Case study · Survey operator

Four grounded aircraft, one bad batch of props.

A corridor-mapping team kept failing altitude hold on new airframes. Six weeks of intermittent aborts, no reproducible cause. The pattern was in the logs the whole time — it took a graded library and a flight-to-flight diff to see it.

Read the analysis Open the app
Outcome
14 days
F/DB
Affected aircraft grade
−71%
Peak Z vibration
4 / 4
Airframes returned
0
Aborts since
142 logs · 6 weeks · QUADROTOR ×4 · ArduPilot
Operator
Corridor mapping · 12-aircraft fleet
Airframe
7 kg survey quad, LiDAR payload
Symptom
Intermittent altitude-hold loss
Found in
One afternoon
01 / The analysis

From 142 logs to one part number.

01

Grade the library first

The team dropped six weeks of .bin files into the library. Grading ran locally on their own laptops — the payload configuration was under NDA, so nothing could be uploaded to a hosted tool. That constraint was why the logs had never been analysed properly in the first place.

Sorted by grade, the fleet split cleanly. Eight aircraft sat at B or better. Four sat at D or F. Not random — four specific tail numbers, every flight, for six weeks.

Library · sorted by grade
142 flights
F
TAIL-07
31 flights VIBE Z
F
TAIL-09
28 flights VIBE Z
D
TAIL-11
24 flights ALT HOLD
D
TAIL-12
19 flights ALT HOLD
B
TAIL-01 … TAIL-06, 08, 10
40 flights NOMINAL
02

Read what the recap already said

Every failing flight carried the same ranked anomaly at the top: Z-axis vibration over threshold, followed by altitude-hold drift a few seconds later. The debrief stated the relationship in one line — the drift was downstream of the vibration, not a separate control problem.

This mattered because the team had spent two weeks tuning the altitude controller. They were correcting a symptom.

Recap · TAIL-07 · 11 Apr
QUAD · 18:24
F
29/100
CRIT
Z-axis vibration over threshold
04:12–17:50 · peak 41.6 m/s²
WARN
Altitude-hold drift
04:18 onward · ±2.4 m

Sustained Z-axis vibration precedes every altitude excursion by 4–7 s. The barometer is reading airframe resonance, not altitude. Investigate rotor balance before retuning the altitude controller.

03

Diff a good aircraft against a bad one

Overlaying TAIL-07 against TAIL-03 — same corridor, same day, same wind — showed the two aircraft tracking together on the ground and diverging the moment the rotors loaded up. TAIL-07 sat three to four times higher on VibeZ for the whole cruise, then fell back in line the instant it descended.

Vibration that scales with rotor loading and disappears at idle points at the rotor rather than the mount. Per-arm current draw was even across all four motors, which ruled out a failing motor or ESC. The four aircraft shared exactly one thing: a propeller batch fitted during the same maintenance window.

Graphs · diff · VibeZ
TAIL-07 TAIL-03
00:00 · ground cruise · 41.6 vs 11.8 m/s² 18:24 · land
04

Confirm the fix with the same measurement

Props were swapped on all four aircraft and each flew the same corridor. The peak went with the props. Grades moved from F and D to B on the first flight back, and the altitude-hold warnings stopped appearing.

The controller tuning was reverted to fleet standard — those two weeks of changes had been compensating for a mechanical fault, and keeping them would have left four aircraft mistuned.

Before · original props
F
41.6 m/s²
Peak VibeZ
After · replacement batch
B
12.1 m/s²
Peak VibeZ
02 / What made the difference

Three capabilities did the work.

→ stayed inside

The logs could be read at all

An NDA payload ruled out every hosted tool. Parsing runs locally, and the AI debrief ran on the operator's own self-hosted deployment — so six weeks of evidence never left their network.

→ graded the fleet

The split was visible in one sort

One score per flight turned 142 logs into four suspect tail numbers without opening a single graph.

→ compared directly

Good against bad, same mission

Overlaying two aircraft showed vibration tracking rotor load — the detail that separated a prop fault from a mount fault.

“We had the data for six weeks. What we didn't have was a way to look at 142 logs at once and see that four aircraft were different.”

Flight operations lead · corridor mapping programme

Grade your own backlog.

Drop a folder of logs and sort by grade. Nothing uploads, so you can start with the aircraft you can't send anywhere.

Open the app →