University of Limerick
Browse
Mongan_2024_Development.pdf (5.15 MB)

Development of hybrid machine learning models for quality assurance and process optimisation of ultrasonic welding and computer numerical control (CNC) machining

Download (5.15 MB)
thesis
posted on 2024-02-29, 12:37 authored by Patrick MonganPatrick Mongan

Data-driven modelling has gained significant interest in the wake of Industry 4.0 as it has scope to create efficient, robust, and sustainable manufacturing systems. Traditionally, physics-based modelling approaches were the sought-after technique as they had demonstrated excellent prediction accuracies. However, as industries are adopting Industry 4.0 technologies, the amount of manufacturing related data is growing exponentially. The abundance of rich data can be transformed into actionable data in the form of predictive models. Furthermore, as a result of the remarkable developments in computing power in recent times, data-driven modelling approaches are becoming practical for many enterprises. Machine learning algorithms are data-driven modelling algorithms which are effective methods of exploiting high dimensional data. It is now feasible for enterprises to develop robust machine learning models by training them on the abundance of manufacturing data. There are various types of machine learning algorithms each with their own advantages and disadvantages. This work investigates how different machine learning algorithms can be combined to mitigate the drawbacks with stand-alone methods in order to create robust hybrid models.

To acquire insights into ultrasonic welding (USW) of aluminium 5754 and to determine the suitability of machine learning models for monitoring the USW process, an experimental investigation was conducted. A full factorial design of experiments was performed and both open and closed-loop predictive models were developed using a hybrid genetic algorithm–artificial neural network (GA-ANN) trained on the experimental data. It was found that the closed-loop predictive model, which formulates predictions based on process inputs and process feedback, reduced the prediction error from 13.17% to 7.51%, compared to the open-loop model which uses process inputs only. Therefore, the use of process feedback acquired from integrated sensor data increases the robustness of the predictive model.

Investigation of data-driven approaches progressed to optimise a USW process for the joining of a dissimilar composite material pair, i.e., Polyetherketoneketone (PEKK) to epoxy. An ensemble neural network (ENN) was developed whose base learners were hybrid GA-ANNs. The ENNs hyperparameters were optimised through Bayesian optimisation. Once trained, the ENN was exploited to identify the optimal parameter permutation that repeatedly produces the maximum joint strength while also producing a defect free joint. The predictions were subsequently validated through experimentation where the prediction error produced was just 3%.

To evaluate the agnostic nature of data-driven approaches, the modelling approach devised to optimise the USW process was applied to a CNC machining process. Initially, a full factorial design of experiments was conducted to acquire insight into the process. ANOVA identified that feed per tooth, cutting speed and depth of cut were significant contributors to surface finish, therefore, all were considered when developing the ENN. Once trained on the experimental data, the ENN was assessed on six distinct scenarios, where the predictions were validated through experimentation. The mean prediction error was just 2.56% where the maximum prediction error was 3.57%.

The findings of this thesis enhance the understanding of data-driven modelling of USW and CNC machining processes. Although the manufacturing use cases investigated were USW and CNC machining, the agnostic nature of the modelling approach is relevant to any manufacturing processes that requires process parameter refinement.


Funding

Confirm Centre for Smart Manufacturing

Science Foundation Ireland

Find out more...

History

Faculty

  • Faculty of Science and Engineering

Degree

  • Doctoral

First supervisor

Conor McCarthy

Second supervisor

Noel O’Dowd

Third supervisor

Eoin Hinchy

Department or School

  • School of Engineering

Usage metrics

    University of Limerick Theses

    Categories

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC