Abstracts Track 2024

Area 1 - Industrial Informatics
Area 2 - Intelligent Control Systems and Optimization
Area 3 - Robotics and Automation
Area 4 - Signal Processing, Sensors, Systems Modelling and Control

Area 1 - Industrial Informatics

Nr: 177
Title:

Inter-Layer Pressure Control Based on Nozzle Reaction Force Sensing for 3D Printing on Uneven Surfaces
Authors:

Shinichi Ishikawa and Ryosuke Tasaki
Abstract: In 3D printing using a material extrusion method, printing parameters such as nozzle position and speed, material flow rate, and contact pressure with the layered surface are important factors that affect the quality of the printing process. This study focuses on controlling the inter-layer contact pressure by adjusting the nozzle speed, ensuring a consistent deposition process. This study aims to achieve a deposition motion for non-planar layered surfaces while maintaining a constant contact pressure to suppress the failure of deposition. In this control approach, on uneven surfaces, the nozzle feed rate is slowed down at areas where the distance between the nozzle and the deposition substrate is large. This approach increases the amount of layering material deposited on the higher groove areas, resulting in constant deposition reaction forces and a smooth layered surface. To measure the reaction force applied to the nozzle during the printing process, a force measurement device was developed using high-resolution load cells with a rated capacity of 1 N. This device enables measurement of micro changes in reaction force at the order of 0.01 N caused by surface variations. In addition, a feedback system was implemented to maintain a constant reaction force by controlling the nozzle height in response to changes in the reaction force caused by changes in the shape of the deposition substrate. Surface smoothing experiments on uneven surfaces using mortar materials were conducted to confirm the reaction force response for two shape patterns with a maximum unevenness of 5 mm. The experimental results with three-layer deposition showed that nozzle speed control performance maintained a constant reaction force even at points with large grooves. In terms of the unevenness of the layered surface, the maximum unevenness was reduced from 5 mm to 0.91 mm, which confirms a smoothing effect of 81.8%. Feedback control experiments on inclined substrates showed that deposition force was maintained with a 9.75% error rate relative to the target force. These results concluded that, as a 3D printing approach for uneven surfaces, it is possible to maintain a constant reaction force by controlling the nozzle motion.

Area 2 - Intelligent Control Systems and Optimization

Nr: 182
Title:

Evaluating Deep Learning Models for Data-Based Feedback Linearization in Nonlinear Systems
Authors:

Joanna Piasek Skupna
Abstract: Feedback linearization is a fundamental control technique used to convert nonlinear systems into linear forms, facilitating more straightforward control design. However, traditional feedback linearization methods require precise system models, which may not always be available in practice. This challenge has led to the exploration of data-driven approaches, where deep learning models can learn the nonlinear transformations required for feedback linearization directly from data. In this paper, we investigate the effectiveness of various deep learning architectures in solving the feedback linearization problem for data-based systems, focusing on evaluating the performance and limitations of existing models rather than proposing a new approach. Our study examines several neural network architectures, including encoder-decoder structures, invertible neural networks, and normalizing flows. These architectures are analyzed in terms of their ability to learn the nonlinear state and input transformations, which map the original system dynamics into a known linear form, as well as their inverse transformations, which return the system from linearized coordinates to the original state and input space. The primary objective of this work is to provide a comparative analysis of different deep learning models applied to feedback linearization. We explore the following key question: Which architectures are best suited to learning the complex nonlinear mappings involved in feedback linearization? Our experiments focus on applying these architectures to simulated dynamic systems with nonlinear behavior. The training process relies on large datasets of system trajectories, and each model is evaluated based on its ability to accurately linearize the system and invert the transformation. Finally, we discuss the implications of this analysis for practical control system design. Our findings highlight areas where further research is needed, such as improving training efficiency and enhancing the robustness of these models in real-world control applications.

