Abstract
Rapid technological progress is leading to an increase in the amount of all types of data generated in areas such as medicine, industry, and finance. One use of this data can be to monitor and control processes. In doing so, it is often necessary to deal with classification tasks, which is a big challenge due to the complexity and dynamics in the data. There are many tools that attempt to overcome this challenge, one of them is the hidden Markov model (HMM). HMMs are statistical models that describe the state changes in a system, where only indirect information about the underlying states is available. Based on a sequence of observable emissions from the system, HMMs allow an estimation of the most probable associated sequence of states.
This thesis first gives a detailed introduction to the topic of HMMs and examines the related issues in detail. Afterwards the concept of HMMs is applied to a practical appli- cation in the injection molding industry. A 3D accelerometer attached to the injection molding machine provides data of an injection molding process, which can be displayed in time series. The aim of this thesis is to use HMMs to identify the states of injection molding processes from 15 different data sets based on their acceleration data.
In order to achieve the best possible results, different approaches of data preprocessing and transformation are applied to obtain meaningful metrics from the time series for the model. The results of the conducted experiments give an idea of the complexity of the problem. It can be seen that the performance of the HMM varies for different datasets, which is an indication of the difficulty in developing an effective general classifier for all datasets. It is shown that increasing the number of HMMs used to four models leads to considerable improvements in the overall results. Furthermore, an approach is presented that uses additional specific knowledge about the injection molding process to post-correct the classification results.
This thesis first gives a detailed introduction to the topic of HMMs and examines the related issues in detail. Afterwards the concept of HMMs is applied to a practical appli- cation in the injection molding industry. A 3D accelerometer attached to the injection molding machine provides data of an injection molding process, which can be displayed in time series. The aim of this thesis is to use HMMs to identify the states of injection molding processes from 15 different data sets based on their acceleration data.
In order to achieve the best possible results, different approaches of data preprocessing and transformation are applied to obtain meaningful metrics from the time series for the model. The results of the conducted experiments give an idea of the complexity of the problem. It can be seen that the performance of the HMM varies for different datasets, which is an indication of the difficulty in developing an effective general classifier for all datasets. It is shown that increasing the number of HMMs used to four models leads to considerable improvements in the overall results. Furthermore, an approach is presented that uses additional specific knowledge about the injection molding process to post-correct the classification results.
| Translated title of the contribution | Versteckte Markov-Modelle zur Zustandserkennung in Spritzgussprozessen unter Verwendung von Beschleunigungsdaten |
|---|---|
| Original language | English |
| Qualification | Master |
| Supervisors/Reviewers |
|
| Award date | 24 Oct 2023 |
| Publication status | Published - 24 Oct 2023 |
Fields of science
- 101019 Stochastics
- 101 Mathematics
- 101018 Statistics
- 101014 Numerical mathematics
- 101024 Probability theory
JKU Focus areas
- Digital Transformation