Deep learning-based analysis of fuel spray images

doctoral thesis

• Title (croatian): Analiza slika spreja goriva dubokim učenjem
• Author: Fran Huzjan
• Mentor: Sven Lončarić (mentor)
• Committee member: Marko Subašić (predsjednik povjerenstva)
• Committee member: Zoran Kalafatić (član povjerenstva)
• Committee member: Milan Vujanović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-07-18, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing
• Universal decimal classification (UDC): 004 - Computer science and technology. Computing. Data processing

Abstract

Direct rail injection systems in internal combustion engines depend on various parameters, including nozzle diameter, injection pressure, and injector geometry, to shape the spray strategies. These factors influence macroscopic parameters such as spray cone angle, penetration length, and spray area, all crucial for optimizing engine efficiency and reducing pollutants and harmful emissions. To estimate these parameters from spray images, we have utilized computer vision methods, particularly deep learning techniques. In this research, two primary methods were proposed for spray image analysis: spray image segmentation and direct parameter estimation from image sequences. A lightweight segmentation neural network, termed Min U-Net, was developed for the first method. Min U-Net achieved state-of-the-art results while being faster and requiring fewer parameters, outperforming other tested methods in estimating spray macroscopic parameters. For the second method, a regression deep neural network was employed, achieving a more accurate estimate by using a sequence of images as input rather than a single image. An extension was proposed to the second method, incorporating an additional fully connected layer to enhance prediction accuracy. These methods offer promising future research avenues in spray analysis and could help optimize engine efficiency and emission reduction.

Abstract (croatian)

Sustavi s izravnim ubrizgavanjem goriva u motore s unutarnjim izgaranjem ovise o različitim parametrima, uključujući promjer mlaznice, tlak ubrizgavanja i geometriju injektora, kako bi se oblikovale strategije spreja. Ti faktori utječu na makroskopske parametre kao što su kut spreja, duljina penetracije i površina spreja, od kojih su svi ključni za optimizaciju efikasnosti motora i smanjenje zagađenja i štetnih emisija. Kako bi se procijenili makroskopski parametre iz slika spreja, korištene su metode računalnog vida, posebice metode dubokog učenja. U ovom istraživanju predložene su dvije metode za analizu slika spreja: segmentacija slike spreja i izravna procjena parametara iz sekvenci slika spreja. Za prvu metodu razvijena je neuronska mreža niske složenosti za segmentaciju, nazvana Min U-Net. Min U-Net postigao je suvremene rezultate, dok je pritom bio brži i manje složenosti od suvremenih metoda, nadmašivši druge testirane metode u procjeni makroskopskih parametara spreja. Za drugu metodu koristila se regresijska duboka neuronska mreža, postižući točniju procjenu korištenjem sekvence slika spreja umjesto jedne slike kao ulaz. Također je i predloženo proširenje druge metode, uključivanjem dodatnog potpuno povezanog sloja kako bi se poboljšala točnost predviđanja. Ove metode nude obećavajuće pravce budućih istraživanja u analizi spreja i mogu pomoći u optimizaciji učinkovitosti motora i smanjenju štetnih emisija.

• Keywords: Computer Vision; Artificial Intelligence; Neural Networks; Deep Learning; Image Segmentation; Image Analysis; Regression; Spray; Diesel; Macroscopic Parameters
• Keywords (croatian): Račulalni vid; Umjetna inteligencija; Neuronske mreže; Duboko učenje; Segmentacija slika; Analiza slika; Regresija; Sprej; Dizel; Makroskopski parametri
• Language: english
• URN:NBN: urn:nbn:hr:168:190245
• Promotion: 2024
• Study programme: Title: Computer Science; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti (Doktor znanosti)
• Type of resource: Text
• Extent: ix, 77 str. : graf. prikazi ; 31 cm.
• File origin: Born digital
• Access conditions: Open access
• Created on: 2024-02-13 15:23:30