Area 3 - Robotics and Automation

Nr: 178
Title:

Localization of Muscle Knots by Kneading Technique for Physical Therapy Massage Robot
Authors:

Naoya Harada and Ryosuke Tasaki
Abstract: Our research aims to propose a method for the robot to quickly recognize and localize muscle knots on the recipient’s body. It is essential to make an accurate assessment and analysis of the massage recipient’s overall muscle condition. This is done by localizing muscle knots using fingertip reaction force sensors during robotic kneading action. The robot arm is equipped with an end effector and a force sensor that can detect varying levels of muscle stiffness based on the measured reaction force. Muscle knots are detected by mimicking the circular kneading motion of the human thumb and determining the direction of the highest stiffness gradient. The experiment was done on a test piece; a PLA hemisphere; within a 70 x 70 mm sponge. The results of the experiment for a muscle knot with a radius of 25 mm revealed a distance error of 2.6 mm for the center of the muscle knot and an error rate relative to the diameter of the muscle knot of 5.1%. The localization of the center of the muscle knot was completed in a ten-finger-pressing kneading cycle. In the experiments to localize muscle knots, the proposed method for estimating the position of knots using the kneading technique allowed for the accurate localization of knots with a reduced number of presses. This confirmed the utility of the proposed method for rapid and accurate estimation of the knots during massage. In the future, we will replicate the shape, movement and deformation of the muscle knots and surrounding tissues in order to make the test piece more similar to a real human body. In addition to the localization of the muscle knots, our objective is to accurately estimate the overall distribution of stiffness and elasticity of muscles.

Area 4 - Signal Processing, Sensors, Systems Modelling and Control

Nr: 147
Title:

Addressing Hardware Imperfections in Optical Neural Networks Based on a Machine Learning-Driven Self-Correction Mechanism
Authors:

Yelim Kim, Minjoo Kim and Won Il Park
Abstract: Optical Neural Network (ONN) has been designed to address the energy consumption issue of Artificial Neural Networks (ANNs). ONN enables large-scale parallel processing, leading to lower power consumption and reduced heat generation in devices. Recent studies have focused on fast and accurate computations using digitally pre-trained weights, with advanced optical components and systems employed to minimize measurement errors and non-ideal parameters. However, given that these imperfections are hard to eliminate completely, the fabrication of imperfection-free hardware remains challenging. In this study, we developed an ONN that generates optical signals corresponding to the matrix-vector multiplication (MVM) outputs by utilizing each pixel of a liquid crystal display (LCD) panel as updatable weights and biases (WBs). Instead of constructing an imperfection-free ONN, we introduce a self-correction mechanism that addresses hardware problems through a machine-learning algorithm. Indeed, the optical classification of our ONN using a digitally pre-trained model (i.e., WBs with a 100% accuracy for 500 handwritten digits) showed a 20-30% reduction in the recognition accuracy due to the system imperfections associated with optical alignment, non-ideal functions of optical components, etc. However, after compensating for these hardware defects through our self-correction approach, the recognition accuracy and loss of our ONN achieved the levels of the digitally pre-trained model (for instance, we achieved 100% recognition accuracy for 500 handwritten digits after 59 times of epochs). This study highlights the capability of constructing defect-tolerant hardware through machine learning techniques.
Nr: 148
Title:

Efficient Color Image Recognition Using Optical Convolution Operations
Authors:

Minjoo Kim, Yelim Kim and Won Il Park
Abstract: Optical Neural Networks (ONNs) have recently been explored for energy-efficient and high-throughput image processing and inference tasks, leveraging optical fully connected matrix-vector multiplication (MVM). However, current ONN architectures mainly rely on monochromatic light from digital images, making it challenging to classify real-world images with diverse colors. In this research, we introduce an ONN system that integrates with conventional display technology, utilizing a red, green, and blue (RGB) pixel array. In a proof-of-concept demonstration, we conducted 22 epochs of iterative training using a rank-4 kernel for standard convolution, allowing for accurate classification of color images from five different fruit categories. The ONN achieved 100% accuracy on the training dataset and maintained strong performance on modified test datasets, even with added noise. Our analysis demonstrated high operational efficiency, achieving over 94% classification accuracy per MAC operation while consuming less than 17 aJ of optical energy on average. This setup provided substantial throughput, reaching approximately 2.36 teraMAC/s. Additionally, we showcased depth-wise convolution using a rank-3 kernel, separating the system into R, G, and B channels. This architecture successfully classified complex patterns consisting of three MNIST handwritten digits encoded in RGB. Our approach represents a significant advancement in optical computing and neuromorphic vision, emphasizing the potential of optical convolution operations across multiple wavelength channels.