Repository of FERIT Osijek
Not a member yet
4053 research outputs found
Sort by
Optimization of Deep Learning Algorithms for the Cardiovascular System Image Processing Using Residual Units
No full textIzraz kardiovaskularne bolesti (KVB) odnosi se na brojne funkcionalne abnormalnosti srca i krvožilnog sustava. KVB uzrokuju gotovo jednu trećinu (33%) smrtnosti u suvremenom svijetu, što predstavlja najveći udio u odnosu na sve druge bolesti. Rana dijagnoza i odgovarajuće liječenje kardiovaskularnih bolesti mogu značajno smanjiti smrtnost I poboljšati kvalitetu pacijentova života. Postavljanje dijagnoze temelji se na cjelokupnoj slici kardiovaskularnog sustava, uključujući anatomiju i fiziologiju srca. Dijagnostički proces obično se sastoji od dva glavna dijela. Prvi dio odnosi se na prikupljanje slika srca pomoću medicinskih uređaja. Razvijene su brojne invazivne i neinvazivne tehnike medicinskog snimanja koje omogućuju uvid u anatomiju i funkcionalnost srca. Drugi dio dijagnostičkog procesa je kvantifikacija i interpretacija prethodno dobivenih slika pomoću naprednih metoda obrade slike. Razvoj učinkovitih metoda za obradu medicinskih slika je složen zadatak, s obzirom da podrazumijeva obradu ogromne količine visokodimenzionalnih podataka. Napredak u razvoju algoritama obrade slike, računalnog vida i umjetne inteligencije, kao i dostupnost grafičkih procesorskih jedinica (GPU-a), značajno su olakšale i ubrzale razvoj takvih metoda.
Segmentacija medicinskih slika ima važnu ulogu u procjeni, dijagnozi te postavljanju prognoze različitih kardiovaskularnih bolesti. Opsežna istraživanja i kliničke primjene pokazale su da računalna tomografija (CT) i magnetska rezonanca (MRI), kao osnovne tehnike prikupljanja medicinskih slika, imaju izrazito važnu ulogu u procjeni kardiovaskularnih bolesti. Njima je omogućeno kvantificiranje bolesti, mjerenje volumena kao i analizira morfologije različitih organa. Prema tome, segmentaciju srca i srčanih struktura predstavlja osnovu za širok spektar kliničkih primjena. Primjerice, često se koristi se za modeliranje i analizu anatomije i funkcionalnosti kao i za lokalizaciju različitih patologija. Izrada trodimenzionalnog (3D) modela srca specifičnog za pojedinog pacijenta predstavlja izrazit potencijal za poboljšanje kirurškog planiranja za pacijente s urođenom srčanom manom. Kako bi se takvi 3D modeli mogli izraditi, potrebno je imati segmentirane različite srčane strukture, uključujući pojedine srčane komore, epikardijalnu površinu, aortu kao i pojedine žile kardiovaskularnog sustava. Segmentacija lijeve i desne klijetke ima izrazito važnu ulogu u kvantitativnoj analizi globalnih i regionalnih informacija, odnosno pokazatelja rada srca, poput volumena na kraju dijastole (VKD), volumena na kraju sistole (VKS), frakcije izbacivanja (FI), debljine stijenke ili mase. Primjerice, ventrikularna hipertrofija uzrokovana je abnormalnim povećanjem srčanog mišića koji okružuje lijevu ili desnu klijetku. Prema tome, segmentacija cijelog srca i srčanih komora iz volumetrijskih medicinskih slika igraju bitnu ulogu u procjeni cjelokupnog kardiovaskularnog zdravlja. Nadalje, radiolozi često trebaju ocrtati aortu kako bi dobili njezinu morfologiju, što je bitno za otkrivanje i dijagnosticiranje aneurizme aorte. Ručna segmentacija srca i srčanih struktura je vremenski veoma zahtijevan posao, podložan subjektivnosti. Prema tome, razvoj točnih i robusnih automatskih algoritama za segmentaciju je neophodan za primjenu u kliničkoj praksi.
Duboko učenje predstavlja najsuvremeniju metodu za različite zadatke obrade slike poput raspoznavanja, segmentacije i klasifikacije. Metode dubokog učenja temelje se na umjetnim neuronskim mrežama. Najčešće upotrebljena vrsta neuronske mreže su konvolucijske neuronske mreže (CNN). FCNs predstavljaju specifičnu vrstu CNN-a bez potpuno povezanog sloja, kojima se obrađuje cijela slika te nije potrebno korištenje patcheva. Razvijene su različite varijante FCN-a, od kojih su najznačajnije varijante koje koriste koder-dekoder arhitekture. U biomedicinskoj obradi slika, za segmentaciju, najčešće se koristi U-Net arhitektura neruonske mreže kao i njezina odgovarajuća 3D verzija. U-Net arhitektura ima snažnu reprezentativnu snagu te je u mogućnosti zabilježiti značajke niskih razina što je izrazito važno prilikom treniranja mreže sa malom količinom podataka. Iako U-Net ima snažnu reprezentativnu snagu, dugoročni odnosi između značajki su slabi zbog upotrebe konvolucijskih operacija. Prema tome, potrebno je razvijati naprednije mehanizme kao i dodatne blokove koji će biti u mogućnosti ispraviti nedostatke U-Net arhitekture. Tehnike i blokovi poput veza za preskakivanje ili dubokog nadzora, omogućuju izgradnju dubljih arhitektura neuronskih mreža koje pružaju apstraktnije rezultate učenja te postižu veću točnost prilikom segmentacije medicinskih slika. S obzirom da povećanje broja slojeva osigurava veći prostor parametara koji omogućuje učenje apstraktnijih značajki, dublje arhitekture neuronskih mreža pružaju apstraktnije učenje koje rezultira boljim performanse i većom točnost u zadacima medicinske segmentacije. Unatoč tome, kako se dubina mreže povećava, informacije o gradijentu prolaze kroz mnogo slojeva te mogu nestati ili nakupiti velike pogreške do trenutka kada gradijet dosegne kraj mreže. To dovodi do uobičajenih prepreka treninga dubokih arhitektura neuronskih mreža kao što su problem nestajajućih gradijenta, ekstenzivnog rasta parametara, kao i smanjenja točnosti, što dovodi do računalno zahtjevnih modela.
