Kochi University of Technology Academic Resource Repository

菊森, 剛

Japanese

クラウドコンピューティングにおける 可用性と消費電力を考慮した 仮想マシン配置手法

渡邊, 翔太

ゲーム「2048」へのMonte-Carlo Softmax 探索の適用

松﨑 峻平

酸化亜鉛ナノロッドの合成及び光分解効果の評価

高知工科大学博士（学術）doctoral thesi

Wang Chen

English

Hermeneutical Studies for Human-Engaged  Computing: Theoretical Justification, Practical  Reflections, and Next-Generation User Interface Design

Annisa REZA, Kento IWAI, Nagatoshi NISHIWAKI

EVALUATION OF THE BULKINESS OF PERI-SUBSTITUENTS DISTORTING 1,8-DIMETHYLNAPHTHALENE FRAMEWORK

Liping WU, Koichi OKA, Ning LIU, Feng SUN

The 8th International Symposium on Frontier Technology (ISFT)
Performance Study of Single DOF Linear Motor with E-shaped Electromagnet
Variable Structure
Liping WU*, Koichi OKA**, Ning LIU*, Feng SUN*
Shenyang University of Technology*
Kochi University of Technology**
Corresponding author: Koichi OKA (oka.koichi@kochi-tech.ac.jp)
ABSTRACT: In order to study the effect of structural changes on the performance of linear motors, we take
the linear motor system with electromagnet (EM) as a slider composed of E-shaped iron core as the research
object, and analyze the levitation force Fy, driving force Fx, and torque around the center of mass T of the
E-shaped EM. By changing the length of the U, V, W poles of the E-shaped core to change the structure of
the EM to investigate the variation rules of the levitation force, driving force and torque around the center of
mass of the E-shaped EM. Firstly, the structure change is divided into 2 groups, respectively, change 1 pole
and change 2 poles, and the Fx, Fy and T of the E-shaped EM are simulated by using JMAG software. The
results show that when changing the core structure of the E-shaped EM, the Fx, Fy and T will be affected, in
which the Fy is affected more when the length of a single pole increases, and the Fy is reduced obviously at
the beginning when a single pole decreases, but the trend of the Fy reduction tends to be stable with the
increase of the reduction of a single pole; in terms of the Fx and the T, when the In terms of Fx and T, when
the altered structure is symmetrically altered, there is little effect, and when the altered structure is
asymmetric, the Fx and T are more affected.
KEYWORDS: linear motor, E-shaped electromagnet, structural changes, driving force, levitation force
1. INTRODUCTION
Magnetic levitation technology is widely used in
many industrial fields due to its non-contact,
non-lubrication and easy control. Especially, it is
applied in some special occasions such as clean
room or vacuum environment. Linear motor drive is
a commonly used technology in clean transfer. Guo
Y. [1] et. al designed a non-contact transfer device
based on magnetic levitation technology, which is
controlled by brushless DC control scheme. Most of
the existing magnetic levitation linear drives are
composed of a levitation mechanism and a drive
mechanism [2], which is good to realize dust-free
transmission and drive, however, this will lead to a
larger device, and for the levitation type of conveyor
equipment, its own gravity is too large, which will
reduce the goods carried and affect the efficiency of
transmission. And too large equipment will also lead
to an increase in energy consumption.
Based on this, this paper proposes a magnetic
levitation linear motor with an E-shaped core that is
integrated with levitation and drive, and this device
can realize the miniaturization of dust-free
transmission and improve the transmission efficiency.
However, this structure poses a big challenge to the
control system, and in order to be able to better
The 8th International Symposium on Frontier Technology (ISFT)
control the device, its performance and mechanical
properties need to be studied. Therefore, this paper
takes the structure of the E-shaped electromagnet
(EM) core as the research object, and changes the
structure of the EM by changing the length of the
three poles of U, V, and W. It explores the effect of
changing the structure of the EM on the magnetically
levitated linear drive system.
2 E-SHAPED EM VARIABLE STRUCTURE
DESIGN
The linear motor drive unit is shown in figure 1. It
consists of a frame made of aluminum profiles, a
fixed guide made of a permanent magnet array, an
E-shaped EM, and a laser displacement sensor.
Wherein both the drive and levitation of the device
are provided by the attraction between the E-shaped
EM and the permanent magnets of the linear array.
Figure 1 Linear motor drive
The force analysis of an E-shaped EM is shown in
figure 2, where Fy provides the levitation force, Fx
provides the driving force, and T around the center
of mass causes the EM to rotate.
Figure 2 Force diagram of EM
The Fx, Fy and T can be calculated from equation (1).





