Fault Detection in Supermarket Refrigeration Systems Using Machine Learning
Translated title
Fejldetektion i Supermarkedskølesystemer ved hjælp af Maskinlæring
Authors
Sathiyaseelan, Gowsikan ; Martin Garcia, Antonio Javier
Term
4. term
Education
Publication year
2021
Submitted on
2021-06-03
Pages
109
Abstract
Supermarkeders køleanlæg holder varerne friske ved at sikre stabile temperaturer. Fejl i udstyret kan give madspild og økonomiske tab. Løbende manuel overvågning er upraktisk og dyr, så dette speciale undersøger automatisk fejldetektion for at vurdere anlæggenes tilstand og forudsige fejl, før de påvirker driften. Vi har udviklet en Support Vector Machine (SVM), en maskinlæringsklassifikator der skelner mellem mønstre, trænet på simulerede data til at genkende systemtilstande i realtid. For at lette klassifikationen reducerede vi antallet af variabler med Principal Component Analysis (PCA) og Linear Discriminant Analysis (LDA), to metoder der komprimerer data og bevarer de mest informative træk; LDA søger samtidig at adskille klasserne bedst muligt. Metoden er testet under forskellige begyndelsesbetingelser og fejlscenarier. Resultaterne viser, at LDA er den bedste metode til dimensionsreduktion i denne anvendelse, og at SVM er en anvendelig løsning til automatisk fejldetektion med tæt på 100 % nøjagtighed i de fleste simulationstests.
Supermarket refrigeration keeps food safe by holding steady temperatures. When equipment fails, it can cause food spoilage and financial losses. Constant manual monitoring is impractical and costly, so this thesis explores automated fault detection to determine system status and predict failures before they disrupt operation. We developed a Support Vector Machine (SVM), a machine-learning classifier that separates patterns, trained on simulated data to recognize system states in real time. To make the classifier’s job easier, we reduced the number of variables using Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA), two techniques that compress data while retaining the most informative features; LDA also seeks to separate classes as clearly as possible. We tested the approach under a range of initial conditions and fault scenarios. Results show that LDA is the better choice for dimensionality reduction in this application, and that the SVM is a practical solution for automatic fault detection, achieving close to 100% accuracy in most simulation tests.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
Documents