U ovoj doktorskoj disertaciji, predložen je niz metoda dubokog učenja za automatsku segmentaciju srca i srčanih komora. Fokus disertacije je na poboljšanju metoda dubokog učenja za segmentaciju cijeloga srca, lijeve i desne klijetke i miokarda kao i aneurizme abdominalne aorte. S obzirom na karakteristične probleme koji se javljanju prilikom dizajniranja metoda dubokog učenja za segmentaciju medicinskih slika, poput problema visoke dimenzionalnosti slika koje rezultiraju treniranim modelima s velikim brojem parametara kao i nedostatkom anotiranih podataka za treniranje, cilj ove disertacije je ublažiti navedene izazove predlaganjem novih i robusnih arhitektura neuronskih mreža koje smanjuju broj korištenih parametara, ali zadržavaju izrazito visoku točnost krajnjih rezultata segmentacije.
Prvi i najvažniji znanstveni doprinos predstavlja nova struktura povezivanja rezidualnih jedinica, koju nazivamo rezidualna jedinica za spajanje značajki (FM-Pre-ResNet). FM-Pre-ResNet struktura povezivanja rezidualnih jedinica dodaje konvolucijski sloj na vrh i na dno već postojećih prethodno aktivirajućih rezidualnih jedinica. Pri tome, gornji sloj uravnotežuje parametre dviju grana rezidualne jedinice, dok donji sloj smanjuje dimenzije kanala. Na ovaj način predložena struktura povezivanja rezidualnih jedinica omogućuje kreiranje značajno dubljih modela uz održavanje iste ili čak manje količine parametara u odnosu na originale rezidualne jedinice.
Nakon toga, u drugom znanstvenom doprinosu, predložena je nova 3D arhitektura neuronske mreže bazirana na koder-dekoder arhitekturi koja uspješno integrira FM-Pre-ResNet jedinice s varijacijskim autokoderima (VAE) za segmentaciju srca i srčanih komora iz CT i MRI slika. Metoda se sastoji od tri osnovna dijela. U prvom dijelu, prethodno predložene FM-Pre-ResNet jedinice koriste se za učenje nisko-dimenzionalnog prikaza ulaza u fazi kodiranja. U drugom dijelu, VAE rekonstruira ulaznu sliku iz nisko-dimenzionalnog latentnog prostora, osiguravajući da su sve težine modela snažno regulirane, kako bi se izbjegnula neželjena pojava pretreniranja. VAE dio koristi se samo tijekom treniranja mreže. Konačno, u trećoj fazi dekodiranja ponovno su integrirane FM-Pre-ResNet jedinice pomoću kojih se stvaraju konačne segmentacije srca. Predložena nova arhitektura evaluirana je na testnom skupu podataka koji se sastoji od 40 različitih pacijenata dostupnih kroz MICCAI Multi-Modality Whole Segmentation Challenge (MM-WHS) izazov. Naša metoda ostvarila je prosječni DSC, JI, SD i HD za cijelo srce od 90,39%, 82,24%, 1.1093 i 15,3621 na CT snimkama, odnosno 89,50%, 80,44%, 1,8599, 25,6558 na MRI snimkama. Predloženi pristup ostvario je približno slične rezultate kao i najsuvremenije metode za segmentaciju cijelog srca na CT slikama dok su rezultati na MRI slikama bolji od rezultata prethodno objavljenih najsuvremenijih metoda.
