Real Data Sources
for AI Models

MDT allows operators to collect RF measurements (RSRP, RSRQ, SINR) directly from subscriber UEs without deploying drive test teams. Defined in 3GPP TS 32.422, MDT comes in two modes: Immediate MDT (real-time reporting for connected UEs) and Logged MDT (UE stores measurements in idle mode and uploads later). A single cell can generate 10K-100K MDT samples per day, providing coverage maps that no drive test could match.

Drive tests use professional UE scanners (TEMS, Nemo, XCAL) mounted in vehicles to collect high-resolution RF measurements along predefined routes. Unlike MDT, drive tests capture scanner data (all cells visible, not just serving), layer-3 messages (RRC, NAS), and application-level metrics (throughput, latency, MOS). The tradeoff: expensive ($500-2000/day) but ground truth for model validation.

Performance Management (PM) counters are the most widely used data source in telecom AI. Every eNB/gNB reports hundreds of counters every 15 minutes (configurable to 5 min): RRC.ConnEstab.Att, ERAB.EstabInit.Succ, HO.ExeSucc, PRB.Used.DL.Avg, DL.THP.Time. These counters feed KPI calculations (HOSR, Drop Rate, CSSR, Throughput) and are the primary input for most ML models.

UE traces capture every Layer 3 signaling message for individual calls or sessions: RRC Setup/Release, Measurement Reports, Handover Commands, NAS Attach/Detach. Unlike PM counters (aggregated per cell), traces are per-subscriber and per-call, enabling precise root cause analysis. A single problematic call generates 50-200 L3 messages. Vendors: Huawei CHR (Call History Record), Ericsson CTR, Nokia CellTrace.

The OSS (Operations Support System) contains the network's configuration database: cell parameters, neighbor relations, antenna settings, feature activations. The BSS (Business Support System) holds subscriber data, service plans, billing. For ML, OSS provides the configuration context that explains why a cell behaves the way it does. A cell with downtilt=8 and power=43dBm behaves differently from downtilt=4 and power=46dBm — OSS tells you which is which.

CDRs were originally designed for billing but are now a goldmine for AI. Every voice call, data session, and SMS generates a CDR containing: start/end time, duration, serving cell, data volume, release cause, QoS class, and IMSI/MSISDN (anonymized for ML). CDRs enable subscriber-level analytics: churn prediction, usage pattern clustering, and service quality scoring.

Alarms are the network's cry for help. Defined in 3GPP TS 32.111, they come in four severity levels: Critical, Major, Minor, Warning. A typical network generates 10K-100K alarms per day. The challenge for ML: alarm storms. A single fiber cut can generate 500+ correlated alarms across multiple cells. AI alarm correlation reduces noise by 80-90%, surfacing only root-cause alarms.

DPI probes sit at strategic points in the network (Gn/S1-U, SGi/N6) and classify traffic by application: YouTube, Netflix, WhatsApp calls, gaming, OS updates. This gives operators visibility into what is consuming bandwidth, not just how much. For ML, DPI enables traffic-type-aware capacity planning, QoE prediction (video buffering ratio, voice MOS), and application-specific optimization.