 

211
,;,,
,
LFLFLFT
FFFFF
xuywy
yxjwvui
wjvjujijyx
(1)
The scheme of structural change is shown in figure 3.
The purpose of changing the structure of the EM is
achieved by changing the length of the U, V and W
poles. Here the structure change is divided into 2
groups. (1) changing only one pole, i.e., dU=dW=0
and changing dV (changing U and W poles is similar
and not listed here); (2) changing 2 poles, i.e., dU=0
and changing dV and dW; and then analyzing its Fy,
Fx, and T through JMAG simulation.
Figure 3 Variable structure design
3 SIMULATIONANALYSIS
According to the EM variable structure scheme
designed in Section 2, simulation analysis is carried
out using JMAG and the simulation parameters are
shown in table 1.
Table 1 Simulation parameters
Name Value
Current 0 A
Permanent magnets N35
Iron core type SS40
Air gap 9 mm
The simulation conditions are as follows: keep dU
constant and vary the lengths of dV and dW ranging
from -1mm to 1mm, varying by 1mm each time. let
the EM and permanent magnet move relative to each
other, moving by a distance of 1mm at each step,
moving by a total of 60mm. solve for Fx, Fy, and T at
The 8th International Symposium on Frontier Technology (ISFT)
each position, and the results are shown in figure 4.
(a) Fx
(b) Fy
(c) T
Figure 4 Simulation results
From figure 4(a), it can be seen that case1 and case2
change less, while case3 and case4 increase
significantly, indicating that dV has a greater effect
on the Fx. Figure 4(b) shows that case1 decreases
significantly, case4 increases significantly, and case2
and case3 do not change much, indicating that the
average air gap has a greater effect on Fy. Figure 4(c)
shows that case1 and case3 decrease and case2 and
case4 increase, indicating that the change in dW has
a more pronounced effect on T.
4 CONCLUSION
This paper takes the E-shaped electromagnet of
single DOF linear motor as the research object. It
analyzes the change pattern of levitation force,
driving force and torque around the center of mass
after changing the structure of E-shaped EM, and
obtains the following conclusions: (1) When the
structure of E-shaped electromagnet is changed, it
has an effect on Fx, Fy and T. (2) The Fy is most
affected by the change of electromagnet structure
and is closely related to the average air gap.
REFERENCES
Guo Y, Jin J, Zhu J and Lu H. 2007. Design and
analysis of a prototype linear motor driving system
for HTS maglev transportation. IEEE transactions
on applied superconductivity, 17(2): 2087-2090.
(Journal Articles)
Sun F, Pei W, Zhao C, Jin J, Xu F and Zhang X.
2022. Permanent Maglev Platform Using a Variable
Flux Path Mechanism: Stable Levitation and Motion
Control[J]. IEEE Transactions on Magnetics, 58(7):
1-10. (Journal Articles)


Performance Study of Single DOF Linear Motor with E-shaped Electromagnet Variable Structure

