Quality monitoring for ultrasonic welding: unlike continuous condition monitoring of turbines, this captures the in-process signal of each single weld (power/amplitude/frequency/displacement) and judges OK/NG with CNN plus a rule engine.
Weld Monitoring ≠ Turbine Monitoring: Two Kinds of Signals
Factories often ask: 'You do equipment monitoring — can you monitor welders too?' Yes — but first understand two fundamentally different signals:
| Turbine PHM (condition monitoring) | Weld quality (process monitoring) | |
|---|---|---|
| Signal nature | Continuous state (vibration/temp/RPM) | Single event (power/amplitude/frequency/displacement) |
| Time scale | Seconds-to-days continuous sampling | One weld, 50-200ms window |
| Object | Equipment health, trend degradation | Each weld's quality, judged instantly |
| Logic | Anomaly detection + life prediction | Curve-shape comparison + feature offsets |
Turbine monitoring watches 'equipment health'; weld monitoring watches 'was this weld good or not' — the latter must capture and judge an entire energy curve within tens of milliseconds. A very different problem.
The Pain Point: Why Weld Quality Is Hard to Manage
Typical issues in ultrasonic welding (tab welding, wire-harness welding, plastic welding):
- Too fast to see: one weld lasts 50-200ms — the human eye cannot follow, so only sampling is possible
- Sampling misses defects: destructive pull testing samples a small fraction; bad welds flow downstream
- No records to trace: when a problem surfaces, there is no per-weld record — nobody knows which batch, which machine, which weld
- Parameter drift: horn wear, air-pressure fluctuation — the energy curve changes silently until a batch fails
Solution: Four-Channel Process Signal, One Weld One Judgment
This solution captures power, amplitude, frequency and displacement (10kHz sampling) *while welding happens*, judging the quality of every single weld:
- Four-channel full capture: power curve (energy output), amplitude (ultrasonic vibration), frequency (resonance state), displacement (horn stroke) — four dimensions reconstructing the complete physical process of one weld
- CNN curve classification: a lightweight 1D-CNN (25KB, 0.3ms) infers in real time, learning the curve shapes of normal / cold weld / over-weld / insufficient power, judged instantly at the edge
- Explainable rule engine: complements the CNN by computing physical feature offsets — weld-time offset, peak-power offset, power-curve offset, first-ultrasonic offset — translating 'something is wrong' into concrete metrics an operator understands (e.g. 'power offset 78%')
- Unknown triple judgment: low-CNN-confidence welds enter a human review page with rule-engine signals and waveform visualization for fast triage
- Auto-labeling loop: high-confidence results are auto-labeled (conf > 0.95), accumulating training data — a self-improving loop of label → train → re-infer
Value: Every Weld Has a Record
| Aspect | Traditional | After Deployment |
|---|---|---|
| Coverage | 5%-10% sampling | 100% of welds |
| Judgment basis | Destructive tests + visual | Four-channel curves + CNN + rules |
| Defect discovery | After batch failure | Instant alert on drift |
| Traceability | No records | Full curve per weld, replayable |
| Improvement | Guesswork from experience | Feature-offset data drives improvement |
Reuse with Turbine PHM
Both solutions share the same edge inference platform (capture → infer → judge → review → label loop); only signal type, model input and judgment logic differ:
- Turbine PHM: vibration/temp/RPM → health/life (continuous state) | for equipment makers, operations
- Weld monitoring: power/amplitude/frequency/displacement → OK/NG (single event) | for battery/harness/plastic weld lines
Reuse the edge platform, lightly customize models and rules, and both scenarios are covered quickly — a real 'reuse + light customization' story.
Suitable for: EV battery tab welding, wire-harness terminal welding, plastic housing welding, 3C electronics assembly — any scenario demanding strict weld quality.
Every Weld, Replayable and Traceable
Ultrasonic weld quality should not rely on 'sampling luck'. Four-channel signals plus three-way fused judgment give every weld a full record, instant verdict and explainable basis.
Is your welding line still on sampling and visual checks? Contact us for a solution demo.
