Remaining Lifetime Estimation in Semiconductor Scenarios

Results of this project are explained in detail in the final report.


Apply to this project here

Thermomechanical fatigue is one of the root causes for failures of transistors as can be found in power modules for wind turbines, photovoltaic systems or electric vehicles. Because of temperature cycles with high amplitude that take place within a short time period (seconds or minutes), the various materials within a transistor experience thermomechanical stress. A combination of Spectral Residual and Convolutional Neural Network is employed for the detection of anomalies that are detrimental to the lifetime of the transistor (arxiv:1906.03821). The overall temperature stress cycles are determined through a rainflow analysis of the time series data (ASTM E1049, DOI:10.1520/E1049-85R17) and a remaining lifetime estimation is achieved
utilizing reference data and the Palmgren-Miner rule (DOI:10.1115/1.4009458).

Data and Analysis Goals
You will access data that correspond to a time series of temperatures measured by a sensor. The simulation is carried out using the open source project TimeEval-GUI. A rainflow algorithm for the counting of thermomechanical fatigue cycles needs to be implemented in an efficient and scalable way. Thus, you produce the rainflow matrix and the corresponding load profiles. A remaining lifetime estimation is achieved by the Palmgren-Miner rule and reference data. The reference data will be provided by literature. For a rigorous monitoring of the transistor’s health, you implement an anomaly detection service developed at Microsoft and explore the viability of this approach.

Visualization and interface
The results are visualized in well-designed plots. Matplotlib is a reliable package for this purpose. Finally, a web interface is designed that includes dashboards for:

  • Plots of the input data
  • Plots of the rainflow matrix and the load profile
  • Display of remaining lifetime estimation and anomaly warning

The database for the time series data (10GB), the analysis scripts and the web interface are hosted on AWS using established technologies for scalable cloud applications. The analysis tool is triggered as an on-demand service accessing the database and publishing the results to the dashboards.

Tech Stack
AWS, Docker, Flask, Python, Matplotlib, Dash, TimeEval-GUI, Rainflow algorithm, Anomaly detection

The data science team members at PROCON IT have backgrounds in computer science, mathematics,
statistics and physics. We love to share our passion for data innovation with you.

Important notice

Accepted students to this project should attend online workshops at the LRZ in April 2023 before the semester starts, unless they have proven knowledge. More information will be provided to students accepted to this project.