谷久保, 泰樹

アクリロニトリル系π共役分子の架橋構造デザインと その固体光物性

Zhen ZHANG

－1－ 
 
論 文 内 容 の 要 旨 
 
Electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) are 
neuroimaging techniques used to study brain function. fMRI provides high spatial resolution, allowing 
for detailed visualization of brain structures, while EEG offers high temporal resolution, allowing for 
the study of rapid brain processes. These techniques are crucial for understanding various cognitive 
processes, such as perception, attention, memory, and language. Deep learning (DL) algorithms as 
high-performance tools for pattern recognition and regression problem have been increasingly applied 
to fMRI and EEG data. DL-based techniques can automatically learn complex and meaningful 
representations from the raw data, which can extract high-level features from the voxel-level fMRI 
volumes and be applied to tasks such as brain state classification or decoding cognitive processes. For 
EEG analysis, DL-based techniques can learn temporal patterns and relationships in the electrical 
signals, facilitating tasks like emotion recognition or seizure detection. Therefore, this thesis conducted 
research based on deep learning algorithms from three perspectives, including P300 detection (a type of 
EEG signal), EEG denoising, and fMRI classification, and proposed a spatial-temporal neural network 
(STNN), a multi-module neural network (MMNN), and a multi-pooling 3D convolutional neural 
network (MP3DCNN) to resolve these challenges. Specifically, the first chapter provided a brief 
overview of EEG, fMRI, and deep learning algorithms, where the definition, characteristics, and 
application of EEG and fMRI signals are given. Then the principle, key concepts, and real applications 
of deep learning algorithms are introduced. Finally, the potential and risks of EEG and fMRI signal 
processing based on deep learning are listed. From the second to the fourth chapter, the background, 
materials, methods, experiments, and discussion about the proposed algorithms will be described, 
respectively. The fifth chapter summarizes this thesis. The main works are brief described as follows: 
1) Spatial-temporal neural networks (STNN) for P300 detection: The P300 signal, also known 
as P3 or P300 component, is an electrophysiological response observed in the event-related potential 
(ERP) waveform recorded from the scalp using EEG tools. It typically occurs around 250-500 
milliseconds after the presentation of a rare or unexpected stimulus and is associated with attention, 
cognitive processes, and decision-making. P300 spellers are common brain-computer interface (BCI) 
systems designed to transfer information between human brains and computers. In most P300 
detections, the P300 signals are collected by averaging multiple electroencephalographic (EEG) changes 
to the same target stimuli, so the participants are obliged to endure multiple repeated stimuli. 
Therefore, an STNN is proposed to detect P300 signals, which can detect signals by combining the 
outputs from a temporal unit and a spatial unit. The temporal unit is a flexible framework consisting of 
several temporal modules designed for analyzing brain potential changes in the time domain. The 
spatial unit combines one-dimensional convolutions (Conv1Ds) and linear layers to generalize P300 
features from the space domain, and it can decode EEG signals recorded using different numbers of 
electrodes. We demonstrate the effectiveness of our model using three public databases: P300 speller 
with ALS patients, covert and overt ERP-based BCI, and BCI Competition III-dataset II. In the dataset 
of P300 speller with ALS patients, EEG signals from eight ALS patients were recorded, and every 
participant in the study went through 35 trials, with 10 rounds of repeated stimuli in each trial. Every 
round of stimuli contained two target stimuli and 10 nontarget stimuli. In the dataset of covert and 
overt ERP-based BCI, 10 healthy subjects took part in the experiment. The EEG data are recorded on 
Farwell and Donchin’s paradigm and the Geometric Speller. Each experiment included three sessions, 
with six trials in each session. Each trial contained eight rounds of repeated stimuli with 12 stimuli 
－2－ 
within each round of stimuli. In the dataset of BCI Competition III-dataset II, both EEG signals from 
two subjects (A and B) were divided into a training set (85 trials) and a testing set (100 trials). Every 
trial contained 15 rounds of repeated stimuli. We implemented a within-subject P300 detection and a 
cross-subject P300 detection, respectively using the dataset of P300 speller with ALS patients as well as 
covert and overt ERP-based BCI. 
The results showed that both amyotrophic lateral sclerosis (ALS) patients and healthy subjects 
can benefit from this study. In the within-subject P300 detection and the cross-subject P300 detection, 
the proposed STNN gained higher performance with fewer repeated stimuli than other comparative 
approaches. Furthermore, we applied the proposed STNN in the P300 detection challenge of BCI 
Competition III. The accuracy score was 89% in the fifth round of repeated stimuli, outperforming the 
best result in the literature (accuracy = 80%) to the best of our knowledge. The results demonstrate that 
the proposed STNN performs well with limited stimuli and is robust enough for various P300 
detections. The main reasons are as follows: 1) the temporal unit, as a flexible DL-based network 
dedicated to time-domain modeling, can capture the temporal dependencies from brain potential 
changes by constructing an end-to-end multi-level sequential mapping, so it is more sensitive than the 
previously mentioned approaches when detecting P300 signals; 2) the spatial unit can constantly 
generalize and compress P300 features in the space domain, which hedges complex noise interference 
to a certain extent; 3) a joint decision-making mechanism is built into the network by connecting the 
temporal unit and the spatial unit concurrently, which can utilize the above advantages of the two 
units, thus achieving both better performance and stronger robustness. In the future, the proposed 
STNN is predicted to reach a high information transfer rate (ITR) when implementing online P300 
detection. Moreover, we consider that this network has potential for applications in EEG-BCI systems 
and some other areas of signal processing, such as Electrocardiogram (ECG) classification, seeing that 
it is designed with a flexible structure and can be fast training and testing with limited data. 
2) Multi-module neural networks (MMNN) for EEG denoising: This study proposed an MMNN 
to remove ocular artifacts (OAs) and myogenic artifacts (MAs) from noisy single-channel 
electroencephalogram (EEG) signals. This network consists of multiple denoising modules connected in 
