Acoustic Emission for In-Process Weld Quality Monitoring: What the Research Actually Shows

Welding is one of the few manufacturing processes where the defect is created and hidden in the same second. By the time a seam reaches inspection, the gap that caused the porosity is closed and invisible. Research our founder led and published on acoustic monitoring of joint-gap formation in laser beam butt welding, set out to catch the event as it happens (Crystals 2023). This post explains the approach in plain terms and what it means for production lines.

The problem: gaps you cannot see afterwards

In laser butt welding, a small gap between the parts changes how the keyhole and melt pool behave. The result is a defective seam, but the gap itself is gone once the parts are joined. Existing monitoring approaches based on optical signals or single process parameters did not generalise across process settings: change the feed rate or laser power and the thresholds had to be re-tuned.

The approach: listen to the process

The welding process emits sound, and that sound changes when a gap forms. The team recorded airborne acoustic emission in the welding cell, transformed the signal into time-frequency representations (short-time Fourier transform features), and trained a neural network classifier to recognise gap formation. Two decisions made the difference.

• Training under process variability. Instead of a single fixed parameter set, the model saw data across the process window, with augmentation to simulate the variation it would meet in production. This is what let it generalise where threshold-based methods failed.
• A monitoring framework developed with the process engineers. The output was designed to answer the question a welding engineer asks, not to maximise a benchmark score. The work was published in the journal Crystals in 2023 (DOI 10.3390/cryst13101451), with a companion paper on the temporal resolution of the same process emissions in Applied Sciences (DOI 10.3390/app131810548). We will happily walk through the papers and their figures with anyone evaluating the method.

Why this transfers to production

Airborne sensing means no contact with the workpiece and no change to the fixture. The signal processing runs comfortably on an edge device at the cell. And because the model was trained across the process window, it does not need re-tuning every time the recipe changes. Those three properties are exactly what a production line needs and what most lab results lack.

The same pattern, acoustic signature plus a model trained under variability, is transferable to other joining and forming processes: resistance spot welding, ultrasonic welding of battery tabs, friction stir welding, and extrusion. The same founder-led group showed it for quality tracking in corn extrusion (INTER-NOISE 2024), which could not be further from laser welding in every respect except the physics of the signal.

What a plant project looks like

1. A feasibility study on a few hours of recordings from your cell, with known-good and known-bad seams, to confirm the signature is present in your acoustic environment.
2. A proof of concept with a proper measurement campaign, labels tied to your inspection results, and evaluation against your current monitoring.
3. A pilot cell with a multi-modal sensor unit and edge inference, integrated with the cell PLC so that a detected event can flag the part or stop the line.

We report performance numbers only from the published evaluation or from your own data, never from a brochure. If you would like the paper, ask us.

If any of this matches a problem on your line, the fastest way to find out what is possible is a free discovery call followed, where it makes sense, by a feasibility study of two to ten days.