Challenges of ML in Industrial Laser Welding
ML in laser welding faces three hard problems: data availability and quality across different machines and materials, model generalisation and compute cost, and the real-time inference constraints imposed by the welding process itself.
Applying ML to laser welding in a production environment is not straightforward. The challenges fall into three categories: data, model, and system integration.
Data Availability and Quality
Labelled training data in welding is expensive to generate. Producing a defective weld for labelling purposes requires controlled degradation of an otherwise functional process. Destructive testing (cutting the weld open to verify penetration depth or porosity) is the only way to confirm ground truth for some quality measures.
The problem compounds across variation sources:
- Different machines: the same nominal process on two different welding systems produces different sensor signals due to hardware variation
- Different materials: aluminium and stainless steel have different thermal conductivity, reflectivity, and melting points, which changes the relationship between process parameters and weld quality entirely
- Different sensors: a model trained on data from one sensor configuration may not transfer to a different sensor setup
This is why domain adaptation and transfer learning are active research topics in this area: the goal is to reduce the labelled data requirement when moving to a new machine, material, or sensor setup.
Model Generalisation and Compute
ML models trained on data from one welding configuration tend not to generalise to others. A model that predicts defect likelihood for aluminium butt welds on machine A will perform poorly on stainless steel lap welds on machine B, even if the process parameters overlap.
Deep learning models that handle image and time-series data require significant compute for both training and inference. In a production setting, training time is manageable offline, but inference must happen fast enough to be useful.
Real-Time Inference and System Integration
Laser welding processes range from milliseconds (pulsed micro-welding in electronics) to minutes (thick plate structural welds). For online control applications, the ML inference must complete within the timescale of the process.
This creates two practical constraints:
- Hardware limitations: edge inference hardware must be fast enough, low-latency enough, and reliable enough for an industrial environment
- PLC integration: the ML output must interface with the Programmable Logic Controllers that actually drive the welding equipment. PLCs operate on deterministic real-time operating systems with strict timing requirements. Feeding ML predictions into a PLC control loop requires careful integration and latency guarantees that are not native to most ML frameworks
These integration challenges are often the bottleneck between a working research prototype and a deployed production system.