Treći znanstveni doprinos, predstavlja novu automatsku metodu za segmentaciju miokarda (MiO), lijeve (LK) i desne klijetke (DK) iz cineMRI slika. Predstavljena je nova arhitekturu koja integrira SERes blokove u 3D U-net arhitekturu (3D SERes-U-Net). SERes blokovi upotrebljavaju operacije stiskanja i uzbude u rezidualne jedinice. Sposobnost ponovne kalibracije značajki operacija stiskanja i uzbude povećava reprezentativnu snagu mreže, dok ponovna upotreba značajki koristi učinkovito učenje o značajkama, što poboljšava performanse segmentacije. Predloženu metodu evaluirali smo na testnom skupu podataka MICCAI Automated Cardiac Diagnosis Challenge (ACDC). Naša predložena metoda za segmentaciju pomoću 3D SERes-U-Net ostvarila je prosječni DSC za LK, DK i MiO na kraju dijastole od 95%, 90%, 83%. Slično, prosječni DSC za LK, DK i MiO na kraju sistole je 86%, 83%, 85%. Dodatno, izračunati su volumeni LK, DK i MiO na temelju kojih su dalje računate značajne kliničke metrike te su uspoređeni rezultati s referentnim rezultatima. Navedeno uključuje kliničke metrike, odnosno pokazatelje funkcionalnosti srca, uključujući volumen lijeve klijetke na kraju dijastole (VLKKD), volumen lijeve klijetke na kraju sistole (VLKKS), frakciju izbacivanja lijeve klijetke (FILK), volumen desne klijetke na kraju dijastole (VDKKD), volumen desne klijetke na krajnjoj sistoli (VDKKS), frakciju izbacivanja desne klijetke (FIDK), volumen miokarda na krajnjoj sistoli (VMiOKS) kao i masu miokarda na kraju dijastole (MiOKD). Bland-Altman analiza pokazuje visoki koeficijent korelacije od R = 0,99 za VLKKD i VLKKD, dok je R = 0,95 za FILK. Korelacije VDKKD, VDKKS i FIDK su R = 0,97, R = 0,93, R = 0,69. Konačno, R = 0,96 za VMiOKS i R = 0,95 za MiOKD dodatno pokazuju snagu točnosti i preciznosti naše predložene metode.
Konačno, četvrti znanstveni doprinos predstavlja novi automatski pristup za segmentaciju aneurizme abdominalne aorte (AAA). 3D U-Net arhitektura modificirana je uvođenjem rezidualnih jedinica u koder dijelu kao i mehanizmom dubokog nadzora u dekoder dijelu. Kako bi se povećala točnost rezultata, mreža je trenirana i validirana na 19 preoperativnih AAA CTA volumena različitih pacijenata primjenom 4-ostrukog pristupa unakrsne provjere valjanosti. Naša metoda postiže DSC rezultat od 91,03% za segmentaciju aneurizme abdominalne aorte.
Tijekom rada na ovoj doktorskoj disertaciji, objavljeno je 5 radova u časopisima (od čega 3 kao prvi autor), 10 radova objavljeno je na međunarodnim konferencijama (od čega 5 kao prvi autor) te 1 rad kao dio knjige (ko-autor).The term cardiovascular disease (CVD) refers to numerous dysfunctions of the heart and circulatory system. Cardiovascular disease accounts for nearly one-third (33%) of all deaths in the modern world, which is the highest proportion of all diseases. Early diagnosis and appropriate treatment can significantly reduce mortality and improve quality of life. The diagnosis of heart disease is based on the complete cardiovascular picture, including anatomy and physiology. The diagnostic process usually consists of two main parts. The first part refers to obtaining images of the heart using imaging devices. Numerous invasive and non-invasive imaging techniques have been developed to characterize the anatomy and functionality of the heart. The second part of the diagnostic process is the quantification and interpretation of the images using advanced image processing methods. Developing efficient medical image processing and analysis methods is a complex task, mainly because it involves processing large amounts of high-dimensional data. Advances in the development of image processing, computer vision, and artificial intelligence, as well as the widespread availability of powerful graphical processing units (GPUs), have made this challenging task manageable.
Medical image segmentation plays an important role in the as-assessment, diagnosis, and prognosis of various cardiovascular diseases. Extensive research and clinical applications have shown that computed tomography (CT) and magnetic resonance imaging (MRI) play an important role in the non-invasive assessment of cardiovascular disease. They help quantify disease, measure the volume of structures, and analyze organ morphology. Therefore, segmentation of whole heart is an important step for a variety of clinical applications. For example, itis used for modelling and analysing the anatomy and function of the heart and for localizing pathologies. The creation of a patient-specific3D heart model holds excellent potential for improving surgical planning for patients with congenital heart defects. It requires delineation of all cardiac structures, including heart chambers, epicardial surface, entire blood pool, and great vessels. Segmentation of the left and right ventricles plays a critical role in quantitative analysis of global and regional information, i.e., indicators of cardiac function, such as end-diastolic volume (EDV), end-systolic volume (ESV), ejection fraction(EF), wall thickness, and mass. For example, ventricular hypertrophy is caused by abnormal enlargement of the myocardium surrounding the left or right ventricle. Therefore, segmentation of the whole heart and heart chambers from volumetric medical images plays an essential role in cardiac assessment. In addition, radiologists often need to delineate the aorta to obtain its morphology, which is essential for the detection and diagnosis of aortic aneurysms. Manual segmentation of cardiac structures is a time-consuming process that depends on observer variability. Therefore, the development of accurate and robust automatic segmentation algorithms is critical for clinical practice.
