The thesis will be written in cooperation with Georgsmarienhütte GmbH (http://www.gmh.de/).

In today’s world, quality inspections are indispensable to meet the ever-increasing requirements of industry [1]. Beyond other applications, Industry 4.0 elevates the potential for product quality inspections through the acquisition of vast amounts of data through sensors. These sensor data can provide, under certain circumstances, information about the quality of a manufactured product. Data-driven, unsupervised anomaly detection algorithms allow analyzing these data to, e.g., identify abnormal quality [2]. Unsupervised anomaly detection algorithms do not require labeled data and can generally be distinguished into nearest-neighbor, clustering, statistical, and subspace techniques [3].

In the steel manufacturing industry, cracks can be measured via non-destructive testing techniques such as eddy current testing [4]. Several measures and parameters can be recorded for every crack, e.g., length or depth, but it is not trivial to identify abnormal quality or fault clusters. To determine a defect, these sensor values can be assessed with unsupervised anomaly detection algorithms. However, finding the best algorithm for such a task is rather challenging, as its success depends on several factors, such as sample size, dimensionality, noise, or anomaly percentage [3]. To tackle this challenge, the following research questions should be answered in the Master thesis:

1. Which criteria of steel bar manufacturing quality control are relevant for the application of anomaly detection algorithms?
2. Which unsupervised anomaly detection algorithms exist?
3. How can these algorithms be classified and assessed (e.g., suitability to certain data types)?
4. Which algorithms perform best (e.g., computational complexity, accuracy) for improved steel bar manufacturing quality control?

To detect anomalies, Georgsmarienhütte currently uses the method CAPA from the R package ‘anomaly’. With this thesis, it should be investigated whether other algorithms can perform better. The student should have good knowledge of R and machine learning (preferably anomaly detection). The thesis offers exciting insights and a valuable contribution to quality control.

References

[1]    Y. Zhang, P. Peng, C. Liu, and H. Zhang, “Anomaly Detection for Industry Product Quality Inspection based on Gaussian Restricted Boltzmann Machine,” in 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC), Bari, Italy, 2019, pp. 1–6.

[2]    R. Domingues, M. Filippone, P. Michiardi, and J. Zouaoui, “A comparative evaluation of outlier detection algorithms: Experiments and analyses,” Pattern Recognition, vol. 74, pp. 406–421, 2018, doi: 10.1016/j.patcog.2017.09.037.

[3]    M. Goldstein and S. Uchida, “A Comparative Evaluation of Unsupervised Anomaly Detection Algorithms for Multivariate Data,” PloS one, vol. 11, no. 4, e0152173, 2016, doi: 10.1371/journal.pone.0152173.

[4]    J. García-Martín, J. Gómez-Gil, and E. Vázquez-Sánchez, “Non-destructive techniques based on eddy current testing,” SENSORS, vol. 11, no. 3, pp. 2525–2565, 2011, doi: 10.3390/s110302525.