Back Room — Data
The transparency layer. Everything the model knows, and how it knows it.
Data sources
| Source | Fields | Cadence | Status |
|---|---|---|---|
| PM weekly summary | 8 normalised KPIs per NE | Weekly | Active — last batch 3d ago |
| CM weekly snapshot | SW version, param fingerprint, drift score, 6 fields | Weekly | Active — last batch 3d ago |
| Inventory static | Vendor, model, age, EoL status, site class, region, lat/lng, 11 fields | Monthly | Active — updated 18d ago |
Vendor normalisation map
5 canonical KPIs mapped to per-vendor counter names. All KPIs scaled to 0–1 before NHI composition.
| Canonical KPI | Ericsson | Nokia | Huawei | Samsung |
|---|---|---|---|---|
| Availability % | pmCellDowntimeAuto (inv.) | RRC.ConnEstabAtt cluster | Cell_Available_Time / Period_Duration | CELL_AVAIL_RATE |
| RRC Setup Success % | pmRrcConnEstabSucc / pmRrcConnEstabAtt | RRC.ConnEstabSucc / RRC.ConnEstabAtt | RRC_CONN_SETUP_SUCC_RATE | RRC_SETUP_SR |
| DL Util % | pmPdcpVolDl + capacity est. | PDCP.UEThpDl / capacity | L_PDCP_DL_THROUGHPUT / cap | PDCP_DL_THPUT_UTIL |
| HO Success % | pmHoPrepSuccOuttraHo / pmHoPrepAttOuttraHo | HO.PrepSucc.IntraFreq / HO.PrepAtt | HHO_SUCC_RATE | HO_SR_INTRA |
| CQI Mean | pmCqiDistr (computed) | L.ChMeas.CQI.Distr (computed) | L_CQI_CQI_MEAN (direct) | CQI_MEAN (direct) |
NHI formula weights
NHI = 0.35·performance + 0.30·config + 0.20·lifecycle + 0.15·alarm
performance_score35%
PM KPI composite (avail, RRC, DL util, HO, CQI)
config_conformance30%
CM fingerprint vs golden baseline
lifecycle_risk20%
HW age, SW currency, EoL proximity
alarm_burden15%
Weighted active alarm count by severity
Weights calibrated quarterly against field incident rate. Last calibration: Q1 2026.
Archetype clustering inputs
22 fields used by k-means (k=6, weekly refit). Model confidence: 74%.
PM weekly summary
kpi_availability_pctkpi_rrc_setup_success_pctkpi_dl_util_pctkpi_ho_success_pctkpi_cqi_meankpi_dl_throughput_mbpskpi_ul_throughput_mbpskpi_active_alarm_count
CM weekly snapshot
sw_versionsw_release_age_daysparam_fingerprint_hashdrift_score_pctsrPeriodicity_valueprachConfigIndex_value
Inventory static
vendormodelne_typeinstall_datehw_age_monthshw_eol_datesw_eol_datesite_classregionlatlng
Model confidence
74%
Last calibration
Q1 2026 — field incident rate
Archetype model
k-means · k=6 · weekly refit