Deep learning has emerged as a state-of-the-art method for various image processing tasks such as recognition, segmentation, and classification. Deep learning methods are based on deep artificial neural networks. The most common type of deep neural network is convolutional neural networks (CNNs). Fully convolutional neural networks (FCNs) are a special type of CNNs that do not have a fully connected layer and are trained and applied to the entire image so that no patch selection is required. Several variants of FCNs have been pro-posed to transfer features from the encoder to the decoder to increase segmentation accuracy. The most widely used FCNs for biomedical image segmentation are the U-net architecture and its corresponding three-dimensional counterpart, the 3D U-net architecture. The ability of U-Net architecture to capture low-level features makes them very useful in scenarios with a small amount of training data. Although it has strong representational power, long-range relationships are weak due to the inherent localization of convolutional operations, so more advanced mechanisms and building blocks are required. Techniques and building blocks such as residual connections and deep supervision enable the construction of deeper architectures that provide more abstract learning results and higher accuracy for medical segmentation tasks. The increment in the number of layers provides larger parameter space enabling learning of more abstract features. Therefore, deeper architectures could provide more abstract learning that results in better performance and higher accuracy in medical segmentation tasks. Nevertheless, when the depth of CNN increases, information about the gradient passes through many layers, and it can vanish or accumulate large errors by the time it reaches the end of the network. This leads to common obstacles of training deep neural network architectures such as appearance of vanishing gradients, accuracy degradation, and extensive parameter growth, which results in computationally intensive models.
In this Thesis, we propose a set of deep learning methods for automatic heart and heart chambers segmentation. We focus on improving deep learning segmentation methods for the whole heart, both ventricles, myocardium, and abdominal aortic aneurysm. Several unique challenges and issues arise in developing deep learning methods for medical image segmentation and analysis. For example, the high im-age dimensionality leads to trained models with a high number of parameters, and the lack of expert annotation makes the models more susceptible to overfitting. Therefore, we aim to alleviate these challenges by proposing new and robust CNNs that reduce the number of parameters so that they can be trained with smaller training sets and are less prone to overfitting. One of the most important scientific contributions of this work is the novel connectivity structure of residual units, which we call the feature merge residual unit (FM-Pre-ResNet). The FM-Pre-ResNet unit attaches two convolution layers at the top and at the bottom of the pre-activation residual block. The top layer balances the parameters of the two branches, while the bottom layer reduces the channel dimension. The proposed connectivity allows the construction of notably deeper models while maintaining the same or smaller number of parameters than the pre-activation residual units. Following that, the second scientific contribution is a novel three-dimensional (3D) encoder-decoder architecture that successfully integrates FM-Pre-ResNet units and is additionally guided with variational autoencoders (VAE) for the task of whole heart segmentation from CT and MRI images. The architecture includes three stages. First, in an encoding stage, FM-Pre-ResNet units learn a low-dimensional representation of the input. Second, in the VAE stage, an input image is reduced to a low-dimensional latent space and reconstructs itself to provide a strong regularization of all model weights. This ensures that all model weights are strongly regularized while avoiding overfitting the training data. Third, the decoding stage creates the final whole heart segmentation. We evaluate our method on the 40 test subjects of the MICCAI Multi-Modality Whole Heart Segmentation (MM-WHS)Challenge. Our method achieves an average Dice score (DSC), Jaccard index (JI), surface distance (SD), and Hausdorff distance (HD) for WHS of 90.39%, 82.24%, 1.1093, and 15.3621 on CT images and 89.50%,80.44%, 1.8599, 25.6558 on MRI images, respectively. The proposed approach obtains highly comparable DSC to the state-of-the-art for whole heart segmentation tasks on CT images while outperforming the current state-of-the-art on the MRI images. The third scientific contribution is a new automatic method for left ventricle (LV), right ventricle (RV), and myocardium (Myo) seg-mentation and quantification from cine-MRI images. We introduce a new architecture that incorporates SERes blocks into 3D U-net architecture (3D SERes-U-Net). The SERes blocks incorporate squeeze-and-excitation operations into residual learning. The adaptive feature recalibration ability of squeeze-and-excitation operations boosts the network’s representational power while feature reuse utilizes effective feature learning, which improves segmentation performance. We evaluate the proposed method on the MICCAI Automated Cardiac Diagnosis Challenge (ACDC) testing dataset. Our method obtains an average DSC for LV, RV, and Myo at end-diastole of 95%, 90%, 83%,respectively. Similarly, we obtain an average DSC for LV, RV, and Myo at end-systole of 86%, 83%, 85%, respectively. Additionally, we calculate significant clinical metrics, i.e., indicators of hearts’ function, including volume of the left ventricle at end-diastole (LVEDV), the volume of the left ventricle at end-systole (LVESV), left ventricles’ ejection fraction (LVEF), the volume of the right ventricle at end-diastole(RVEDV), volume of the right ventricle at end-systole (RVESV), right ventricles’ ejection fraction (RVEF), myocardium volume at end-systole(MyoLVES), and myocardium mass at end-diastole (MyoMED). The Bland-Altman analysis shows a high correlation coefficient of R=0.99for LVEDV and LVESV, while