Automated generation of pipelines for machine learning in industrial manufacturing

Created by W.Langdon from gp-bibliography.bib Revision:1.8862

@PhdThesis{Zoeller:thesis,

• author = "Marc-Andre Zoeller",
• title = "Automated generation of pipelines for machine learning in industrial manufacturing",
• school = "Fakultaet Konstruktions, Universitaet Stuttgart",
• year = "2024",
• address = "Germany",
• month = "19 " # jul,
• keywords = "genetic algorithms, genetic programming, Maschinelles Lernen, Automation, 2025ZollerMarc, CRISP-DM, CASH, AutoML, AVATAR, ML pipeline, DSWIZARD, auto-sktime, HPO, XAutoML, CRISP-DM",
• publisher = "Stuttgart : Fraunhofer-Institut fuer Produktionstechnik und Automatisierung IPA",
• URL = "https://elib.uni-stuttgart.de/items/dee2b434-a5ee-4a7a-80d7-dbf97b4815df",
• URL = "https://elib.uni-stuttgart.de/server/api/core/bitstreams/d58663f8-07d2-4b8c-9dfc-763940982cc6/content",
• URL = "http://nbn-resolving.de/urn:nbn:de:bsz:93-opus-ds-162670",
• URL = "https://elib.uni-stuttgart.de/handle/11682/16267",
• URL = "https://www.ipa.fraunhofer.de/de/Publikationen/dissertationen.html?&%253Bno_cache=1#19674820_908387356",
• URL = "https://katalog.dnb.de/EN/resource.html?id=1374190004&pr=0&sortA=bez&sortD=-dat&v=list",
• DOI = "10.18419/opus-16248",
• size = "412 pages",
• abstract = "In recent years, machine learning (ML) has become a vital part of the fourth industrial revolution, known as Industry 4.0. The pervasive integration of sensors and the digital transformation of processes have elevated data to a crucial resource within modern manufacturing. Yet, harnessing this data to craft dependable and well-performing ML models remains a challenge. This task often requires a cross-functional team of domain experts who understand the underlying business domain and trained data scientists with extensive ML expertise to create a top-performing ML model successfully. Automated machine learning (AutoML) aims to revolutionize the iterative, manual model-building process by automating the tasks involved. Despite AutoML promising results across diverse domains, its integration within the manufacturing area remains essentially non-existent.

  This thesis delves into the challenges preventing AutoML adoption within Industry 4.0. Therefore, we introduce two related technical contributions to efficiently build highly adaptable, well-performing, and robust ML models, leading to a single generic approach for the diverse requirements in the manufacturing landscape. Additionally, we propose an adaptation of AutoML to time series data, an integral data type within manufacturing. Beyond technical innovation, this thesis examines potential barriers preventing the practical use of AutoML in Industry 4.0. We investigate AutoMLŠs alignment with the standard process for model-building (CRISP-DM) and practitioners motivations for using AutoML methods. Furthermore, we introduce a novel visual Analytics tool to explain and validate AutoML optimizations, leading to increased trust in AutoML and remedying a major obstacle for the practical use of AutoML. Finally, we evaluate our contributions to AutoML on two real-world Industry 4.0 use cases. First, we show that AutoML is able to outperform manually created ML models for predicting the remaining useful life of engineering systems, a crucial step in predictive maintenance. Second, we use AutoML to effectively generate price predictions for used heavy construction equipment in a case study with a strong focus on enabling domain experts with no ML expertise.",

• language = "English",
• notes = "urn:nbn:de:bsz:93-opus-ds-162670

  Record ID 1374190004

  DDC notation 658.5

  Supervisor: Marco Huber",

}

Genetic Programming entries for Marc-Andre Zoeller

Citations