ML Applications in Laser Welding: Quality, Defects, and Process Control
ML in laser welding is applied across four areas: process optimisation, quality monitoring and defect detection, parameter optimisation, and real-time process control. Each area uses different models suited to its data type and response time requirements.
ML in laser welding is not a single application. It is applied across four distinct areas, each with different data, different models, and different latency requirements.
Process Optimisation
SVM is the model of choice when labelled data is scarce, a common situation in industrial welding where generating labelled defect data requires destructive testing. SVMs generalise well from small datasets.
CatBoost on time-series is used when process signal data is available in sequence and the relationships between variables evolve over time.
Quality Monitoring and Defect Detection
ANN (Artificial Neural Networks) for modelling or predicting weld geometry from process parameters. Given power, speed, and focus position, the model predicts the resulting pool width, penetration depth, or HAZ extent.
Deep learning models for high-dimensional sensor data:
- CNN: applied to high-speed camera images and thermal imaging frames for detecting spatter, porosity, or geometric anomalies in the weld pool
- RNN / LSTM: applied to time-series signals (acoustic, thermal, electrical) where the temporal sequence carries diagnostic information
- Thermal imaging sequences: treated as video data, processed with CNN or CNN-LSTM hybrid architectures
Parameter Optimisation
Random Forest for optimising process parameters. Given a target weld quality specification, RF models map the input parameter space (power, speed, focus) to quality outcomes and identify the parameter combination that achieves the target.
Process Control
Reinforcement learning and adaptive parameter tuning for closed-loop control during the weld. The system measures the current state of the process through sensors, computes a deviation from the target state, and adjusts laser power, welding speed, or focus position in real time.
This is the most demanding application in terms of inference speed and hardware integration. The control loop must close within the timescale of the process (which can be milliseconds) and communicate with PLC systems that govern the physical equipment.