PulsCar analysis of NHTSA public data, September 2026 — "Strange car noises and what they mean"
License: CC BY 4.0. Credit "PulsCar analysis of NHTSA public data, September 2026" with a link to
https://www.pulscar.io/blog/strange-car-noises-and-what-they-mean

SOURCE FILES (downloaded from https://static.nhtsa.gov/odi/ffdd/ in September 2026)
- Complaints: COMPLAINTS_RECEIVED_2015-2019.zip, 2020-2024.zip, 2025-2026.zip. Unit = one ODINO
  (one consumer report). Received dates 1 Jan 2015 to 8 Sep 2026. 777,040 unique vehicle complaints.
  Field dictionary: https://static.nhtsa.gov/odi/ffdd/cmpl/CMPL.txt
- Manufacturer communications (TSBs): TSBS_RECEIVED_2015-2019, 2020-2024, 2025-2026. One row per
  make/model/year; documents counted by DISTINCT NHTSA document id. 187,827 documents, 6 Jan 2015 to
  12 Sep 2026.
- Recalls: FLAT_RCL_POST_2010, campaigns reported 2010 to 9 Sep 2026, 15,241 distinct campaigns.
- Investigations: FLAT_INV.

1. SOUND WORDS. Case-insensitive Python regular expressions on the complaint free text (CDESCR),
   applied AFTER the idioms in section 2 were blanked out. SQLite FTS5 was used only to pre-filter.
   knock  \bknock(s|ing|ed)?\b        click  \bclick(s|ing)\b        grind  \bgrind(s|ing)\b
   squeal \bsqueal(s|ing)?\b          whine  \bwhin(e|es|ing)\b      tick   \bticking\b
   rattle \brattl(e|es|ing)\b         hum    \bhumming\b             clunk  \bclunk(s|ing|ed)?\b
   shake  \b(shak(e|es|ing)|vibrat(e|es|ing|ion|ions)|shudder(s|ing)?)\b   (a feel, not a sound; kept
          because the article's tenth item is shaking)
   A complaint is counted once per word. A complaint can carry several words, so the words do not
   sum to a number of complaints.

2. IDIOMS blanked before matching (not a mechanical sound):
   ticking (time|flame)[- ]?bomb | click(s|ing|ed)? (on|the|a) (button|lock|unlock|link|key|remote) |
   (pump|nozzle|handle)s? (keeps |kept |that )?click | key ?fob | clicks? off | clicking off |
   knock(s|ed|ing)? (\w+ ){0,3}(unconscious|out|over|down|off|loose)\b | knock(s|ed|ing)? (me|him|her|them|us|it)\b |
   knock sensor | knock on wood | \bno (funny |strange |weird |unusual |abnormal |odd |loud )?(noises?|sounds?)\b
   Effect: tick 2,802 raw FTS matches -> 2,524 after idioms; click 7,555 -> 5,787; knock 10,052 -> 8,371.

3. HAND AUDIT (30 random matches per word read in full, two rounds; the second round after the
   idiom list was extended):
   knock 29/30 describe a mechanical sound (1 was the "knock detection system");
   click 27/30 (turn-signal click, seat belt "not clicking", "no clicking from the ignition");
   grind 27/30 (a sliding door, a pedal that "felt grinding", grinding at the moment of impact);
   squeal 27/30 (3 were squealing tires under hard acceleration);
   whine 30/30; tick 30/30 after the idiom fix (before it, 6/30 were "ticking time bomb");
   rattle 30/30 (includes cabin, sunroof and trim rattles — still a rattle);
   hum 30/30; clunk 30/30; shake 30/30 (a vibration or shudder is described in every one).
   We did not correct the counts for these residual false positives; treat each count as an upper
   bound with roughly the error rate above.

4. FLAGS. towed, crash and fire are the owner's yes/no answers on the complaint form (fields
   TOWED, CRASH, FIRE). Baseline over all 777,040 complaints: towed 39,739 (5.1%), crash 33,940
   (4.4%), fire 14,121 (1.8%). Percentages are shares of complaints containing the word. They are
   co-occurrences in the same report, not causes.

