PULSCAR ANALYSIS OF NHTSA PUBLIC DATA — GRINDING NOISES
PulsCar analysis of NHTSA public data, September 2026. License: CC BY 4.0.
Files in this folder: grinding_complaints_summary.csv, grinding_complaints_by_system.csv,
grinding_timing_vs_system.csv, grinding_complaints_by_model.csv.

SOURCE FILES
1. NHTSA Office of Defects Investigation complaint flat file (COMPLAINTS_RECEIVED_*.txt),
   downloaded from https://static.nhtsa.gov/odi/ffdd/cmpl/ — field definitions:
   https://static.nhtsa.gov/odi/ffdd/cmpl/CMPL.txt
   Rows loaded: 777,040 complaints, received 1 January 2015 through 8 September 2026.
   One row per complaint (ODINO), de-duplicated on ODINO.
2. NHTSA manufacturer communications (TSB) flat file, FLAT_TSBS.txt — 187,827 distinct
   NHTSA document ids, one row per make/model/year per document.
3. NHTSA recall flat file, FLAT_RCL.txt — 245,578 rows, counted as DISTINCT campaign.
4. NHTSA investigation flat file, FLAT_INV.txt.
All four were loaded into a local SQLite database with FTS5 full-text indexes on the
complaint description and the bulletin summary. Queries below are the exact ones run.

QUERY 1 — THE GRINDING SET
  select c.* from complaints c join complaints_fts f on f.rowid = c.rowid
  where complaints_fts match 'grind OR grinds OR grinding OR grinded';
Result: 6,248 complaints.

EXCLUSIONS
Each match was re-read in Python. For every complaint we extracted every window of
+/-90 characters around every occurrence of /grind\w*/i. A complaint was DROPPED only if
ALL of its windows matched the machining-and-idiom regex:
  (resurfac|re-?surfac|machin|turn(ed|ing)?\s+the\s+rotors?|grind(ing)?\s+(down\s+)?(the\s+)?
   (rotors?|brake\s+rotors?|discs?)|grind(ing)?\s+to\s+a\s+halt|daily\s+grind|grind(ing)?\s+of\s+traffic)
That is, a complaint that says "the shop ground the rotors" and nothing else is dropped;
a complaint that says "it grinds when I brake, so they ground the rotors" is kept.
Dropped: 12. Kept: 6,236. This is the denominator for every percentage in the article
unless another denominator is named.

SAMPLE READING (false-positive rate)
Before publishing any count we drew 40 matches at random (Python random.seed(7),
random.sample over the 6,248 raw matches) and read the grinding sentence in each.
39 of 40 described a noise, a sensation or a gear clash the owner or a technician
experienced. 1 of 40 was a machining use ("Toyota's solution was to resurface/grind the
rotors", ODINO 10821007) — a 2.5% false-positive rate in the raw set. That single failure
mode is what the exclusion regex above targets. We did not tighten further: narrower
phrase queries such as '"grinding noise" OR "grinding sound"' lose real reports that say
"it grinds when I brake" or "the gears grind", and the residual error after exclusion is
smaller than the rounding we publish at.

QUERY 2 — SYSTEM WORDS (grinding_complaints_by_system.csv)
Case-insensitive regex over the FULL complaint text, one flag per family, NON-EXCLUSIVE:
a complaint that names both brakes and a bearing counts in both rows, so the shares add to
more than 100%. Patterns, verbatim:
  Brakes: pads, rotors, calipers            \bbrake|\bpads?\b|rotor|caliper
  Transmission, CVT, clutch or gears        transmission|\bcvt\b|clutch|gearbox|shift|\bgears?\b
  Wheel bearing or hub                      wheel bearing|hub assembl|\bhub\b|bearing
  Tires or wheels                           \btires?\b|\bwheels?\b|lug nut|rim\b
  Differential, axle, driveshaft, t-case    differential|\baxle|driveshaft|drive shaft|transfer case|\bcv joint|cv axle
  Starter or flywheel ring gear             starter|flywheel|ring gear|bendix|solenoid
  Steering: rack, pump or column            steering|\brack\b|power steering|column
  Engine internals or accessories           \bengine\b|timing chain|\bpiston|camshaft|valve train|valvetrain|water pump|alternator|serpentine|tensioner|idler|pulley
  ABS, traction control, automatic braking  \babs\b|traction control|automatic emergency|collision|\baeb\b|stability control
  Suspension: struts, springs, control arms strut|suspension|shock absorber|control arm|ball joint|sway bar
  A/C compressor, blower or HVAC            \ba/?c\b|air condition|compressor|blower|hvac
These are WORD counts, not diagnoses. A complaint lands in "Brakes" because the owner wrote
"brake", which may be the system that failed, the system they suspected, or the system a
shop replaced without fixing the noise.