R=0.95 for LVEF. Correlations ofRVEDV, EVESV and RVEF are R=0.97, R=0.93, R=0.69, respectively. Finally, R=0.96 for MyoLVES and R=0.95 for MyoMED further show our proposed methods’ strength of accuracy and precision. Finally, the fourth scientific contribution includes a new automatic approach for robust and reproducible abdominal aortic aneurysm(AAA) segmentation. The 3D U-Net network is adapted by introducing residual units in the contracting pathway and a deep supervision mechanism in the expanding pathway. We conduct an ablation study to demonstrate the effect of the addition of residual units and deep supervision for this particular clinical application. To increase the robustness of the results, networks are trained, validated, and evaluated on 19 pre-operative CTA volumes from different patients using a 4-foldcross-validation approach. Our pipeline achieves a Dice score of 91.03%for AAA segmentation. The work conducted during this Thesis resulted in 5 journal publications (of which 3 as the first author), 10 papers are published at international conferences (of which 5 as the first author), and 1publication in book chapters (as co-author)
Improvement of Interlocking Railway System by Using Cyber-Physical Model
No full textSignalno-sigurnosni sustav željeznica je temeljni sustav za upravljanje željezničkim prometom. Stalne promjene u tehnologijama i sigurnosnim zahtjevima prometa rezultirale su brojnim unaprijeđenjima. Danas su glavni oslonci razvoja novih naprednih, pa tako i željezničkih sustava, Internet stvari, oblak računala, umjetna inteligencija, analiza podataka, Industrija 4.0 i kibernetsko-fizikalni sustavi. Željeznice su kao složeni sustav upravljanja svakako zahvaćene utjecajem tih promjena. Ipak, zbog svoje važnosti, načina gospodarenja i obimnosti, željeznice pokazuju visoki stupanj sporosti u promjenama. Ovim specijalističkim radom nastoji se barem jednim dijelom odgovoriti na izazove koje stvara najavljena revolucija u upravljanju industrijskim okruženjima, te promjene standarda ERTMS. Naglasak je stavljen na transformaciju postojećih signalno-sigurnosnih sustava željeznice kao krutih industrijskih komunikacijsko-upravljačkih mreža u prilagodljivije otvorene komunikacijske sustave u smislu razmjene i prikupljanja podataka, te proširenje nadzora stanja infrastrukture i vlaka zbog unaprijeđenja procesa upravljanja prometom i održavanja opreme. U radu je detaljno obrađen nadzor jednokolosječne otvorene pruge te su analizirani problemi koji se na njoj pojavljuju. Glavni cilj analize postojećih sustava rezultirao je izradom kibernetsko-fizikalnog modela signalno-sigurnosnog sustava jednokolosječne otvorene pruge zajedno s potrebnim programskim implementacijama komunikacijskih sučelja. Na temelju uspostavljenog modela provedena je usporedna analiza prednosti takvog načina upravljanja prometom, te tehno-ekonomska analiza investicije koja opravdava uvođenje novog slojevitog pristupa opremanja željezničke infrastrukture s ciljem stvaranja jedinstvene kibernetsko-fizikalne mreže. U radu su prikazane i dodatne analize prednosti koje donosi predloženi kibernetsko-fizikalni model, a koje se očituju u povećanju sigurnosti, pouzdanosti i učinkovitijem održavanjuThe railway signaling and safety system is the basic system for managing rail traffic. Continuous changes in technology and traffic safety requirements have resulted in numerous improvements. Today, the main pillars of the development of new advanced systems, including railway, are the Internet of Things, cloud computing, artificial intelligence, data analytics, Industry 4.0 and cyber-physical systems. Railways, as a complex management system, are certainly affected by the impact of these changes. However, due to their importance, the way they are managed and the size of the railway, they show a high degree of slowness in change. This paper seeks to address, at least in part, the challenges posed by the announced revolution in the management of industrial environments, and changes to the ERTMS standard. Emphasis is placed on the transformation of existing signaling systems as rigid industrial communication and control networks into more adaptable open communication systems in terms of data exchange and collection, and the expansion of infrastructure and train monitoring to improve traffic management and equipment maintenance processes. The paper deals with the supervision of a single-track railway and analyzes the problems that arise on it. The main objective of the analysis of the existing systems has resulted in the development of a cyber-physical model of the single-track interlocking system along with the necessary software implementations of the communication interfaces. On the basis of the established model, a comparative analysis of the advantages of this method of traffic management was carried out, as well as a feasibility analysis of the investment which justifies the introduction of a new stratified approach to equipping the railway infrastructure with the aim of creating a unified cyber-physical network. The paper also presents additional analyzes of the benefits of the proposed cyber-physical model, which are manifested in increasing security, reliability and more efficient maintenanc
The principle of operation of the static electricity meter in the transmission power system
No full textU završnom radu dokumentirana je tema o načelu rada statičnog električnog brojila u prijenosnom
elektroenergetskom sustavu te parametrizacija brojila i postavljanje za rad. Opisane i prikazane su
sve vrste brojila koja se koriste u elektroenergetskom sustavu te njihova načela rada i sheme. Zatim
je na modelu punog imena LANDIS+GYR ZMD405CT44.2409 S3 osim načela rada, strukture i
funkcija opisana njegova parametrizacija i postavljanje istoga za rad u elektroenergetskom
sustavu. Pri parametrizaciji opisuju se glavne mogućnosti i postavke zadanog brojila te korištenje
softvera MAP110 i MAP120.The final thesis documents the topic of the operation principle of a static electric meter in a portable
power system, as well as the meter's parameterization and installation for operation. It describes
and presents all types of meters used in the power system, their operating principles, and schemes.
