ML Frameworks: Custom models for KPI analysis, pattern recognition, and threshold optimization
Custom AI pipelines are implemented to detect performance anomalies, classify interference patterns, and refine decision thresholds dynamically.

• Trained models continuously learn from historical KPI behavior, comparing trends across multiple time windows (daily, monthly, yearly).

• Adaptive logic recalibrates thresholds based on seasonal network load variations, ensuring consistent accuracy across varying traffic and environmental conditions.
  This enables real-time self-optimization — the model evolves with the network, reducing false positives and improving prediction confidence over time.

Visualization Stack: Web-based UI with interactive frequency tables and map integration
The front-end visualization framework translates AI outputs into an intuitive, engineer-friendly interface.

• Interactive frequency tables display “Good” (interference-free) and “Bad” (interfered) channels for every Base Station.

• A feedback module allows engineers to validate AI recommendations with thumbs-up/down ratings, enhancing the tool’s learning loop.

• Integrated performance cards summarize key network insights — including link quality, CPE counts, and interference rankings — enabling rapid root cause analysis.
  This ensures that data-driven intelligence is directly usable by network operations teams without needing to interpret raw analytics.

Geospatial Mapping: Lat-long analysis for azimuth overlaps and interference zone detection
Spatial intelligence is built into the system to visualize and analyze interference topologies.

• Each Base Station and CPE is mapped using latitude and longitude coordinates.

• Azimuth overlap algorithms calculate directional beam intersections across sectors to detect potential self-interference zones.

• A dynamic 360° interference heat map helps engineers identify coverage overlaps, cluster hotspots, and optimize antenna alignment or beamwidth settings.
  This provides a visual, real-world context to the analytical model outputs, merging spatial and signal-level intelligence.

NMS Integration: APIs for automatic frequency update and trap-based event handling
The SON Tool is integrated with the customer’s Network Management System (NMS) via RESTful APIs.

• It can automatically push recommended frequency changes based on real-time link quality triggers.

• Trap-based event handling allows the SON system to react to degradation alerts (e.g., LQI alarms) by instantly re-evaluating frequency options.

• Manual override and validation options remain available through the UI to ensure safe network operations and human-in-the-loop governance.
  This integration bridges AI insights with actual network control, enabling a closed-loop optimization workflow from detection → recommendation → execution.

Database Layer: Structured data repository for KPI histories and frequency recommendations
A structured relational data store retains all historical KPI data, interference patterns, and AI recommendation logs.

• It acts as a knowledge base for model retraining and for benchmarking frequency stability across time.

• The repository enables query-based retrieval of BS and CPE performance histories, supporting predictive diagnostics.

• Versioned recommendations allow auditability — showing which frequency changes were made, when, and under what network conditions.
  This foundation ensures that every model output is traceable, explainable, and continuously improving through cumulative intelligence.