Search the digital library catalog

Abstract: We explore the sparsity and prunability of multi-layer perceptrons (MLPs) trained using the Forward-Forward (FF) algorithm, an alternative to backpropagation (BP) that replaces the backward pass with local, contrastive updates at each layer. We analyze the sparsity of the weight matrices during training using multiple metrics, and test the prunability of FF networks on the MNIST, FashionMNIST and CIFAR-10 datasets. We also propose FFLib—a novel, modular PyTorch-based library for developing, training and analyzing FF models along with a suite of FF-based architectures, including FFNN, FFNN+C and FFRNN. In addition to structural sparsity, we describe and apply a new method for visualizing the functional sparsity of neural activations across different architectures using the HSV color space. Moreover, we conduct a sensitivity analysis to assess the impact of hyperparameters on model performance and sparsity. Finally, we perform pruning experiments, showing that simple FF-based MLPs exhibit significantly greater robustness to one-shot neuron pruning than traditional BP-trained networks, and a possible 8-fold increase in compression ratios while maintaining comparable accuracy on the MNIST dataset.
Keywords: Forward-Forward, sparsity, pruning, model compression, machine learning, neural network
Published in DKUM: 20.08.2025; Views: 0; Downloads: 2
Full text (7,01 MB)

Abstract: Improving the power density of an interior permanent magnet synchronous machine requires a complex and comprehensive approach that includes electromagnetic and thermal aspects. To achieve that, a multi-objective optimization of the machine’s geometry was performed according to selected key performance indicators by using numerical and analytical models. The primary objective of this research was to create a computationally efficient and accurate alternative to a direct finite element method-based optimization. By integrating artificial neural networks as meta-models, we aimed to demonstrate their performance in comparison to existing State-of-the-Art approaches. The artificial neural network approach achieved a nearly 20-fold reduction compared with the finite element method-based approach in computation time while maintaining accuracy, demonstrating its effectiveness as a computationally efficient alternative. The obtained artificial neural network can also be reused for different optimization scenarios and for iterative fine-tuning, further reducing the computation time. To highlight the advantages and limitations of the proposed approach, a multi-objective optimization scenario was performed, which increased the power-to-mass ratio by 16.5%.
Keywords: interior permanent magnet synchronous machine, artificial neural network, metamodel, multi-objective optimization, finite element method
Published in DKUM: 08.08.2025; Views: 0; Downloads: 17
Full text (2,87 MB)

Abstract: This Dissertation focuses on the optimization of an Interior Permanent Magnet (IPM) machine for handheld battery-powered tools, aiming to enhance performance and efficiency. The research integrates multi-physics modeling, including electromagnetic Finite Element Method (FEM) and thermal models, to evaluate machine performance under various operating conditions. The performance is evaluated according to selected Key Performance Indicators (KPIs). Further, different control methods, such as Field Oriented Control and Square-Wave Control, impact the performance significantly and are incorporated into the optimization process. Due to the computational challenges of FEM-based performance evaluations in Multi-Objective Optimization (MOO), this work utilizes Artificial Neural Network (ANN)-based meta-models, to accelerate the optimization process while preserving accuracy. The developed meta-models capture nonlinear machine characteristics from the FEM model. These meta-models are then used to evaluate machine performance through a combination of analytical and numerical post-processing methods. Four MOO scenarios are presented, each aimed at optimizing the cross-sectional design of IPM machines, to enhance performance and efficiency while reducing mass and cost. Additionally, these scenarios modify the machine’s electromagnetic behavior, to ensure better alignment with the selected control method. By comparing the optimization process of Scenario 1, which uses direct FEM-based evaluation without time reduction measures, to the approach incorporating Artificial Neural Network based meta-models, the total number of individual FEM evaluations decreased from 2.35×10^9 to 2.03×10^5, without almost any loss of accuracy. This reduced the computation time from 297 years to 9.07 days on our standard desktop computer. The obtained ANN-base meta-models can be used further for other optimizations without the need for additional FEM evaluations. In all four optimization scenarios, the use of meta-models enabled the generation of a Pareto front of the optimal solutions, leading to improved KPIs compared to the reference design. The highest relative improvement occurred in Scenario 1, where the selected optimized machine design achieved a 30% increase in power density compared to the reference design.
Keywords: Interior Permanent Magnet (IPM) Machine, Artificial Neural Network (ANN), Meta-Modeling, Multi-Objective Optimization (MOO), Finite Element Method (FEM), Multi-Physics Modeling, Field Oriented Control (FOC), Square-Wave Control (SWC)
Published in DKUM: 15.05.2025; Views: 0; Downloads: 138
Full text (17,79 MB)

Abstract: This study focuses on the segmentation part in the development of a potato-sorting system that utilizes camera input for the segmentation and classification of potatoes. The key challenge addressed is the need for efficient segmentation to allow the sorter to handle a higher volume of potatoes simultaneously. To achieve this, the study employs a region-based convolutional neural network (R-CNN) approach for the segmentation task, while trying to achieve more precise segmentation than with classic CNN-based object detectors. Specifically, Mask R-CNN is implemented and evaluated based on its performance with different parameters in order to achieve the best segmentation results. The implementation and methodologies used are thoroughly detailed in this work. The findings reveal that Mask R-CNN models can be utilized in the production process of potato sorting and can improve the process.
Keywords: image segmentation, potato sorting, neural network, mask RCNN, object detection, production process, machine learning, AI
Published in DKUM: 27.03.2025; Views: 0; Downloads: 21
Full text (5,97 MB)
This document has many files! More...

Abstract: In the heavy industry, the value of cargo transported by rail is very high. Due to high value, poor security and volume of rail transport, the theft cases are often. The main problem of securing rail transport is predicting the location of a high probability of risk. Because of this,the aim of the presented research was to predict the highest probability of rail cargo theft for areas. It is important to prevent theft cases by better securing the railway lines. To solve that problem the authors' model was developed. The model uses information about past transport cases for the learning process of Artificial Neural Networks (ANN) and Machine Learning (ML).The ANN predicted the probability for 94.7% of the cases of theft and the Machine Learning identified 100% of the cases. This method can be used to develop a support system for securing the rail infrastructure.
Keywords: rail transport security, supply chain disruption, drones, security support systems, cargo theft, predicting, logistics, artificial neural network, drone monitoring, machine learning
Published in DKUM: 28.01.2025; Views: 0; Downloads: 4
Full text (1,93 MB)
This document has many files! More...