Predictive maintenance based on event-log analysis

Kljun, Maša (author), Demšar, Jure (mentor), D’Alconzo, Alessandro (comentor)

The success of manufacturing companies is highly reliant on the performance of their machinery. Unplanned downtimes of the machines may cause severe profit loss, so it is important to prevent such events. One of the ways of achieving an undisturbed manufacturing process is with predictive maintenance, which allows factories to switch from reactive to proactive action taking by predicting a machine failure before it even happens. Predictive maintenance can be performed by utilizing sensor measurements, however, there are a lot of costs associated with installing and maintaining new sensors. A promising and more cost-friendly alternative is predictive maintenance based on machine logs. In this thesis, to tackle the problem of failure prediction, we performed a thorough literature review and found two suitable state-of-the-art approaches that utilize machine logs. We implemented both approaches and made several improvements. To better assess the quality of both approaches, we created 6 toy data sets, each with its own data-generating process and complexity. Our results on the toy data show that we can achieve decent results on data with none or some random noise. Yet, on the real-world data both approaches performed poorly which suggests that the machine logs at hand are weakly related to the failures and as such are not informative enough for successful failure prediction.

• Language: English
• Keywords: predictive maintenance, machine logs, maintenance logs, time series
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2022
• PID: 20.500.12556/RUL-138469
• COBISS.SI-ID: 120316675
• Publication date in RUL: 22.07.2022

Secondary language

• Language: Slovenian
• Title: Napovedno vzdrževanje na podlagi analize dnevnika dogodkov

Abstract

V tovarnah je zelo pomembno, da proizvodnja nemoteno teče, saj lahko vsaka nenapovedana prekinitev proizvodnje povzroči nezaželene stroške. Ena izmed možnih rešitev za doseganje nemotenega delovanja strojev je napovedno vzdrževanje, ki omogoča, da se lahko napako na stroju napove vnaprej in se prepreči, da se stroj pokvari. Napovedno vzdrževanje se lahko izvaja na podagi meritev tipal ali pa na podlagi dnevnika dogodkov. Prednost slednjega je, da so navadno dnevniki dogodkov prosto dostopni in pridobitev teh podatkov ne predstavlja dodatnih stroškov nameščanja tipal. V tem magistrskem delu smo za reševanje problema napovedovanja napake izvedli pregled literature in identificirali dva pristopa. Oba pristopa smo implementirali in predlagali številne izboljšave. Da bi bolje ocenili oba pristopa, smo generirali 6 umetnih podatkovnih množic, vsako s svojo stopnjo kompleksnosti in naključnosti. Opazili smo, da en pristop deluje dobro v primeru enostavnih podatkov, medtem ko v primeru naključnega šuma, pričakovano, noben pristop ne deluje dovolj dobro. Analizo smo izvedli tudi na realnih podatkih, pridobljenih s strani podjetja Siemens. Na realnih podatkih žal noben pristop ni dosegel dobrih rezultatov, saj so le-ti zelo podobni naključnemu šumu in kot takšni niso dovolj informativni za uspešno napovedovanje napake.

• Keywords: napovedno vzdrževanje, dnevnik dogodkov, dnevnik vzdrževanj, časovne vrste