parallel. Each denoising module is built using one-dimensional convolutions (Conv1Ds) and fully 
connected (FC) layers, and it estimates not only clean EEG signals but also artifacts. The proposed 
MMNN has two main advantages. First, the multiple denoising modules can purify noisy input EEG 
signals by continuously removing artifacts in the forward propagation. Second, the parallel architecture 
allows the parameters of each denoising module to be updated concurrently in the backpropagation, 
thereby improving the learning capacity of neural networks. We tested the network denoising 
performance using a recent public database, namely, EEGdenoiseNet. This database provides 
large-scale clean EEG and artifact epochs, involving 4514 clean EEG epochs, 3400 EOG epochs, and 
5598 EMG epochs. These epochs were used to synthesize the training and testing data. Among them, 
we implemented the model evaluation using 3000 pairs and 400 pairs of training and testing epochs for 
the OA removal, as well as 5000 pairs and 598 pairs of training and testing epochs for the MA removal. 
These epochs are synthesized at ten different noise levels, aiming to simulate the real applications. And 
we performed a 10-fold cross-validation on the training samples for hyperparameter tuning. 
The testing results revealed that the proposed network reduced the temporal relative root mean 
square error (T-RRMSE) and spectral relative root mean square error (S-RRMSE) by at least 6% and 
enhanced the correlation coefficient (CC) by at least 3% over the state-of-the-art approaches. Observing 
the deviation distribution between the denoised and clean signals confirmed these significant 
performance improvements. Furthermore, the proposed network achieved a similar performance 
efficiency with only 60% of the training data compared to the existing DL models. Finally, the proposed 
model was compared with the non-deep learning techniques. According to the ANOVA results with 
Holm-Bonferroni correction, the performance improvement is significant (all p-values < 0.001) in both 
－3－ 
the OA and MA removals. In the future, there are some challenges worth exploring using the proposed 
model. For example, OAs and MAs are entangled with motion artifacts in a real EEG epoch, however, 
there is no available public database to evaluate the model performance for the mixed signals. Given 
that the proposed model offers significant advantages over the conventional and DL models in this 
study, the related research is within the scope of further work. 
3) Multi-pooling 3D Convolutional Neural Networks (MP3DCNN) for fMRI Classification: 
Neural decoding of visual object classification via functional magnetic resonance imaging (fMRI) data is 
challenging and is vital to understand underlying brain mechanisms. The previous study performed 
categorical (face vs. object), face sub-categorical (male face vs. female face), and object sub-categorical 
(natural object vs. artificial object) classifications via a classic three-layer 3DCNN, revealing that the 
human visual system recognizes objects following the principle of going from categories into 
sub-categories. However, the previous classification model did not significantly present a high accuracy 
even with 9-fold fMRI data averaging, especially for sub-categorical classification tasks. Therefore, a 
novel multi-pooling 3D convolutional neural network (MP3DCNN) is proposed, which is expected to 
reach a higher accuracy than the previous model and play a valuable role in decoding brain 
mechanisms. The proposed MP3DCNN included a feature extraction, a feature combination, and a 
classifier, where the feature extraction has a mainchain and two branches. The mainchain is a 
three-layer 3DCNN, where each 3D convolution is combined with a batch normalization, an average 
3D pooling layer, and a rectified linear unit (ReLU). The first and second 3D convolutions each have a 
branch connection, where an average 3D pooling layer and a linear layer are used to generalize further 
and connect the extracted features. In the end, through the feature combination and classifier, the 
model can provide a sophisticated decision using multi-level features in fMRI classification.  
In the work, the fMRI dataset is from the previous study, where 53 healthy subjects 
participated in the visual stimulus task, and each one experienced an average of 9.96 ± 2.88 rounds of 
visual stimuli, where the subject clicked the button corresponding to a random visual stimulus (an 
image of a male face, female face, natural object, or artificial object) within 0.5s. During this period, a 
Siemens 3T MRI scanner recorded the subject’s brain states as T1-weighted 3D fMRI volumes. 
Through SPM12, the fMRI volumes were realigned, co-registered, normalized to the standard Montreal 
Neurological Institute (MNI) template, and resampled to 2-mm isotropic voxels. After data cleaning, 
there are 17306 fMRI volumes from 50 subjects available, including 4453, 4399, 4214, and 4240 
volumes corresponding to the visual stimuli of male face, female face, natural object, and artificial 
object, respectively. To suppress the background noise and the irrelevant neural activities, the fMRI 
dataset of each subject was multi-fold averaged as an option to improve the data quality (for example 
9-fold fMRI data averaging). We performed 9-fold cross-validation with 25 training iterations, where 
the batch size was 64, the learning rate was 0.00001, and the loss function was binary cross entropy 
(BCE). Within the 25 iterations, the model parameters with the highest accuracy score on each 
validation dataset were used for model testing. Finally, we used a majority voting scheme to ensemble 
the nine results in determining the classification accuracy.  
The results showed that this model can improve the classification accuracy for categorical (Face 
vs. Object), face sub-categorical (Male face vs. Female face), and object sub-categorical (Natural object 
vs. Artificial object) classifications from 1.684% to 14.918% over the previous study in decoding brain 
mechanisms. The main reasons are speculated as follows: 1) we considered that the multiple 3D 
average pooling layers can against the local feature redundancy in the feature extraction process and 
pass the global information to the classifier as much as possible; 2) through the branch connections, the 
model decision can depend on the merged features of the three 3D convolutions, thereby improving the 
model’s robustness. In future research, we look forward to using grid-search to optimize the model 
hyperparameters and exploring the visual explanations based on the reached classification results.  
Overall, this thesis proposed high-performance deep learning algorithms for P300 detection, 
－4－ 
EEG denoising, and fMRI classification, respectively. And they are expected to improve the efficiency in 
various brain decoding tasks in the future. 
 
