Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Unsupervised Machine Learning for Process Optimization

Broms, Erik LU (2025) DABN01 20251
Department of Economics
Department of Statistics
Abstract
Unsupervised learning techniques are gaining traction in industrial applications due to the increasing volume of unlabeled data in manufacturing environments. This study explores whether such methods can help address two inefficiencies identified in a cheese production process: Uncovering hidden operational modes related to machine performance through clustering and detecting one-off anomalies that may indicate unusual or faulty production runs.

The K-Prototypes algorithm revealed that the clustering structure in the dataset was primarily influenced by recipe-related features, offering limited insight into deeper process driven variations. DBSCAN however, when applied separately to each recipe sub-group, was able to isolate potential... (More)
Unsupervised learning techniques are gaining traction in industrial applications due to the increasing volume of unlabeled data in manufacturing environments. This study explores whether such methods can help address two inefficiencies identified in a cheese production process: Uncovering hidden operational modes related to machine performance through clustering and detecting one-off anomalies that may indicate unusual or faulty production runs.

The K-Prototypes algorithm revealed that the clustering structure in the dataset was primarily influenced by recipe-related features, offering limited insight into deeper process driven variations. DBSCAN however, when applied separately to each recipe sub-group, was able to isolate potential anomalies that warranted further investigation in terms of product quality. Additionally, a pipeline combining DBSCAN and autoencoders was proposed as a fully unsupervised approach, using DBSCAN to filter noise and outliers before training, and autoencoder to flag anomalous runs in production.

The findings suggest that unsupervised learning holds promise for early detection of irregular process behavior. However, further validation with larger datasets is necessary to transition from proof-of-concept stage to robust, production-ready solution. (Less)
Please use this url to cite or link to this publication:
author
Broms, Erik LU
supervisor
organization
course
DABN01 20251
year
type
H1 - Master's Degree (One Year)
subject
keywords
Unsupervised Learning, Manufacturing, DBSCAN, Autoencoders
language
English
id
9201439
date added to LUP
2025-09-12 09:03:41
date last changed
2025-09-12 09:03:41
@misc{9201439,
  abstract     = {{Unsupervised learning techniques are gaining traction in industrial applications due to the increasing volume of unlabeled data in manufacturing environments. This study explores whether such methods can help address two inefficiencies identified in a cheese production process: Uncovering hidden operational modes related to machine performance through clustering and detecting one-off anomalies that may indicate unusual or faulty production runs. 

The K-Prototypes algorithm revealed that the clustering structure in the dataset was primarily influenced by recipe-related features, offering limited insight into deeper process driven variations. DBSCAN however, when applied separately to each recipe sub-group, was able to isolate potential anomalies that warranted further investigation in terms of product quality. Additionally, a pipeline combining DBSCAN and autoencoders was proposed as a fully unsupervised approach, using DBSCAN to filter noise and outliers before training, and autoencoder to flag anomalous runs in production. 

The findings suggest that unsupervised learning holds promise for early detection of irregular process behavior. However, further validation with larger datasets is necessary to transition from proof-of-concept stage to robust, production-ready solution.}},
  author       = {{Broms, Erik}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Unsupervised Machine Learning for Process Optimization}},
  year         = {{2025}},
}