Then, using the LANDIS+GYR ZMD405CT44.2409 S3 model, it describes the parameterization
and installation of the meter, along with its installation for operation in the power system. With
parameterization, thesis explains the main features and settings of the given meter, as well as the
use of the software MAP110 and MAP120
Control system for autonomous robots with wireless and CAN communication
No full textThis thesis presents the development of a control system for Autonomous Transport Robots (ATR)
which will be used in the Volvo Factories. The research also evaluates the timing and frequency of
the control messages sent over the Controller Area Network (CAN) bus of the ATR. The ATR is a
part of a cyber-physical system having control parameters over Wi-Fi and CAN communication.
The ATR has a distributed control system with two controllers, which work simultaneously and
interdependently. One controller is the center of the decision-control system, and the other
controller handles peripherals, buttons, and sensors. The ATR is developed as a state machine
having different operating states. Each operating state has its own tasks to perform and a certain set
of state transitions. The ATR uses CAN communication to communicate with different nodes
present on the ATR. To provide reliability of the CAN message, an interrupt service routine is
developed for both the controllers. This interrupt service routine helps the controller to ensure that
all the necessary CAN messages arriving on the CAN bus are received by the controller and no
important messages are missed or lost in transmission. For Wi-Fi communication, MQTT Protocol
is used to communicate with the edge controller.
The response time of both the controllers as well as the response time of the ATR is calculated in all
the ATR states and a constant response time of the ATR for every state is proposed. The frequency
of the CAN messages is evaluated and a bus load on the CAN bus is calculated at different baud
rates. A suitable baud rate for the CAN bus on the ATR is proposed.
Overall, this thesis contributes to the development of a robust control system for ATR, ensuring
reliable communication over both CAN and Wi-Fi interfaces. The analysis of response times and
bus loads aids in optimizing the performance and efficiency of the ATR in various operational
scenarios
System for automatic control of the vehicle depending on the condition of the road
No full textU ovom diplomskom radu napravljen je sustav za automatsko upravljanje vozilom u ovisnosti o stanju na cesti. Kao osnova sustava korištena je Nvidia neuronska mreža zvana PilotNET. Zadatak sustava je da mora na odnosu podatka o stanju ceste, predvidjeti brzinu vozila i kut zakreta volana. Potrebno je bilo napraviti potpuno novi skup podataka koji će sadržavati potrebne podatke. Nakon treniranja, razvijene su dvije neuronske mreže koje, u odnosu na pregled stanja na cesti, predviđaju brzinu i kut skretanja. Neuronske mreže su objedinjene u jedan model kako bi se sinkronizirano estimirale vrijednosti zakreta volana i brzine vozila. Evaluacija je provedena u dva koraka. Prvi korak evaluacije je testiranje predviđanja sustava na temelju podataka testnog skupa koji nije korišten u procesu treniranja. Ovim testom dobiven je pregled osnovne funkcionalnosti sustava. Sustav je testiran u virtualnom okruženju preko CARLA simulatora. Ovim testom prikazane su mane i pogreške koje sustav radi u stvarnom vremenskom okruženju. Sustav je postigao autonomnost od 88.7 % i 3.72 grešaka po kilometru.In this thesis, a system was created for automatic vehicle control depending on the road conditions. An Nvidia neural network called PilotNET was used as the basis of the system. The task of the system is to predict the speed of the vehicle and the angle of rotation of the steering wheel based on the data on the road condition. It was necessary to create a completely new data set that would contain the necessary data. After training, two neural networks were developed which, in relation to the road condition overview, predict the speed and the turning angle. Neural networks are combined into one model in order to synchronously estimate the values of steering wheel rotation and vehicle speed. The evaluation was carried out in two steps. The first step of the evaluation is to test the predictions of the system based on the data of the test set that was not used in the training process. This test provided an overview of the basic functionality of the system. The system was tested in a virtual environment using the CARLA simulator. This test shows the flaws and errors that the system makes in a real time environment. The system achieved an autonomy of 88.7 % and 3.72 errors per kilometer
WiFi lights with programmable LEDs
No full textWiFi rasvjeta sa programibilnim svjetlećim diodama je jednostavan IoT sustav koji koristi
ESP8266 mikroupravljački modul za povezivanje na WiFi mrežu. Osim WiFi povezivanjem,
sličan sustav može se postići Bluetooth tehnologijom. Uređaji povezani na mrežu mogu upravljati
trakom svjetlećih dioda koja je spojena sa mikroupravljačem. Zadatak završnog rada bio je izraditi
jedan takav sustav u kojem se putem WiFi mreže upravlja rasvjetnim tijelom. Istraživanjem
trenutnog stanja tehnike zaključeno je što se sve treba napraviti kako bi se realizirao jedan takav
sustav. Dizajnirana je sheme sklopa te su testirane sve komponente i sklop je po potrebi
nadograđivan tijekom izrade. Napisan je kod i testirane su biblioteke koje su po potrebi
zamijenjene. Testiranjem te pozitivnim i negativnim rezultatima dostignut je cilj ovog završnog
rada. Jedina stvar koja je bila preostala je rad upotpuniti spajanjem komponenti sa PCB pločicom.