 
 
 
 
 
 
 
 
 
 
 


A Study on Deep Learning Algorithms for Electroencephalography (EEG) and Functional Magnetic Resonance Imaging (fMRI) Signal Processing

楠瀬　慶太

市民科学による地域文化資源継承の実践と方法 －歴史学の社会実装に向けて－

研究論文（査読あり論文）departmental bulletin pape

石谷, 康人

アントレプレナーによるアイデンティティ・マネジメントと希少／異質な資源蓄積の関係性（査読あり論文）

Kochi University of Technology

Not a member yet

クラウドコンピューティングにおける可用性と消費電力を考慮した仮想マシン配置手法

ゲーム「2048」へのMonte-Carlo Softmax 探索の適用

酸化亜鉛ナノロッドの合成及び光分解効果の評価

Hermeneutical Studies for Human-Engaged Computing: Theoretical Justification, Practical Reflections, and Next-Generation User Interface Design

EVALUATION OF THE BULKINESS OF PERI-SUBSTITUENTS DISTORTING 1,8-DIMETHYLNAPHTHALENE FRAMEWORK

Performance Study of Single DOF Linear Motor with E-shaped Electromagnet Variable Structure

アクリロニトリル系π共役分子の架橋構造デザインとその固体光物性

A Study on Deep Learning Algorithms for Electroencephalography (EEG) and Functional Magnetic Resonance Imaging (fMRI) Signal Processing

市民科学による地域文化資源継承の実践と方法－歴史学の社会実装に向けて－

アントレプレナーによるアイデンティティ・マネジメントと希少／異質な資源蓄積の関係性（査読あり論文）

2,359

2,367