Show document

A+ | A- | | SLO | ENG

Title: Umetna inteligenca v internetu stvari: od zajema podatkov do napovedovanja dogodkov : diplomsko delo visokošolskega strokovnega študijskega programa prve bolonjske stopnje Spletne in informacijske tehnologije Authors: ID Adam, Christian (Author) ID Kokot, Tomaž (Mentor) More about this mentor... Files: Adam_Christian_dd_2026.pdf (2,18 MB) MD5: AA05340091E5E2C6B9F54607028AE0B9   Language: Slovenian Work type: Bachelor thesis/paper Typology: 2.11 - Undergraduate Thesis Organization: UAMEU - Alma Mater Europaea University Abstract: Diplomsko delo obravnava uporabo umetne inteligence pri napovedovanju degradacijskih procesov v okoljih interneta stvari. V raziskavi sem želel preveriti, ali kompleksnejši sekvenčni modeli, kot je LSTM, v praksi res prinašajo boljše napovedne rezultate, ali pa so lahko enostavnejši modeli, na primer linearna regresija, v določenih primerih primerljivejši in stabilnejši. Empirični del vključuje tri medsebojno povezane eksperimente. Prvi eksperiment temelji na simuliranem, pretežno linearnem signalu, kjer razlike med modeloma niso bile izrazite (RMSE: linearna regresija 0,0551; LSTM 0,0563 v normaliziranem prostoru pri oknu 50). V drugem eksperimentu je bil uporabljen signal z izrazito režimsko spremembo, ki je povzročila odstopanje testnih podatkov od učne porazdelitve. V tem primeru se je pokazalo, da se lahko napaka modela LSTM močno poveča (RMSE: linearna regresija 0,0905; LSTM 0,7812), medtem ko linearna regresija ohrani stabilnejšo napovedno uspešnost. Tretji eksperiment temelji na realnem industrijskem podatkovnem naboru NASA C-MAPSS (FD001), kjer je LSTM dosegel zmerno prednost pri napovedovanju preostale življenjske dobe v ciklih (linearna regresija: MAE 13,67; RMSE 16,29; LSTM: MAE 13,31; RMSE 15,73). Rezultati raziskave kažejo, da večja kompleksnost modela sama po sebi ne zagotavlja boljše napovedne uspešnosti. Pomembno vlogo imata predvsem struktura podatkov in ustrezna metodološka zasnova eksperimentov, ki vključuje kronološki razrez podatkov, preprečevanje prenosa informacij med učnim in testnim delom, skaliranje samo na učni množici ter dosledno poročanje metrik v pravilni skali. Pri analizi realnih podatkov se je dodatno pokazalo, da morajo biti metrike in grafični prikazi vezani na iste napovedi po povratni transformaciji, sicer lahko pride do neskladij med programsko implementacijo in poročilom rezultatov. Keywords: internet stvari (IoT), umetna inteligenca, globoko učenje, LSTM, linearna regresija Place of publishing: Maribor Place of performance: Maribor Publisher: C. Adam Year of publishing: 2026 Year of performance: 2026 Number of pages: 36 str., [8] f. pril. PID: 20.500.12556/ReVIS-14088  COBISS.SI-ID: 282488067  UDC: 004.8:004.738.5(043.2) Publication date in ReVIS: 22.06.2026 Views: 38 Downloads: 1 Metadata:    Cite this work Plain text BibTeX EndNote XML EndNote/Refer RIS ABNT ACM Ref AMA APA Chicago 17th Author-Date Harvard IEEE ISO 690 MLA Vancouver : Copy citation

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

Secondary language

Language:	English
Abstract:	This thesis investigates the application of artificial intelligence in predicting degradation processes within Internet of Things (IoT) environments. The primary objective of the study was to determine whether more complex sequential models, such as Long Short-Term Memory (LSTM) networks, consistently provide better predictive performance than simpler models like linear regression, or whether simpler approaches can be equally reliable and more stable in certain scenarios. The empirical part consists of three interconnected experiments. The first experiment was based on a simulated predominantly linear signal, where differences between the models were minimal (RMSE: Linear Regression 0.0551; LSTM 0.0563 in normalized space with a window size of 50). The second experiment involved a signal with a pronounced regime change that introduced a distribution shift between the training and test sets. In this case, the prediction error of the LSTM model increased significantly (RMSE: Linear Regression 0.0905; LSTM 0.7812), while linear regression maintained more stable predictive performance. The third experiment used real industrial data from the NASA C-MAPSS (FD001) dataset, where the LSTM model showed a moderate advantage in predicting Remaining Useful Life measured in cycles (Linear Regression: MAE 13.67; RMSE 16.29; LSTM: MAE 13.31; RMSE 15.73). The results indicate that increased model complexity does not automatically guarantee better predictive accuracy. Predictive performance largely depends on data structure and appropriate experimental design, including chronological data splitting, prevention of data leakage, scaling performed exclusively on the training set, and consistent reporting of evaluation metrics. In the analysis of real-world data, it was further observed that performance metrics and graphical results must be based on identical predictions after inverse transformation, otherwise inconsistencies between implementation and reported results may occur.
Keywords:	internet of things (IoT), artificial intelligence, deep learning, LSTM, linear regression

Back