Cross-Process Anomaly Detection in Multivariate Time Series for Minimising Quality Drift in Electric Powertrain Production: A Predictive Quality Approach

Type: MA thesis

Status: running

Supervisors: Nastassia Vysotskaya, Andreas Maier

Manufacturing processes inherently display variability, no matter how precise and delicate the
process may be. Correctly identifying these variations in production helps enhance product quality,
promote sustainability, and improve employee morale by creating a more consistent and predictable
work environment. This Master’s thesis will use methods in Machine Learning (ML) and Statistics
to trace quality drifts and anomalies in production, potentially caused by complex dynamics of
mechanical and chemical processes in rotor production. This thesis aims to make a meaningful
contribution towards the ultimate goal of Zero Scrap in manufacturing processes. There is minimal
research on applying ML to press-in curve (force/ displacement) anomaly detection. State-of-theart
methods for detecting the generation of quality variations and defects consider only production
processes at a single stage at a time. However, there are research initiatives such as Project-
MUPROD [2014], Meiners et al. [2021], Kornas et al. [2019] to prevent the transmission of defects
between processes at a multistage level. Yet they suffer from not attaining precision and cannot
model temporal dependencies across series production processes. This thesis divides its objectives
into the following to address these shortfalls:

• To conduct a systematic literature review, gather relevant data from distributed sources, and
develop an analysis pipeline.
• To sequentially analyse regularly and irregularly sampled multivariate time series data at
multiple stages and perform anomaly detection with each stage as a quality checkpoint and
End-of-Line(EOL) as the final quality determinant. To compare a few methods.
• To identify root causes of a failure and recommend optimal process parameters/ reach new
process thresholds. Determine whether quality variations stem from a single stage or a
combination of stages in the production process.
• To model cross-process interactions with memory among multiple press-in processes using
Deep Learning architectures such as Recurrent Neural Networks or Transformers.

T. Kornas, R. Daub, M. Z. Karamat, S. Thiede, and C. Herrmann. Data-and expert-driven
analysis of cause-effect relationships in the production of lithium-ion batteries. In 2019 IEEE
15th International Conference on Automation Science and Engineering (CASE), pages 380–385,
2019. doi: 10.1109/COASE.2019.8843185.
M. Meiners, A. Mayr, and J. Franke. Process curve analysis with machine learning on the example
of screw fastening and press-in processes. Procedia CIRP, 97:166–171, 02 2021. doi: 10.1016/j.
Project-MUPROD. Innovative proactive quality control system for in-process multi-stage defect reduction.
CORDIS Europa, Oct. 2014. URL