Time je riješen zadatak završnog rada.WiFi lighting with programmable LEDs is a simple IoT system that uses the ESP8266
microcontroller module to connect to a WiFi network. In addition to WiFi connectivity, a similar
system can be achieved using Bluetooth technology. Devices connected to the network can control
a strip of LEDs that are connected to the microcontroller. The task of the final project was to create
such a system in which lighting is controlled via a WiFi network. Researching the current state of
the technology, it was concluded what needed to be done to make such a system. A circuit diagram
was designed, and all components were tested, with the circuit being upgraded as necessary during
construction. Code was written and libraries were tested and replaced as needed. Testing, both
positive and negative results lead to achieving the goal of this final project. The only thing
remaining was to complete the project by connecting the components to a PCB board. Solving this
task concluded the final project
Application of security mechanisms in the development of network applications for the automotive industry
No full textJedna od najbitnijih stvari kod razvoja svake mrežne aplikacije je briga o sigurnosti mrežne aplikacije. Postoji jako velik broj oblika mrežnih napada i prijetnji koje mogu ugroziti mrežne aplikacije, stoga je primjena odgovarajućih sigurnosnih mehanizama od velike važnosti. Kroz ovaj diplomski rad opisana je važnost provođenja testiranja softvera. Poseban je naglasak stavljen na sigurnosno testiranje softvera. Također opisane su najčešće prijetnje na sigurnost mrežnih aplikacija kao i načini na koje ih je moguće spriječiti. Opisani su i implementirani sigurnosni mehanizmi koje pruža Django radni okvir, te je za potrebe testiranja kreirana mrežna aplikacija pomoću Djanga. Na kraju je provedeno testiranje implementiranih funkcionalnosti kao i sigurnosno testiranje mrežne aplikacije pomoću ZAP alata.One of the most important things in the development of any web application is to take care of the security of the network application. There are a very large number of forms of network attacks and threats that can threaten network applications, therefore the application of appropriate security mechanisms is of great importance. This masterwork describes the importance of software testing. Special emphasis is placed on software security testing. The most common threats to the security of web applications are also described, as well as ways to prevent them. The security mechanisms provided by the Django framework were described and implemented, and a web application using Django was created for testing purposes. At the end, testing of the implemented functionalities was carried out, as well as security testing of the web application using the ZAP tool
PERFORMANCE OF TRANSMISSION COIL IN WIRELESS TRANSMISSION SYSTEM ENERGIES WITH HOMOGENE MAGNETIC FIELD DISTRIBUTION
No full textRazdvajanje frekvencije i neuskladenost impedancije posljedice su promjene faktora magnetske veze izmedu predajne i prijemne zavojnice odnosno
promjene poloˇzaja prijemne zavojnice u odnosu na predajnu zavojnicu u konvencionalnom rezonantnom induktivnom sustavu za beˇziˇcni prijenos energije. Problemi razdvajanja frekvencije i neuskladenosti impedancije u rezonantnom induktivnom prijenosu energije mogu se ublaˇziti generiranjem
homogenog magnetskog polja u ve´cini povrˇsine ravnine punjenja. Homogenost magnetskog polja postiˇze se optimiranjem geometrije predajne zavojnice
i optimiranjem raspodjele struje izmedu namotaja iste zavojnice. U ovoj
disertaciji provodi se postupak odredivanja strukture dvoslojne pravokutne
predajne 3D zavojnice za generiranje homogenog magnetskog polja iscrpnim
pretraˇzivanjem. Mjerenja magnetskog polja na prototipu takve zavojnice
potvrduju homogenu regiju koja ˇcini 55.3% povrˇsine ravnine napajanja. Nakon ˇsto je odredena struktura predajne zavojnice uvodi se pretpostavka da
´ce raspodjela struje izmedu namotaja predajne zavojnice uz optimizaciju geometrije sada poznate strukture pridonijeti poboljˇsanju svojstava predajne
zavojnice. Optimizacija se provodi u dva koraka: optimizacija 2D modela
i optimizacija 3D modela. Razlog tome je smanjenje prostora pretraˇzivanja
147
B. SAZETAK ˇ
odnosno dimenzije problema koji se nastoji rijeˇsiti optimizacijom 3D modela.