QUERY 3 — WHEN IT GRINDS (grinding_timing_vs_system.csv)
Same +/-90-character windows around the grinding words (NOT the whole complaint, which
would catch any mention of braking anywhere in the text). Non-exclusive; a complaint that
says "it grinds when I brake and when I turn" counts twice. Patterns, verbatim:
  When braking                brak|\bpedal\b|stopp?ing|come to a stop|slow(ing)? down
  When turning or cornering   turn(ing|s|ed)?\b|corner|steer
  When starting the engine    start(ing|ed|s|up)?\b|crank|ignition|turn the key|push button
  When shifting or changing gear  shift|chang(e|ing) gear|\bgears?\b|reverse|\bpark\b|\bneutral\b|clutch
  When accelerating           acceler|throttle|gas pedal|speeding up|pulling away
  At highway speed            highway|freeway|interstate|\b[5-8]\d\s?mph|high speed
  At low speed or parking     low speed|parking lot|park(ing)? space|driveway|\b[1-2]?\d\s?mph|backing (up|out)|slow speed
Known weakness, stated in the article: "When starting the engine" also catches the very
common narrative opener "the noise started when...", so that row is an upper bound.
The cross-tab columns are the system-word families above, counted within each timing row.

QUERY 4 — MODELS (grinding_complaints_by_model.csv)
Top 30 make/model pairs by grinding-complaint count. The rate column is
grinding complaints / all complaints for THAT make and model in this file x 1,000. It is a
share of that model's complaint traffic, NOT a failure rate: there is no registration or
vehicles-in-service denominator in these files, and complaint volume follows recalls,
news coverage and fleet size. Model years shown are the top three by count.

MILEAGE, TOW, CRASH, FIRE
Mileage quartiles use only rows with miles > 0 and miles < 400,000 (3,805 of 6,236
grinding complaints carry a usable mileage; 392,935 of 777,040 file-wide). Tow, crash and
fire are the file's own Y/N flags, owner-reported.

"A SHOP COULD NOT FIND IT"
FTS5 phrase list, applied identically to the grinding set and to the whole file:
  "could not duplicate" OR "unable to duplicate" OR "could not replicate" OR
  "unable to replicate" OR "could not reproduce" OR "no problem found" OR
  "could not find the" OR "unable to find the" OR "nothing was found" OR
  "found nothing wrong" OR "no issue found" OR "no problems found"
Grinding: 107 of 6,236 (1.72%). Whole file: 7,337 of 777,040 (0.94%). This measures a
phrase, not an outcome: complaints that describe the same experience in other words are
not counted, so both numbers are floors and only their ratio is the finding.

QUERY 5 — WHAT THE OWNER SAYS WAS DONE (last ten rows of the summary CSV)
Case-insensitive regex over the full complaint text, non-exclusive, sentence-bounded by
[^.]{0,30-60}. Example, verbatim, for the pads row:
  (replac\w+|new)[^.]{0,30}\bpads?\b|\bpads?\b[^.]{0,30}(were |was )?(replac|chang)
and for the transmission row:
  (transmission|cvt)[^.]{0,60}(replac|rebuil|new)|(replac\w+|new|rebuil\w+)[^.]{0,30}(transmission|cvt)
The other eight families follow the same shape around rotors, caliper, wheel bearing/hub,
starter, differential/axle/driveshaft, engine, clutch and compressor. THIS IS NOT A FIX
RATE. It counts complaints that MENTION a part being replaced — including complaints where
the replacement did not cure the noise, and including parts replaced before the grinding
started. Most complaints are filed before any repair, so all ten rows are floors.

BULLETINS
  select distinct t.* from tsbs t join tsbs_fts f on f.rowid = t.rowid
  where tsbs_fts match 'grind OR grinding OR grinds';
502 distinct NHTSA document ids. "Software-ish" = the summary matches
software|reprogram|calibrat|flash|update the|pcm update|module update — 30 of 502 (6.0%).
Bulletin counts are per document, not per vehicle.

RECALLS AND INVESTIGATIONS
  select * from recalls where defect like '%grind%' or consequence like '%grind%'
  or notes like '%grind%';
4 distinct campaigns, of which 2 are industrial grinder trailers (Vermeer) and 2 are
passenger vehicles: 10V656000 and 11V244000. Investigations were searched the same way.

LIMITATIONS
Complaints are unverified owner reports. They are not a failure rate, not a sample of all
cars, and not a sample of all grinding noises — only of grinding noises somebody chose to
report to the federal government. Counts here must not be read as "this make grinds more
than that one." Word counts measure vocabulary. Nothing in this folder contains a VIN, a
city, a dealer name or any free text longer than a short quoted snippet.

CONTACT: support@pulscar.io