Training / Course 05
Aims — what you can do after
- Map every data source a 5G/IMS network emits to its schema and its join keys.
- Ingest heterogeneous telemetry into tidy tables and build a time-aligned feature store.
- Assess data quality: missingness, clock skew, duplication, and the defects that silently poison a pipeline.
- Analyse with the right statistics: distributions, percentiles, and standardized KPIs from raw events.
- Work with time series: resampling, seasonality decomposition and change-point detection.
- Visualise for two audiences — engineering diagnostics and operational status.
- Engineer features and avoid the traps that make ML lie: leakage and improper time-series splitting.
- Detect anomalies without labels, classify faults from shape, and forecast load — each judged honestly.
- Operationalise: evaluation campaigns, drift detection, and serving analytics back to the network.
From packets to predictions
IngestParse captures, logs, metrics and database state into tidy tables.
↓
Feature storeSessionize events into calls, time-align, and build one clean feature table.
↓
AnalyticsPercentiles, standardized KPIs, seasonality and change-points.
↓
Machine learningAnomaly detection, fault classification and forecasting.
↓
ServeFeed the results back to the analytics function — the loop closed.
Three model families
Anomaly · Isolation Forest → Classify · RandomForest + DTW → Forecast · seasonal + backtest
The same models the reference analytics function and the digital-twin agent run — an Isolation Forest for anomalies, time-series matching for root cause, and a seasonal baseline for forecasting.
The honesty rule
A model that scores 0.99 on your own data ... -> is usually a leak, not a result Every ML module ends by asking: -> what would be true on REAL traffic, not synthetic? # the skill is the pipeline and the scepticism, not the score
The habit the ML modules train: trust the honest lower number.
Two editions
SyntheticA generator grounded in the real schemas — runs anywhere, no lab, for learning the pipeline.
↓
Live-labThe same pipeline on real telemetry: it deploys a collector, drives traffic, and injects real faults to build a labelled dataset.
Module breakdown
0
Foundation & the data engine25 min
The scientific stack and the synthetic-telemetry generator.
1
The data landscape30 min
Sources, schemas, identifiers, and the cross-plane join problem.
2-3
Ingestion & the feature store80 min
Parsing, sessionization, time-alignment — the load-bearing steps.
4
Data quality30 min
Missingness, clock skew, dedup, and a validation gate.
5-6
Analytics & time series75 min
Percentiles, standardized KPIs, seasonality and change-points.
7-8
Visualization65 min
KPI dashboards, call-flow timelines, and audience-appropriate views.
9
Feature engineering & leakage35 min
Honest features and chronological splitting.
10
Anomaly detection40 min
Baselines and an Isolation Forest — the analytics-function model.
11
Fault classification45 min
Tree models and time-series matching — the root-cause method.
12
Forecasting35 min
Seasonal baselines, backtesting, and honest skill.
13
MLOps & closing the loop35 min
Evaluation campaigns, drift, serving, and the human in the loop.
What modelling the network looks like
- Build a feature store and know why a shuffled split or a leaked feature is cheating.
- Detect anomalies without labels, and classify a fault by its shape.
- Quote a KPI as a percentile, never a bare average — the tail is the user experience.