Acoustic monitoring of laser welding: what the sound of a weld tells you

Acoustic monitoring of laser welding sounds almost too simple. Put a microphone near the weld, record, and listen for the bad ones. In practice the method works, but only if you understand where the sound comes from, which part of the spectrum carries the information, and how to turn a noisy signal into a quality grade you can defend.

I work on the data and model side of this at NeuroControls. This post summarises what the peer-reviewed research shows, what the signal does not tell you, and how we approach training and validation.

Where the sound of a laser weld comes from

In deep penetration laser welding, the beam vaporises metal and opens a narrow cavity, the keyhole, surrounded by molten metal. The keyhole is not static. Vapour pressure, surface tension and melt flow make it oscillate, partially collapse and reopen. The melt pool and the escaping vapour plume move with it.

These pressure fluctuations radiate sound in two ways. Some of it travels through the air as airborne sound, from the audible range well into ultrasound. Some travels through the workpiece as structure-borne sound, usually called acoustic emission (AE) when measured at high frequencies with a contact sensor.

When the process changes regime, the sound changes. A stable keyhole produces a different spectrum from conduction welding, where the beam only melts the surface, and from an unstable keyhole that collapses and traps gas as pores.

Airborne sound and acoustic emission compared

The two sensing routes have different strengths. A review of acoustic methods in welding monitoring by Yusof, Ishak and Ghazali highlights the airborne method for its simple, non-contact and low-cost measurement system, and also notes a research gap: how the emitted sound changes with parameters, materials and surrounding noise is still not well documented.

Fraunhofer IDMT has worked on airborne-sound-based monitoring for MAG, laser, friction stir and stud welding. For MAG welding the institute reports detection of process irregularities with up to 95 % accuracy in feasibility tests, which shows that airborne sound carries usable information even in loud processes.

What the signal carries: evidence from the research

Energy input and gap

The most direct study for battery production I know of is by Basile et al. (2025), who welded 1 mm aluminium 1050 busbar-to-terminal joints with a remote laser and recorded airborne sound with a membrane-free optical microphone sampled at 2.5 MHz. Their main findings:

• sound power correlated with laser power with a coefficient of 0.95 at zero gap, and filtered sound power with a coefficient of 0.82 across all gaps
• sound power correlated negatively with part-to-part gap (-0.797)
• a spectral peak around 63 kHz was linked to stable keyhole formation
• two thresholds on filtered sound power separated lack of fusion, lack of connection and sound connections in their test set

They removed a band from 8 to 26 kHz with a band-stop filter before computing sound power, to suppress background noise. That detail matters: much of the useful information sat above the audible range.

The same paper summarises related work. Authier et al., combining OCT with a microphone, used a 40 to 90 kHz band for penetration depth. Omlor et al. found spatter-related features in the 10 to 800 kHz range in hairpin welding.

Keyhole events and pores

Shevchik et al. (2020), published in Scientific Reports, took a different route. They welded aluminium AA5005 with a pulsed fibre laser and recorded structure-borne acoustic emission at 10 MHz together with back-reflected laser light. As ground truth they used high-speed X-ray radiography at the ESRF synchrotron, which shows the keyhole and pores inside the metal.

From wavelet spectrograms of 2 ms signal windows, convolutional neural networks classified five categories: conduction welding, stable keyhole, unstable keyhole, blowout and pores. Classification results ranged from 71 % to 99 % across the categories, sensor combinations and network types they tested. Two observations from the paper are worth remembering. Surface-based monitoring can be unreliable because keyhole irregularities happen deep in the melt pool. And pores sometimes merged into a following keyhole and disappeared, which complicates labelling.

What the sound does not tell you

Acoustic monitoring has clear limits, and it is better to know them before a trial than after.

• Background noise. Fume extraction, cross-jets, robots and neighbouring stations all make sound. Basile et al. note that industrial noise and cross-jets remained untested in their set-up.
• Transfer path. Distance, angle and fixture design change the recorded spectrum. A model trained at one sensor position may not transfer to another.
• Indirectness. Sound is a consequence of the process, not a measurement of the joint. It correlates with penetration and gap; it does not measure them the way OCT measures depth.
• Joint-specific training. A model trained on aluminium busbars does not know what a good copper tab weld sounds like.

This is one reason to combine sound with other modalities, such as images of the seam and process light, as discussed in laser welding of battery cells: defects and inline inspection.

From signal to grade: how the models work

The modelling pipeline for acoustic weld inspection has four stages.

1. Segmentation. Cut the recording into one segment per weld, using the laser trigger or the energy envelope.
2. Features or representations. Either compute engineered features (band energies, spectral peaks, envelope statistics) or feed a time-frequency representation, such as a spectrogram or wavelet transform, to a neural network. I covered the trade-offs in feature engineering for industrial signals.
3. Model. Start with anomaly detection trained on good welds, then add classification of known defect types once labelled examples exist, and finally severity grading.
4. Validation. Hold out entire days, batches or machines, not random segments.

The validation step is where most acoustic projects go wrong. If all bad welds in the training set were made on the same afternoon, a model can learn the afternoon, for example a change in extraction noise, instead of the defect. The post on avoiding shortcut learning in industrial AI goes into this in detail, and how much training data acoustic inspection needs covers dataset size.

Labels must come from the joint itself: cross-sections, peel or shear tests, or X-ray. A label that says "parameter set B" is not a quality label.

A practical recipe for a first acoustic trial

• Choose one joint type and one station.
• Mount the microphone at a fixed, documented distance and angle, protected from spatter.
• Record at a sample rate that covers the ultrasonic range you expect to need; a study band up to around 100 kHz is a reasonable start for keyhole-related features.
• Produce deliberate variations: laser power steps, defined gaps, contaminated surfaces.
• Cross-section or mechanically test a representative subset of every variation.
• Record factory background noise separately, with the laser off.
• Train, then validate on welds recorded on a different day.

How NeuroControls approaches this

Our Acoustic Node uses a Microtech Gefell MK 301 E measurement capsule, with a capture band from 2 to 100 kHz; the band from 2 to 20 kHz is production-proven. It is available as a standalone logger and built to order as a hub node. The measurement chain was developed in our research projects, including e-LAS+, a project with Fraunhofer IDMT on ML-based multimodal inline inspection of laser weld seams in battery cell production, which ran from May 2023 to April 2026, according to the Fraunhofer IDMT project page. A summary of that work is in our post on e-LAS+ results.

Models for customers are trained by our engineers on the customer's own weld data, from anomaly detection through to severity grading. On-device inference as ONNX on the node CPU was first proven on our Vibration Node; NPU deployment is coming soon. How we structure that model ladder is explained on our models page. Applications in cell and module production are described on the battery production page.