Predictive maintenance using clustering methods for the use-case of bolted connections in the automotive industry

Date
2023-03
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: The rise of Industry 4.0 and the implementation of “smart machines” have opened new opportunities to utilise all the data gathered from machines in an automotive manufacturing environment. A big obstacle faced by maintenance staff is understanding the data retrieved from the machines as the volume is often overwhelming and needs to be processed to be fully useful. The most prevalent model found in predictive maintenance is data-driven and based on statistical process control, pattern recognition, or machine learning algorithms. This thesis explores the feasibility of predictive maintenance, using clustering, specifically for the use case of bolted connections using electronically controlled nutrunners in a modern, high-volume manufacturing environment. The data for this study is collected from the wheel bolting machine of an automotive factory and consists of process parameters already recorded by the current system. The data retrieved from the system is unlabeled, time-series based and records the torque, angle, and time from the bolting process. During this period, the failure rate of one nutrunner was significantly higher than the others. After changing the socket, the machine showed a 2% improvement and the bolting graphs returned to the expected format. This thesis aims to retrospectively establish if the failure was predictable from the data by using clustering algorithms. The study includes a critical analysis of the mechanical machine-train of the nutrunner system based on literature and domain knowledge. A failure analysis is done to understand the key characteristics of common failures as identified on the bolting process curves. Due to the format of the data, a comprehensive data exploration phase had to be conducted to find and understand the outliers and data quality. Furthermore, the required features for the dataset used during modelling had to be designed. The clustering algorithms investigated are agglomerative hierarchical clustering (AHC), density-based spatial clustering of applications with noise (DBSCAN) and a selforganising feature map (SOFM). Each algorithm underwent extensive parameter optimisation and fine-tuning in order to establish the best clusters. The performance metrics used to compare and evaluate the clusters were the silhouette coefficient score (SC) and variation rate criterion (VRC). The SOFM clustering performed the best and the resulting clusters were used to further perform clustering analysis. The cluster analysis showed promising results as the clusters were well-defined and decipherable using domain knowledge. The results of this thesis show that it is feasible to use clustering to improve the maintenance strategy of nutrunners in the automotive industry.
AFRIKAANSE OPSOMMING: Met die opkoms van die vierde industriële rewolusie en die inspan van “slim masjiene” het nuwe geleenthede oopgegaan om alle data wat van masjiene in die motorvervaardigingsektor ingesamel word, te ontgin. ’n Groot struikelblok waarvoor operasionele personeel te staan kom, is die dikwels oorweldigende omvang van die data wat van die masjiene verkry is en geprosesseer moet word om ten volle nuttig te wees. Die mees algemene model wat gebruik word in voorspellende instandhouding is data-gedrewe en gebaseer op statistiese prosesbeheer, patroonherkenning of masjienleer tegnieke. Hierdie tesis ondersoek die werkbaarheid van voorspellende instandhouding deur gebruik te maak van groepering, spesifiek vir toepassing in ’n moderne, hoë volume vervaardigingsomgewing. Die data vir die studie is versamel van ’n masjien wat wiele monteer in ’n motorvervaardigingsfabriek en is parameters wat die masjien reeds genereer. Die data is ongemerk, tyd-reeks gebasseer en neem die wringkrag, hoek en tyd van die aanhegting proses op. In die tydspan van die ondersoek is bevind dat die falings van een moeraanjaer beduidend hoër vertoon as dié van die ander. Toe die sok vervang is, het die masjien ’n verbetering van 2% getoon en die grafiek het na die normale formaat teruggekeer. Die tesis mik om te bepaal of die meganiese faling van die moeraanjaers beter voorspel kan word deur die gebruik van groeperings algoritmes. Die studie sluit ook ’n kritiese analiese in van die meganiese komponente van die moeraanjaer wat gebaseer is op veldkennis en literatuur. ’n Falings-analise is gedoen om die sleuteleienskappe van algemene falings te verstaan soos gesien op aanhegtingsgrafieke. Weens die aard van die data is omvattende data eksplorasie gedoen om die uitskieters te vind en te verstaan. Daarna is die nodige kenmerke vir verdere modellering. Agglomeratiewe hiërargiese groepering (AHC), density-based spatial clustering of applications with noise (DBSCAN) en selforganiserende afbeelding (SOM) groepering algoritmes is gebruik. Elke algoritme se parameters is ge-optimiseer en verfyn om die beste groeperings te vorm. Die resultate is geëvalueer op grond van die inter- en intragroepering afstande en die bes presterende algoritme is gekies vir verdere analiese. Die SOFM algoritme het die beste presteer en is verder gebruik vir groepontleding. Die groepontleding het belowende resultate gewys omdat die groepe goed-gedefinieerd en ontleedbaar met die gebruik van veldkennis is. Die resultate bevestig dat dit haalbaar is om ’n groeperings algoritme benadering te gebruik as hulpbron om die instandhouding plan van moeraanjaers te verbeter.
Description
Thesis (MEM)--Stellenbosch University, 2023.
Keywords
Citation