Machbarkeitsanalyse für einen Machine Learning Klassifikator an einer Serienendmontagelinie
Feasibility analysis of a machine learning classifier on a series assembly line
- Machine Learning methods have shown great potential in production during recent years and have subsequently played a significant role in driving the Industry 4.0 revolution. However, there is still a lack of systematic guidelines for identifying and implementing viable ML use cases in the manufacturing industry . The aim of this work therefore, is to identify whether machine learning classifiers can be used in addition to a traditional process monitoring method (tolerance window) in order to improve quality control of staking processes on an assembly line.
Applied methods:
Extensive literature research will be initially conducted in order to understand ongoing research in this field and to clearly define the research gap. For a scientific and structured execution, CRoss Industry Standard Process for Data Mining (CRISP-DM) method will be used as the backbone of the thesis. The first step of Business Process Understanding would include learning about the staking and quality testing processes. This will be followed by data extraction and Exploratory Data Analysis (EDA) as part of Data Understanding. The Data Preparation step would include cleaning, filtering, feature engineering and scaling in order to use as input for the model.
Preparing the Machine Learning model will include selecting, modeling, training and testing the appropriate classification algorithms e.g. Näive Bayes, XGBoost, Feed Forward Neural Network and Random Forest. Depending on a pre-defined success criteria such as Accuracy, Precision etc. the training and test results will be evaluated and the best performing model will be selected. Further, subject to the model performance and business requirments , possible methods for deployment of the classifier in real time production will be discussed.
Expected results:
This work will help to identify if process curve data from staking processes can be used to train machine learning models and improve quality monitoring on an assembly line. Good model performance will indicate potential to reduce waste and decrease unnecessary processing time at the testing station. Additionally, it will also assist the process experts understand unknown influences on the product quality. A systematic approach to implement machine learning for similar use cases involving curve data of staking processes and will be developed.