5. MILEAGE. Field MILES, kept only when 0 < MILES < 400,000. Quartiles on the kept values; the
   "with_mileage" column is the denominator.

6. SHOP COULD NOT FIND (regex on the same text):
   (could ?n[o']?t|can ?n[o']?t|could not|cannot|unable to|were not able to|was not able to|failed to|
   did not|didn'?t) (duplicate|replicate|reproduce|find (anything|any problem|the problem|a problem|
   the issue|an issue|the cause|the source|the noise)|locate (the )?(problem|issue|noise|source)|
   hear (the|any|it)|verify) | no (problem|trouble|issue|fault)s? (was |were )?found | nothing (was )?
   (found|wrong) | could not be (duplicated|replicated|reproduced|determined|found) | operating as designed |
   normal (operation|characteristic) | is normal
   NO CODE OR LIGHT MENTIONED:
   \bno (trouble |error |fault |dtc |stored |check engine |engine |warning )?(codes?|lights?)\b |
   no check engine light | check engine light (did not|didn'?t|never|has not|hasn'?t) (come|came|turn|illuminate|light)
   Absence of a mention is not proof that no light came on.

7. COMPONENT SYSTEM. COMPDESC top-level system before the colon; a complaint can carry several
   systems and is counted once per system, so percentages within a word can exceed 100 in total.

8. WHAT OWNERS SAY WAS REPLACED (what_owners_say_was_replaced.csv). Among complaints containing the
   word AND a repair word (replac*, rebuil*, repair*, fix*, install*, new), a part category counts
   when its pattern occurs within 100 characters of a repair word. Categories and patterns:
   engine replaced/rebuilt; wheel bearing/hub (\b(wheel )?bearing — NOTE this also matches "rod
   bearing" and "crank bearing", which is what it mostly is under knock); struts/shocks; sway bar
   links/bushings; control arm/ball joint/tie rod; CV axle/driveshaft; brake pads/rotors/calipers;
   transmission/torque converter/CVT; steering rack/column/pump; lifters/rockers/camshaft/cam phaser;
   timing chain/belt/tensioner; serpentine belt/pulley/idler; alternator/starter/battery; water
   pump/cooling; heat shield/exhaust/catalytic/manifold; motor or transmission mount; spark plugs/
   coils/injectors; differential/transfer case/driveline; tires/wheel balance/alignment; software
   reflash/reprogram; low oil/oil level. Hand audit of 120 tagged snippets (12 per word): the tag
   named a part that the text says was replaced, diagnosed or needed in about two thirds of cases;
   the rest name a part mentioned near a repair word for another reason (a recall, a prior repair,
   a part the owner suspects). Read these as "parts named next to a repair", not as verified fixes.
   The count is a floor: a complaint that says only "the dealer fixed it" is not tagged.

9. TSB COUNTS (tsb_noise_documents_by_make.csv). Distinct documents whose NHTSA summary matches the
   FTS5 query: noise OR rattle OR rattling OR knock OR knocking OR squeal OR squealing OR whine OR
   whining OR clunk OR clunking OR ticking OR grinding OR humming OR clicking. 12,839 of 187,827
   documents (6.8%). Per-word document counts in the article use "<word> OR <word>s OR <word>ing".
   A document is a manufacturer communication received by NHTSA; it covers named models and does
   not measure how often anything fails.

10. RECALLS AND INVESTIGATIONS. Recall campaigns whose defect description contains noise, rattl,
    knock, grind or clunk: 163 distinct campaigns of 15,241 (2010 to 9 Sep 2026). Investigations
    whose subject line contains noise, rattl, knock or grind: 1 (DP86017, a 1984 Mazda B-2000, 1986).

11. LIMITS. Complaints are self-reported and unverified; volume tracks fleet size and owner
    motivation, not failure rates; nothing here ranks makes by reliability. A make's share of a word
    is a share of complaints that use the word, nothing more. Counts change daily as NHTSA adds
    reports; ours are frozen at the 8 Sep 2026 file.