Network Alarm & Fault Analytics · Wisdom notebook
Ask the follow-up.
Agentic conversational BI · from the alarm storm to the band and sector
For the top sites by alarm volume this week, what is the category breakdown and how much of it is really one fault?
RKDe-duplicated alarms per element per window, clustered by signature, then compared raw counts against distinct faults for the ten busiest sites.
Volume and fault count are almost unrelated — the busiest site is not the worst problem:
Raw alarms by site — and how many faults they really are
Chicago and Kansas City have many small independent faults and no subscriber impact — the classic noise profile. The New York sites have one fault each and real impact.
Which of those actually correlate with customer complaints?
RKJoined the alarm clusters to care contacts and 5G fallback on the same districts and windows, keeping only correlations that hold.
Three sites, and the correlation is tight enough to prioritise on:
The three New York sites sit together in the corner. The two noisy sites sit with the quiet ones, which is the point.
Is it whole sites, or one band? Break it by frequency band and sector.
RKSplit the 19 affected sites by frequency band and sector, and checked availability per carrier rather than per site, so a single-carrier fault separates from a site-wide one.
One band, and not every sector — n78 only, which is why the sites never went down and the fault stayed invisible to availability monitoring:
Availability by frequency band × sector · the 19 affected sites
Site availability stays green because 4G and n1 are carrying the traffic — the degradation is confined to n78 on two of three sectors. That is exactly the shape that produces 5G→4G fallback and speed complaints while every site-level dashboard reads normal.