The title of the presentation is Two stage quality monitoring of a laser welding process using

machine learning.

And there are five parts for the presentation, which are motivation, quality monitoring,
experiment, results&discussion and summary&outlook.

For the Motivation part, the speakers tell us some background information about Laser-welding
technique and their corresponding Advantages&challenges. Laser-welding technique is used in cases
like automotive industry, shipbuiding and aerospace industry. And they have advantages like precise,
high processing rate, complete automation and access to difficult parts of the workpiece. But it also
has challenges like high investment cost, very sensitive to small gaps between workpieces, faults like
humping or false friend occur, quality monitoring at smallcyclic times.

For the Quality monitoring part, the speakers tell some information about Sensors and evaluation
techniques. Typical sensors are photodiode, spectrometer and high speed camera(monitoring during
welding) and X-ray, microscope(monitoring after welding). And there are three systems for quality
monitoring. One is Multi Sensor System (MSS), which combines data of several sensors, has high
inference time and holistic and accurate evaluation. Another one is Cascaded System (CS), which is
the first evaluation on data of single sensor and aim at quick yet accurate evaluation. But remark
that if necessary, further sensor data is included. The last one is Single Sensor System (SSS), which
uses data of one of the Sensors and has quick evaluation, but is not sufficient for holistic evaluation.

For the Experiments part, the speakers tell some information about Setup and Data evaluation
approaches.

For the Hardware setup, Infrared Laser of 250W, Metal plates welded in overlap joint, Welding time
of 338 ms are used. And Photodiode (PD) and High Speed Camera (HSC) are used for monitoring. For
the Measured data, PD has a single intensity value as output, HSC has a 2D image as output and
faults can be detected and located. For the Data evaluation approaches, a general prediction model
structure is used. That means, features Block extracts most important data features(mean,
skewness, geometrical features, and etc.). Classification Block maps can process input data to a
quality-relevant quantity (Yi). Then blocks are realized with multiple Machine Learning algorithms.

For the SSS, SSS&PD formal definition can be achieved via the formular below.

fSSS,PD: ℝ1×13 → {0,1}: XPD,i ↦ Yi

And SSS,HSC formal definition can also be achieved using the formular below.

fSSS,HSC: ℝ100×100 → {0,1}: XHSC,i ↦ Yi

For the MSS, it combines PD and HSC data and its formal definition is shown below.

fMSS: ℝ1×13 × ℝ100×100 → {0,1}: XPD,i × XHSC,i ↦ Yi

For the CS, it is a two-stage quality assessment and combines two SSSin a cascaded way, consults
HSC data only if necessary and is expressed by the formular below.
𝑓𝑆𝑆𝑆,𝑃𝐷 𝑖𝑓 𝑝 < 𝑟 𝑜𝑟 𝑝 > 1 − 𝑟
𝑓𝐶𝑆 {
𝑓𝑆𝑆𝑆,𝐻𝑆𝐶 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝑝 ∈ [0, 1]
𝑟 ∈ (0, 0.5)
For the Results&discussion part, the speakers use the diagram below to help implemention of the
results.

Fig1. Performance of the approaches and their configurations with respect to the accuracy and the
inference time. The four approaches are marked by the four colors: yellow, red, blue and black. The
configurations are marked by the symbols. The gray lines indicate the performance of the CS
depending on the threshold r, where the CS-symbols indicate the maximum value on each line

As shown in the Fig.1, for Single-sensor systems photodiode, among the SSS,PD-configurations,
CNN1 has the highest accuracy of 89.84 %, whereas MLP has the lowest of 85.98 %. Out of the NN-
SSS, PD-configurations, CNN1 leads to the highest inference time of 0.700 ms and MLP to the lowest
of 0.456 ms. Among the DT-based configurations, DT2 performs 0.48 % better than DT1. However,
since in DT2 more features were extracted, the inference time increases by 0.542 ms. Comparing the
NN- and DT-based SSS,PD-configurations, the best DT-configuration DT2 performs by only 0.09 %
worse than the best NN-configuration CNN1 and can conclude that DT-based methods are
competitive with deep-learning-based methods.

For Single-sensor systems high-speed camera, among the NN-based SSS,HSC, RN50 achieves the
highest accuracy of 96.50 %. Comparing SSS,HSC with SSS,PD, all SSS,HSC lead to higher accuracies.
And the accuracy of the NN-based configurations increases when the model size increases as well.

For the Multi-sensor systems, MSS mostly results in higher accuracies than the corresponding SSS,
HSC and can conclude that the main information of the PD signals is included in the HSC images and
can be found by using complex processing algorithms. For the Cascaded systems, every CS is able to
outperform its corresponding SSS,HSC with respect to accuracy and inference time.

With the help of Fig.1, a suitable system can be chosen depend on production process conditions. If
the inference time cannot be longer than 1 ms, CS that combines CNN3 and CNN4 with an accuracy
of 92.92 % might be the best. If the inference time is limited to 2 ms, suitable solutions would be the
CS that combines CNN1 and MN with an accuracy of 94.40 % or the CS that combines CNN1 and
CNN4 with an accuracy of 94.44 %.

For the Summary&outlook part, the speakers say Cascaded system (CS) is as a two-stage quality
monitoring approach. CS are able to outperform SSS and is suitable for short inference time. Future
research could include CS with multiple sensors and is possible for extending quality monitoring
capabilities.