PulsCar analysis of NHTSA public data, September 2026
Topic: what do drivers' own noise reports actually cover - which vehicle systems,
       and which words - measured against what the sound-diagnosis tools claim to cover.
License: CC BY 4.0 (attribution: "PulsCar analysis of NHTSA public data, September 2026")

SOURCES
- NHTSA Office of Defects Investigation flat file COMPLAINTS (FLAT_CMPL), field
  dictionary https://static.nhtsa.gov/odi/ffdd/cmpl/CMPL.txt
  Rows loaded: 777,040 owner complaints received 1 January 2015 - 8 September 2026.
  Downloaded and indexed 17 September 2026; queries run 18 September 2026.
- The COMPONENT field is NHTSA's own categorization of the complaint, not ours.
  One complaint can carry several components separated by a pipe; each is counted
  once, so the component shares sum to more than 100%.

CASE SENSITIVITY
  Complaint narratives are ALL CAPS for complaints received through about 2020 and
  mixed case from 2021 on. Every query here is case-insensitive: SQLite FTS5 is
  case-insensitive, and every Python regular expression carries re.IGNORECASE. A
  capitals-only regex without that flag, or an uppercase SQLite GLOB, silently drops
  most post-2020 matches - on the noise set used here it would have returned 31,409
  complaints instead of 55,169 and kept only 2,008 of the 26,709 post-2020 ones.
  SQLite LIKE is case-insensitive for ASCII and does not have this failure mode.

FILE 1 - noise-complaints-by-component.csv
  Set: the same 55,169 noise complaints used by the recording guide on this blog.
  FTS5 over the narrative for noise, noises, noisy, grinding, grind, squeal(ing),
  squeak(ing), knock(ing), rattle/rattling, clunk(ing), whine/whining, hum(ming),
  tick(ing), screech(ing), growl(ing), chirp(ing), groan(ing) -> 57,204, then a
  negation test (a noise word counts only if it is not preceded within 40 characters
  by no / not / never / without / n't / any) -> 55,169.
  For each complaint, every NHTSA component label attached to it was counted.
  Read this as "which systems do complaints that mention a noise belong to", not as
  "which systems make noise": the component is assigned to the whole complaint, and a
  complaint about an airbag can also mention a rattle.

FILE 2 - owner-noise-vocabulary.csv
  Counted over ALL 777,040 complaints, not only the noise set, so a complaint that
  says only "clicking" is included. Word forms are explicit and whole-word:
  knock/knocks/knocking/knocked, grind/grinds/grinding, rattle/rattles/rattling,
  clunk/clunks/clunking, click/clicks/clicking/clicked, tick/ticks/ticking,
  pop/pops/popping/popped, whine/whines/whining/whiny, squeak/squeaks/squeaking/
  squeaky, squeal/squeals/squealing, hum/hums/humming, screech/screeches/screeching,
  buzz/buzzes/buzzing, thump/thumps/thumping, hiss/hisses/hissing, whistle/whistles/
  whistling, chirp/chirps/chirping, growl/growls/growling, roar/roars/roaring,
  groan/groans/groaning.
  Stem patterns were deliberately NOT used: "\btick\w*" also matches "ticket" and
  "\bhum\w*" also matches "human".
  Two filters, both measured:
   1. negation, as above;
   2. sound context - the match must sit within 100 characters of one of
      noise / noises / noisy / sound / sounds / sounded / sounding / hear / hears /
      heard / hearing / loud / audible.
  The CSV publishes the count after both filters and, in the last column, the count
  after filter 1 only, so the cost of filter 2 is visible. It is largest exactly
  where it should be: "popping" falls from 13,639 to 5,963, because "a message
  popped up" and "a notice pops up" are not sounds.

SAMPLE READING (required before publishing any keyword count)
  12 random matches were read for each of the four riskiest words, before and after
  the sound-context filter.
  - Before the filter, "popping" was roughly half UI language ("warning message
    popped up", "shift to park error randomly pops up"); "ticking" contained the
    metaphor "ticking time bomb"; "clicking" and "knocking" were clean.
  - After the filter, 30 matches drawn across the table read as genuine descriptions
    of a sound in 28 cases. The two exceptions were both "popping": one console
    pop-up that happened to sit near the word "beep", and one rear-window shatter.
    Observed false-positive rate after filtering: about 1 in 10 for "popping" and
    under 5% for the rest, on small samples.
  - A further caution specific to "popping": many of the remaining matches are
    one-off events - a tire blowout, a rear window shattering - rather than the
    repeating noises a diagnostic tool is meant to classify. The count is published
    because it is what owners wrote, not because all 5,963 are diagnosable noises.

LIMITATIONS
- Complaints are unverified, self-selected owner reports. They are not failure rates
  and are not normalized to vehicles on the road. No ranking of makes or models can
  be derived from them, and none is published here.
- The vocabulary table measures the words owners chose, not the acoustic reality. Two
  owners can call the same sound a click and a tick.
- The component table reflects NHTSA's categorization choices, including the large
  "UNKNOWN OR OTHER" bucket.
- No personal data is published: no VIN, city, dealer name, or free text beyond the
  short snippets quoted in this note.