Napovedovalna analiza ravnanja kaljenih kovinskih obdelovancev : doktorska disertacija

Peršak, Tadej

| | SLO | ENG | Cookies and privacy

Bigger font | Smaller font

First page > Show document

Show document

Title:	Napovedovalna analiza ravnanja kaljenih kovinskih obdelovancev : doktorska disertacija
Authors:	ID Peršak, Tadej (Author) ID Klančnik, Simon (Mentor) More about this mentor... ID Vuherer, Tomaž (Comentor) ID Belšak, Aleš (Comentor)
Files:	DOK_Persak_Tadej_2023.pdf (9,37 MB, This file will be accessible after 26.09.2026) MD5: 258FC5AFD0B1BFBC5A11E84EC64C5494
Language:	Slovenian
Work type:	Doctoral dissertation
Typology:	2.08 - Doctoral Dissertation
Organization:	FS - Faculty of Mechanical Engineering
Abstract:	V industriji se kovinski obdelovanci pogosto toplotno obdelujejo z namenom poboljšanja njihovih mehanskih lastnosti, pri čemer pa se pojavljajo neželene deformacije njihove geometrije. Zaradi dosežene visoke trdote (60 HRC ali več) klasični pristopi ravnanja z upogibanjem in valjanjem niso učinkoviti, saj se material poruši. V ta namen smo se v okviru doktorske raziskave ukvarjali z analizo vpliva plastičnih površinskih deformacij na spremembe geometrije kaljenih kovinskih obdelovancev. Izveden je bil laboratorijski eksperiment, v katerem smo na podlagi nadzorovanega vnosa površinskih plastičnih deformacij, visokoresolucijskega zajema geometrije obdelovancev, merjenja pospeškov ter zajema zvoka ravnalnih udarcev (udarci, ki plastično deformirajo površino obdelovanca) pripravili bazo podatkov s 3063 vzorci. Dodatno smo zajemali zvočne odzive nadzorovanih udarcev, ki ne deformirajo površine obdelovanca. Z uporabo U-Net nevronske mreže smo razvili model za napovedovanje spremembe geometrije kaljenega kovinskega obdelovanca glede na vnesene plastične površinske deformacije. V nadaljevanju smo predlagali novo arhitekturo globoke konvolucijske mreže za regresijo, ki omogoča dva vhoda različnih podatkovnih tipov in dimenzij (zvok ravnalnega udarca in reprezentacija geometrije obdelovanca z vključenimi podatki o ravnalnih udarcih) ter večdimenzionalni izhod (napovedana sprememba geometrije obdelovanca). Prav tako smo z uporabo globoke nevronske mreže razvili model, ki izključno na podlagi zvoka nadzorovano izvedenih udarcev, ki ne deformirajo površine obdelovanca, učinkovito napove geometrijo kovinskega obdelovanca. Uspešnost razvitih napovednih modelov smo ocenili z relativno absolutno napako (angl. relative absolute error (RAE)), povprečno kvadratno napako (angl. root mean squared error (RMSE)) in relativno kvadratno napako (angl. relative squared error (RSE)). Najuspešnejši model za napovedovanje oblike obdelovanca je imel na testnih podatkih odlično zmogljivost napovedovanja, saj so povprečne vrednosti RAE znašale 0,0499, RMSE 0,0129 in RSE 0,0040. Pri vključitvi zvoka v napovedovalni model so vrednosti RAE znašale 0,0739, RMSE 0,0185 in RSE 0,0075. Pri napovedi oblike obdelovanca samo iz zvoka pa so povprečne vrednosti RAE znašale 0,7439, RMSE 0,1744 in RSE 0,5638.
Keywords:	proces ravnanja, kaljeni obdelovanec, proizvodnja, konvolucijska nevronska mreža, modeliranje, oblak točk, zvok
Place of publishing:	Maribor
Place of performance:	Maribor
Publisher:	[T. Peršak]
Year of publishing:	2023
Number of pages:	XII, 118 str.
PID:	20.500.12556/DKUM-84173
UDC:	004.8.02:621.983(043.3)
COBISS.SI-ID:	169172739
Publication date in DKUM:	06.10.2023
Views:	632
Downloads:	0
Metadata:
Categories:	KTFMB - FS
:	Copy citation

Average score:	(0 votes)
Your score:	Voting is allowed only for logged in users.
Share:

Hover the mouse pointer over a document title to show the abstract or click on the title to get all document metadata.

Licences

License:	CC BY-NC-ND 4.0, Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International

Link:	http://creativecommons.org/licenses/by-nc-nd/4.0/
Description:	The most restrictive Creative Commons license. This only allows people to download and share the work for no commercial gain and for no other purposes.
Licensing start date:	25.04.2023

Secondary language

Language:	English
Title:	Predictive analytics for hardened metal workpieces straightening
Abstract:	In industry, metal workpieces are often heat-treated to improve their mechanical properties, but it also leads to undesirable deformations of their geometry. Due to the high hardness achieved (60 HRC or more), conventional straightening processes like bending and rolling treatment are not effective, since failure of the material occurs. In this regard, the influence of plastic surface deformation on the geometry changes of hardened metal workpieces was analyzed as part of the doctoral thesis. A laboratory experiment was carried out, creating a database of 3063 samples based on controlled input of plastic surface deformations, high-resolution acquisition of the workpiece geometry, acceleration measurements, and acoustic detection of straightening strikes (strikes that plastically deform the workpiece surface). In addition, we captured the sounds of controlled strikes that do not permanently deform the surface of the workpiece. Using a U-Net neural network, we developed a model to predict the geometry change of a hardened metal workpiece as a function of the applied plastic surface deformations. In addition, we proposed a novel deep convolutional network architecture for regression that allows two inputs of different data types and dimensions (the sound of the straightening strike and the representation of the workpiece geometry with the straightening strike data contained in it) and a multidimensional output (the predicted change in the workpiece geometry). We have also developed a deep neural network model that effectively predicts the geometry of a metal workpiece based only on the sound of controlled strikes that do not deform the surface of the workpiece. The performance of the developed prediction models was evaluated using the relative absolute error (RAE), root mean square error (RMSE), and relative squared error (RSE). The best model for predicting the shape of the workpiece had excellent prediction performance on the test data, with average RAE, RMSE, and RSE values of 0.0499, 0.0129, and 0.0040, respectively. When sound was included in the prediction model, the RAE, RMSE, and RSE values were 0.0739, 0.0185, and 0.0075, respectively, but when the shape of the workpiece was predicted using sound alone, the average RAE, RMSE, and RSE values were 0.7439, 0.1744, and 0.5638, respectively.
Keywords:	Straightening Process, Hardened Workpiece, Manufacturing, Convolutional Neural Network, Modeling, Point Cloud, Sound

Comments

Leave comment

You must log in to leave a comment.

Comments (0)

0 - 0 / 0

There are no comments!

Back