Application of Machine Learning and AI in Industrial Welding

This case study is a brief introduction to some of the solutions we are engineering in Datapao to generate hard savings for our clients by eliminating Non-Value-Added activities such as manual quality checks and improving the overall productivity.

We are going to focus on a specific manufacturing method called laser welding. It is a kind of joining process which brings smaller metal parts to create a larger one and, quite common in the industrial environment. We could even say that the entire global manufacturing is depending on the efficiency of welding. Here is the big picture to make sure everyone is on board and knows where we are before we start:


Two factory workers are performing Arc Welding to fix a metallic part to the main body. Just by looking, we can say that this type of joining process is a bit slow and outdated although it still has its own special use cases.

Laser welding, on the other hand, has deep penetration, minimal distortion by the heat and much faster than the other methods. It is also economical compared to the electron beam welding. It can become quite practical when attached to a robotic arm to automate the entire process:

Lack of penetration and the spatters are the main defect types and we have three root-causes; heat input (optimal, low, high), mismatching parts (loading), and nozzle (laser-head) deviation.

With an automated laser welding system, we will still have defects. However, we don’t want to check them manually after the production simply because it is time-consuming and prone to errors due to human interaction. What we need is a live monitoring system for quality assurance. We will briefly review how we do it step by step below.

We build AI and Machine Learning models running mainly with optical emissions data and visual sensing. The objectives are to detect anomalies accurately and to recognize the patterns: If the model is stable, events can be predicted.

We mentioned the significance of the heat input as a root cause of the defects. The heat is directly correlated with the laser reflection (optical emission). This fact helps us to use the laser beam not only for welding but also for checking the quality. First, the reflected photons are captured by the low-cost photodiode sensors which are located in the laser head. And then the AI & Machine Learning algorithms help us to learn the patterns of bad and good quality of welds with the help of the domain experts.


Here the accuracy will improve as more data flows in to teach, and then we would have smarter reference points to compare real-time in our live system. Essentially the system needs to be more accurate than the baseline or an average employee.

The visual sensing concept is quite straight forward thanks to high-quality cameras. We process the images of the metal vapor plume and spatters captured during the welding. Here the key is to extract features we can use in our machine learning models such as the centroid and height of the plume, the area, perimeter, centroid, average grayscale value and quantity of the spatters.

The challenges associated with these data-driven methods are mainly; not directly monitoring weld defects but acquiring intermediate signals (laser reflection) that can be correlated to the weld defects and, having noisy data because of the dust, external lights etc. during the data collection.

We can depict all we have covered in below architecture we have created in Datapao based on this work.

The framework is designed to leverage the cutting edge tools, algorithms, and libraries for end-to-end Machine Learning capabilities. These tools include:

Apache Spark, Databricks, sci-kit learn, NumPy, SciPy, XGBoost, Support Vector Machines as well as Deep Learning with Keras and H2O.

From what we have reviewed above, it is safe to conclude that the optical sensing is considered the ideal real-time monitoring technology for laser welding. Photodiode sensor, with advantages of low cost and simple structure, provides rich information of high-frequency features. This makes it adaptable to industrial manufacturing. The visual camera gives a great number of spatial information and has high accuracy in detecting weld defects on weld seam surface.

Meanwhile, the extracted features can also be used for establishing an adaptive control system during the welding process to make a real-time adjustment of welding parameters so as to prevent weld defects.

Nevertheless, we should note that the real-time detecting and controlling technologies during laser welding are far from perfect, and we expect the accuracy for different welding statuses and defects to be improved. So far, there is not a compound mode of multiple sensors applicable to detect all kinds of welding statuses and defects. And there are methods we will cover in our next articles.


Resources Used in This Case Study:

  1. Review of laser welding monitoring: Deyong You et al.
  2. Real-time monitoring of laser welding of galvanized high strength steel in lap joint configuration: Junjie Ma et al.
  3. Real-time estimation of CO2 laser weld quality for the automotive industry: Young Whan Park et al.
  4. A framework for physics-driven in-process monitoring of penetration and interface width in laser overlap welding: Erkan Ozkat et al.
  5. Real-Time Error Detection in metallic arc welding process using artificial neural networks: Prashant Sharma et al.