Optimalne vrijednosti varijabli odluka iz optimizacije 2D modela (geometrijske varijable drugog sloja zavojnice, omjer struja) koriste se kao konstante
u optimizaciji 3D modela. Optimizacija 2D modela 3D predajne zavojnice
obavlja se u kosimulaciji Python-FEMM uz pomo´c MIDACO alata i genetskog algoritma te diferencijalne evolucije iz pygmo biblioteke. Najbolja
rjeˇsenja fitnes funkcije ostvarena su diferencijalnom evolucijom. Nuˇzna je i
optimizacija u Ansys Maxwell-u zbog toga ˇsto je to programski paket koji
ima opciju simulacije, a takoder i optimizacije 3D modela s pomo´cu alata
Optimetrics tool. Odabire se konstanta udaljenost prijenosa od 30 mm. Tako
optimirana zavojnica ima bolja svojstva u odnosu na prvotnu. Homogena
regija ˇcini 67.78% povrˇsine ravnine napajanja, dubina zavojnice smanjena je
za 28.8% te je postignuta ve´ca srednja vrijednost jakosti magnetskog polja u
ravnini napajanja.Frequency splitting and impedance mismatching are happening due to
magnetic coupling factor variations. i.e. due to relative shift in receiving coil
position with respect to transmitting coil in a conventional resonant inductive
wireless power transfer system. Frequency splitting and impedance mismatching can be mitigated by generating homogeneous magnetic field intensity at
the majority of the charging plane surface. Homogeneity of the magnetic field
intensity are achieved by optimization of the transmitting coil geometry and
current ratio among the coil turns. In this PhD thesis, the process of determining geometry of the 3D rectangular coil with two layers using exhaustive
search is carried out. Measurements that were carried out with prototype
of such optimized 3D coil verified that homogeneous region occupies at least
55.3% of the charging plane surface. Then, after the 3D coil structure is
determined, a hypothesis is introduced. It is supposed that different current
ratio among turns of such coil along with geometry optimization will contribute to achieve better features of known 3D coil structure. Optimization
is carried out in two steps: optimization of the 2D model and optimization
of the 3D model. Main reason to split optimization is to reduce the search
space, i.e. dimension of the problem that is trying to be solved. Optimal va149
B. ABSTRACT
lues of decision variables (variables that represent coil second layer geometry
and current ratio) obtained from 2D model optimization are used as constants in optimization of 3D model. 2D model optimizations are executed in
Python-FEMM co-simulation along with MIDACO tool. Furthermore, genetic algorithm and differential evolution from pygmo library are also applied in
solving 2D model optimization. Best fitness function value is acheived using
differential evolution strategy. Optimization in Ansys Maxwell is necesarry
because of its feature that 3D models can be simulated and also optimized
using Optimetrics Tool of Ansys Maxwell. Transfer distance in both 2D and
3D models is set to be 30 mm. Such optimized coil own better features
compared to coil obtained by exhaustive search. Homogeneous region now
occupies 67.78% of charging plane surface, coil depth is reduced by 28.8% and
also the higher value of average magnetic field intensity at charging plane is
achieved
Platform for remote monitoring and device management
No full textU diplomskom radu su opisani neki od komercijalno dostupnih platformi za upravljanje i nadzor uređaja te je opisana izrada vlastite platforme. Objašnjena je arhitektura sustava, dostupni tipovi podataka za upravljanje uređajima i osnovna konfiguracija koji uređaji moraju slijediti kako bi bili kompatibilni sa sustavom. Poslužitelj je opisan kroz HTTP zahtjeve koje zaprima i websocket poruke koje odašilje korisnicima i uređajima. Android aplikacija je prikazana sa objašnjenjima i slikama svakog ekrana, objašnjeno je kako korisnici mogu upravljati uređajima, dopuštenjima uređaja, svojim korisničkim računom te okidačima.This master's thesis describes some commercially available platforms for device monitoring and control and describes the development of my own platform. The system architecture is explained, along with the available types of data for device management and the basic configuration that devices must comply with to in order to be compatible with the system. The server is described through HTTP requests it receives and WebSocket messages it sends to users and devices. The Android application is presented with explanations and images of each screen, and it is explained how users can manage devices, device permissions, their user account, and triggers
Smart table
No full textU ovom radu opisan je postupak dizajna i izrade pametnog stola kojem je glavna svrha
dodavanje funkcionalnosti na stol za kavu u dnevnoj sobi. Dizajn je izveden iz više gotovih rješenja
te je prvotni model izrađen u SketchUp programskom alatu, dok je nacrt za izradu, uz dimenzije,
izrađen u AutoCad web aplikaciji. Ploče stola su kupljene i izrezane po mjeri u stolarijskom obrtu te
samostalno spojene uz razne drvodjelne alate. Na Raspberry Pi 4 računalo postavljeno u ovom
pametnom stolu instaliran je Raspberry Pi OS operacijski sustav te su svi programi koji upravljaju
funkcionalnošću stola pisani u Python programskom jeziku. Uz to što kao glavnu funkcionalnost nudi
iskustvo računala u stolu, pametni stol upotpunjava i funkcijske aspekte punjača za mobitel, grijača
za kavu te ambijentalnog svjetla uz glazbu. Svim tim dodatnim funkcionalnostima podiže se korisnost
u odnosu na stol za kavu.In this thesis the design and production process of a smart table, whose main purpose iz
enchancing funcionality of a coffee table, is given. Design was derived from multiple ready-made
solutions and te appearance model was made using SketchUp software, while the design plan was
made by hand. Wooden boards were bought and cut to dimensions by a carpentry trade while the
assembly was done independently using woodworking tools. The computer used for this smart table
is Raspberry Pi 4 which runs on Raspberry Pi OS operating system and all the programs that control
the functionality of the table are programmed in Python programming language. In addition to offering
the computer experience in a table, the smart table fulfills the functional aspects of a mobile phone
charger, coffee heater and an ambient light accompanied by music. All these additional functionalities
increase the usefulness compared to a coffee table.
Keywords: smart table, smart home, Raspberry Pi, Python