AUTOMATED QUALITY ANALYSIS FOR LASER WELDING

This project was carried out as part of the TechLabs “Digital Shaper Program” in Aachen (Summer Term 2020).

PROJECT SUMMARY

As automobiles are increasingly being powered by electrical motors, there is a growing demand for high-performance energy storage systems. Various types of cells are used for the construction of battery modules or packs: round cells, prismatic cells, or pouch cells. The Laser Welding processes required to interconnect the cells require high process stability and reliability irrespective of the type of battery, while at the same time providing low electrical transition resistances and high geometric flexibility.

In addition, the metallurgy of the joining compound can be specifically controlled by this approach. The process is suitable both for the production of electrical and mechanical safe connections from cell to cell as well as for the production of modules from cell networks and large battery packs from modules.

To understand the effect of varying input process parameters on the weld-seam, the cross-section images of the weld-seam are to be inspected and results are to be recorded. We trained a Neural Net based on Mask R-CNN that segments the cross-section images, to extract the features such as length and width of the weld-seam, area of weld-seam, the porosity of the weld, distribution of the pores, and other relevant features. These results augment the process engineer’s understanding of the industrial welding process.

Figure 1 (A) Welding Cross-section

Figure 1 (c ) Feature Identi cation

INTRODUCTION

Mask R-CNN has been the new state of the art in terms of instance segmentation. Mask R-CNN is a deep neural network aimed to solve the instance segmentation problem in machine learning or computer vision. In other words, it can separate different objects in an image or a video. If you give it an image, it will give you the object bounding boxes, classes, and masks.

The primary objective of the project was to identify if Mask R-CNN is a suitable method to analyze Laser Weld processes in Industrial Machines. The application of Mask R-CNN to this problem aims at efficiently detecting the weld-seam as well as the pores in the cross-section image while simultaneously generating a high-quality segmentation mask for each instance.

The secondary objective was to reliably extract the features required for the further analysis of the industrial weld process and to automate it, which would otherwise require extended manual effort.

METHOD

The below-mentioned steps were followed:

1. Get to know the basics of the Laser Welding process.
2. Understand the basics of Mask R-CNN, adapting the Mask R-CNN repository for our project and the Coco Data-set.
3. Annotation of the raw data-set for training the neural network.
4. Creation of synthetic data-set to improve the performance of the neural network.
5. Train the Mask R-CNN.
6. Building a function that calculates the features from the results for further analysis.
7. Record and store the results in a data frame.
8. Analyse the weights of the trained model and calculate the Mean Average Precision (mAP) score to understand the accuracy of the model.
9. Correlate features calculated to the Process Parameters to check if a process corresponds to the expected outcome.

PROJECT RESULTS

1. We were able to successfully train the Mask R-CNN with a limited data-set and produce image segmentation results that identify the seam and pore in a weld cross-section image.

2. The following features were calculated based on the output of the Mask R-CNN:

2.1. Length — the vertical distance between the top and the bottom pixel of the segmented image (weld-seam), translated into microns based on pixel size

2.2. Width — the horizontal distance between the left and the right extreme pixel of the segmented image (weld-seam), translated into microns based on pixel size

2.3. Area of the seam — the sum of all the pixels identified as weld-seam

2.4. Area of the pore — the sum of all the pixels identified as pores

2.5. Porosity — the ratio of the area of pores to the area of weld-seam

2.6. Distribution of pores — identification of no.of pores in top, middle, and lower portion of the weld-seam.

The calculated features were compared against the manually recorded data and the features extracted had an error rate of +/- 2%.

3. We also developed an algorithm to create synthetic data-set which would help improve the accuracy of the Neural Network, in problems where the availability of good quality data-set is limited.

4. We achieved a Mean Average Precision Score of 57% which is a good accuracy score, considering the fact that we did not train our model on a large data-set.

In order to interpret the above results, the following terms have to be understood:

• The Average Precision (AP) is a popular metric in measuring the accuracy of the object detectors like Faster R-CNN, Mask R-CNN, etc., Average precision computes the average precision value for recall value over 0 to 1.
• Precision measures how accurate your predictions are. i.e. the percentage of your predictions are correct.
• Recall measures how good you find all the positives.

Here is the mathematical definition:

For example, in the testing for cancer:

IoU (Intersection over union):
IoU measures the overlap between 2 boundaries. We use that to measure how much our predicted boundary overlaps with the ground truth (the real object boundary). In some datasets, we predefine an IoU threshold (say 0.5) in classifying whether the prediction is a true positive or a false positive.

mAP — mAP is the average precision score over multiple IoU (the minimum IoU to consider a positive match). AP@[.5:.95] corresponds to the average AP for IoU from 0.5 to 0.95 with a step size of 0.05
The below graph depicts the mAP score of all the 60 epochs that were accomplished during the training of our model and applied on 10 random validation data-set. The graph indicates that the scores kept fluctuating after the 12th epoch and this could be resolved through hyper-parameter tuning/ training the model with a larger data-set.

5. The goal of the project was to automate the analysis of the welding cross-section images by using a state of the art image segmentation technique and the results serve as a proof of concept.

TEAM

Kim Nebe — Artificial Intelligence

Prasanth Sathiyamurthy — Artificial Intelligence

Swapnil Deole — Data Science

Marc Ungeheuer — Mentor

TechLabs Aachen e.V. reserves the right not to be responsible for the topicality, correctness, completeness or quality of the information provided. All references are made to the best of the authors’ knowledge and belief. If, contrary to expectation, a violation of copyright law should occur, please contact aachen@techlabs.org so that the corresponding item can be removed.