Bashifu/uav-fault-symptom-reports

UAV Fault Symptom Reports

A synthetic dataset of UAV flight telemetry paired with operator-style symptom reports written by a language model. Each row is one five-second window of a flight: 20 telemetry channels, the fault class, a severity derived from simulated consequences, and a one-sentence report.

split	rows	flights	model-written reports	unique reports
benchmark	10,500	2,100	82.0%	79.7%
challenge	3,500	700	88.1%	91.0%

benchmark is balanced across seven classes (six fault families plus normal). challenge is harder: mostly normal flights, random fault onset, and some flights with two faults.

How it was generated

1. A mechanistic simulator produces telemetry from sampled aircraft, battery, sensor and environment state. Fault events change those latent parameters, never the telemetry directly.
2. Labels are read from the event log afterwards, so the fault class never builds the telemetry.
3. Facts are derived from the telemetry against fixed thresholds.
4. Qwen/Qwen2.5-0.5B-Instruct rewrites those facts as an operator report, using a system prompt, few-shot examples, a JSON output format and explicit constraints.
5. Every report is validated: length, no numbers, no diagnosis, no actions, no inferred causes, no signals the model was not given, no chat-style text. Failures are retried twice, then replaced by a templated sentence built from the same facts.

The model never sees the fault class, the severity or the recommended action.

Exploratory data analysis

Instruction-following failures

82.0% of reports were written by the model and accepted. The rest failed validation three times. Rejection reasons across all attempts:

reason	rejections	share_of_attempts
Too short / too long	21821	0.275
Other	5936	0.075
Inferred a cause	5704	0.072
Linked observations causally	4094	0.052
Changed what an observation said	1131	0.014
Mentioned a signal it was not given	914	0.012
Ignored the given signals	824	0.01
Invented a number	309	0.004

Why model outputs were rejected

Model-written share per fault class

Hallucination

Accepted reports naming a signal they were not given: 0. Reports linking observations causally: 0. Reports changing what an observation says: 0. Earlier runs did contain both: at their worst, 869 reports in one run linked observations causally and 264 changed what an observation said. The validator was extended in response and every failing report was regenerated. 48.2% of accepted reports add no content word beyond their facts.

Inconsistency between text and label

Each report describes one five-second window, and never sees the label. 36.4% of fault windows show no threshold crossing and are described as steady; 15.8% of normal windows cross one during ordinary manoeuvres and are described as irregular. These rows are where the text alone cannot identify the fault.

How each class is described by its text

Reports in TF-IDF space, coloured by class

Duplication, and how the EDA changed the generation

The first full run had 42.8% duplicate reports. The cause was the input: the evidence builder drew from a small fixed pool of facts, which capped how many different reports were possible. Widening the evidence phrasings brought duplicates down to 20.3%. Counting reports that differ only in word order or function words as the same, the near-duplicate rate is 24.5%.

The same review found up to 69 benchmark rows per run whose minimum voltage exceeded their mean voltage, which is impossible. The simulator was corrected; only the affected rows changed.

Every generation run

Read by hand

70 model-written reports, ten per fault class, were read by the team against the facts each was given. This was done on run three, and the rules were tightened in response to what it found, so the counts below describe the run that was reviewed: 36 faithful, 25 uncertain, 7 changed the meaning of a fact, 2 invented a cause. Those three failure modes do not occur in this dataset.

Two habits were recorded rather than treated as errors: a one-sentence report sometimes omits one of several given facts, and sometimes generalises beyond the specific signals it was given.

The cost of stricter rules

Duplicate reports stand at 20.3%. Forbidding a report from inventing a cause, swapping a signal name or adding an intensity word leaves less room to paraphrase the same facts, so reports resemble each other more. Accuracy was preferred over variety, and both are measured here.

Text statistics

Words per report

Top content words and Zipf plot

Word cloud

Signal families mentioned together

Telemetry

Severity is scored per window, so it can escalate within a flight: 43.6% of flights change severity across their five windows, and 92.5% of those only escalate.

Telemetry correlations

Similarity between fault signatures

Persistence of fault evidence across windows

Flagged issues

issue	type	rows affected	share of rows	handling
Model failed its instructions and fell back to a template	instruction following	1885	0.18	Kept, labelled canonical_grounded_fallback; the sentence states the same facts
Accepted report names a signal it was not given	hallucination	0	0	Blocked by the validator before acceptance
Report links observations causally	hallucination	0	0	Blocked by the validator; up to 869 such reports in an earlier run were regenerated
Report changes what an observation says	hallucination	0	0	Blocked by the validator; up to 264 such reports in an earlier run were regenerated
Minimum voltage above mean voltage	telemetry consistency	0	0	Simulator fixed; up to 69 benchmark rows per run were affected before the fix
Fault window described as steady operation	inconsistency (text vs. label)	2785	0.265	Intended: the fault was not visible in that window. Flagged for the recommendation step
Normal window described as irregular	inconsistency (text vs. label)	201	0.019	Intended: ordinary manoeuvres cross thresholds. Flagged for the recommendation step
Exact duplicate report text	duplication	2128	0.203	Reduced from the first run by widening the evidence (Part F)
Near-duplicate report text (differs only in word order or function words)	duplication	2574	0.245	Flagged; each row still carries its own telemetry, but the text adds little on its own
Reports judged unfaithful when read by hand	manual review	9	0.001	Of 70 reports read from v5_production_v3; the rules were tightened in response and these patterns no longer occur
Chat residue in a report	instruction following	0	0	Blocked by the validator

Limitations

• The simulator is a transparent mechanistic prototype, not a validated flight-dynamics model.
• Severity is a declared academic rubric, not a certified safety scale.
• approved_mitigation and approved_inspection are fixed academic placeholders, one per fault family; they are not maintenance procedures.
• A report describes one five-second window and never sees the label, so it may omit one of several given facts or describe a fault window as steady when nothing crossed a threshold.

Files

• 01_synthetic_data_generation.ipynb - the generator
• 02_dataset_and_eda.ipynb - this analysis
• manifest.json - seed, model revision, row counts and SHA-256 of each file