Coping with Limited Data Amounts When Applying Artificial Intelligence in Industrial Process Control
Speaker: Dr. Thomas Schütte – PLASUS GmbH
2025/5/19 pm12:30 – pm12:50 SVC TechCon 2025
ChatGPT 或 DeepSeek 等大型語言模型 （LLM） 的最新進展引起了人們對人工智慧在工業過程控制領域的前景的極大興趣。LLM 和其他基於深度學習架構的方法依賴於用於訓練模型的非常大的數據集。
在工業環境中，生成全面且經過驗證的數據點通常需要花費大量時間和金錢。單個數據點需要在感興趣的過程中處理至少一個樣品，並在之後進行大量測量和測試，以表徵產品的性能。來自各個運行的過程參數必須與表徵結果的結果一起以機器可讀的格式存儲。瓶頸通常是必要的表徵測量量，這可能不是通常質量保證程式的一部分。
因此，將人工智慧 （AI） 方法應用於工業過程控制的一個主要挑戰是根據有限大小的數據集提取有意義的 AI 模型。在本次演講中，我們將介紹在“以高解析度壓電超聲波感測器為例對薄膜材料進行材料研究數位化”（DigiMatUs） 專案的背景下制定的應對策略，該專案是德國研究計劃“MaterialDigital”的一部分。在本專案中，通過機器學習研究了用於超聲換能器的 AlScN 層的反應磁控濺射。應對有限數據量的策略包括使用原位診斷技術，如等離子體監測或原位反射法，求助於物理量和引入專業知識。
Name
Coping with Limited Data Amounts When Applying Artificial Intelligence in Industrial Process Control
Date
2025年5月19日 Monday
Time
pm12:30 – pm12:50
Description
Thomas Schütte1, Jan-Peter Urbach1, Fabian Neuhaus2, Martin Glauer2, Stephan Barth3, Christian Käpplinger4
1PLASUS GmbH, Mering, Germany
2Otto von Guericke University of Magdeburg, Germany
3Fraunhofer Institute for Electron Beam and Plasma Technology FEP, Dresden, Germany
4PVA TePla Analytical Systems GmbH, Jena, Germany
The recent progress of large language models (LLMs) such as ChatGPT or DeepSeek has generated a lot of interest for the prospect of artificial intelligence in the field of industrial process control. LLMs and other approaches based on deep learning architectures rely on very large data sets that are used train the models.
In the industrial context generating comprehensive and validated data points usually requires considerable effort in terms of time and money. A single data point will require to process at least one sample in the process of interest and perform a number of measurements and tests afterwards to characterize the performance of the product. Process parameters from the respective run have to be stored together with the results of the characterization results in a machine-readable format. The bottleneck is most often the necessary amount of characterization measurements that may not be part of the usual quality assurance routine.
So, a major challenge to apply artificial intelligence (AI) methods for industrial process control is to extract meaningful AI models based on data sets of limited size. In this talk we will present strategies for this challenge that are developed in the context of the project “Digitalization of materials research on thin-film materials using the example of high-resolution piezoelectric ultrasonic sensors” (DigiMatUs) which is part of the German research initiative “MaterialDigital”. In this project, reactive magnetron sputtering of AlScN layers for the use in ultrasonic transducers is investigated by means of machine learning. Strategies to cope with the limited amount of data include the use of in-situ diagnostic techniques such as plasma monitoring or in-situ reflectometry, the recourse to physical quantities and the introduction of expert knowledge.
Speakers
Thomas Schütte – PLASUS GmbH