Catch every anomaly, before it leaves the line.
Catch every anomaly,before it leaves the line.
Vibration + near-field acoustics + order tracking + AI defect recognition, delivered at a takt tailored to your line, with raw signals piped to MES. Fits motors, gearboxes & reducers, pumps & compressors, bearings, appliances, fans and switches — anything that makes sound or vibration.
Why defects must be caught before shipping
Whine, wear, loose assembly, bearing and gear defects are invisible to the eye. Once they slip past EOL, a single-unit loss multiplies by three orders of magnitude into complaints, warranty and recall.
Field complaints & warranty
An in-vehicle whine triggers OEM warranty; per-unit cost can hit 3–10× sale price.
Latent batch recalls
Manual listening covers <30%; a bad batch surfaces all at once in the field.
Second-level takt
High-speed lines run ≤ 10 s per unit. Offline analysis is too slow — you need OK/NG online.
Closed-loop traceability
Bind NVH results to RFID and material batches — root cause in hours, not days.
Two walls that ears run into
Ears don't scale. Thresholds don't either.
Listening verdicts drift between people — and within one person across a shift. Skill is hard to quantify and harder to hand to the next shift intact. Even classical NVH spectral analysis hits two walls.
Fault diversity → threshold-matrix blow-up
Bearing spall, gear-tooth damage, imbalance, loose assembly, part friction, EM whine — each fault has its own acoustic signature. Presetting a rule and threshold per class grows an ever-larger matrix that still misses cases, and can't catch a fault it has never seen.
Fault weakness → signal buried in noise
Early-defect features hide under the normal running sound — like catching a distant whisper in a busy market. In an ordinary spectrum they sink below the noise floor, where even the most sensitive sensor may miss them.
How it works
Acquisition → order-domain analysis → AI verdict → self-learning limits — closed within a takt tailored to your line.
Multi-channel sync DAQ
2 / 4-channel tri-axial vibration; takt trigger + tachometer from the dyno bench, hardware-synced with <1 μs jitter.
Order-domain analysis
Angle-domain resampling pins speed-dependent faults (slot, eccentricity, bearing pass) to fixed order coordinates.
AI anomaly verdict
Turns the signal into a high-contrast time–frequency image, then a 1D-CNN / Transformer detects whine, wear, gear-tooth damage and bearing spalls on it — cross-checked against explainable rules. See «Acoustic imaging» below.
Self-learning limits
μ + k·σ on a rolling window of OK units adapts limits to process drift — no daily manual retuning.
Acoustic imaging · make the anomaly visible
See the sound. Box the defect.
Instead of presetting a threshold per fault, turn the sound into a picture. Each unit's running sound becomes a high-contrast time–frequency image: normal sound recedes into a dark background, anomalies light up by severity — obvious at a glance, and AI then boxes the time–frequency location and classifies it.
Aggregate the time–frequency energy of many OK units into a per-SKU «healthy signature» — a reference image of what good looks like.
Take the residual against the healthy signature; a purpose-built colormap pushes normal sound into a dark background and lights anomalies as high-contrast patches — brighter is worse.
The model boxes the anomaly's time span and frequency band, names the pattern — rub, shaft whine, grease, rotor burr — and reports severity and confidence.
▸ a row of short bars near 2 kHz
Switch fault / severity / AI detection
Anomalies read at a glance — verdict in seconds, fast enough for a high-speed line.
Color depth is severity; beyond pass/fail it names the defect type — traceable and countable.
Imaging + learning replaces the «threshold matrix» — no per-fault limit to hand-set.
One imaging-and-verdict core, normalized across SKUs — a model change needs no parameter reset.
▸ From «listen» to «see» — making abnormal sound visible is the first step from subjective skill to a digital standard.
Live waveform · spectrum · order map
Synthetic below — switch between healthy, bearing spall and EM whine. BPFO is derived from shaft speed; EM whine is modelled at a PWM carrier. Verdict tracks the live RMS and case.
Motor case: integrated electrical metrology
As one optional case, the full edition unfolds integrated electrical metrology — trimmed from the standard edition by default. Switch to «Full» to view.
Self-learning · drop-in SKU change
Swap the SKU, keep the line running.
Your line doesn't run a single product — it runs dozens of SKUs interleaved. The self-learning algorithm makes baselines, limits and verdicts «recipe-driven»: a SKU change or line conversion costs no retraining downtime.
RFID / MES reads SKU and process revision; triggers a hot recipe swap.
Each SKU carries its own recipe: order mask, k-multiplier, speed curve and limits.
Load this SKU's latest 30-unit (μᵢ, σᵢ) window — windows belong to a SKU, no cross-contamination.
On OK: Welford-update the SKU baseline on-line. On NG: verdict only, no learning.
Recipe-based baselines
Each SKU owns its (μᵢ, σᵢ) window and threshold k; RFID drives the swap, zero human in the loop. A new SKU never pollutes an existing one.
Cold-start transfer
For the first <30 units of a new SKU, bootstrap from a sibling SKU (same platform, different pole count) and tighten thresholds (k+1) for the first 50 units, then settle to nominal k.
Drift brake
Monitor mean-drift rate dμ/dN — freeze learning + alert above threshold. Prevents the «whole line slowly degrades → baseline slowly loosens → defects pass» failure mode.
Fleet sync
For the same SKU across multiple lines, run «leader / follower»: one line learns, the rest verdict-only and periodically sync the recipe — baselines stay aligned.
▸ SKU swap takes <200 ms (in-memory recipe switch). Operators only maintain the SKU × process-rev table in MES.
3D cutaway (motor example): where sensors actually sit
A motor shown as the example: EOL sensing is intentionally minimal — an accelerometer on the housing and a near-field mic above it. Speed and angle come from the dyno bench; no transducer is mounted on the unit under test. The same layout fits reducers, pumps and fans.
EOL inspection flow
Load to put-away at a takt tailored to your line — every unit gets a traceable record.
- 01
Load · RFID scan
Confirm SKU and recipe version.
- 02
Clamp & align
Servo clamp + concentricity, ±0.02 mm repeatability.
- 03
No-load sweep
0 → rated speed, capture the full band.
- 04
Loaded points
Dyno applies representative N–T operating points.
- 05
AI verdict → MES
Verdict + raw signals pushed back to MES.
- 06
Pass / Retest
NG routed to a retest station, no false-reject loss.
Applications
One platform, fixture swap per product — motors are just one of many.
Motors (EV / BLDC / stepper)
Hairpin / oil-cooled / 800 V to servo & stepper — OEM final & Tier-1 lines.
Reducers / gearboxes
Gear-mesh orders, tooth damage, pitting and bearing scuffing.
Pumps / compressors
Cavitation, imbalance and valve-plate noise in fluid machines.
Bearings / rotating parts
Race spall, rolling-element flaking and cage faults, located by pass frequency.
Appliances & parts
AC / washer / fridge units plus fan and blower parts — noise & imbalance.
Switches / buttons / relays
Contact-close sound, key feel and rattles — sound-quality & consistency verdict.
Contact
On-site demos, custom fixtures and solution scoping — call or add us on WeChat any time.
Building 1, 3rd Floor, No.16 Nanhe Road, Chendajiao Industrial Zone, Beijiao Town, Shunde District, Foshan, Guangdong, China
Douyin
Zhihu
CSDN Blog