Predictive Maintenance from Event Logs Using Wavelet-Based Features: An Industrial Application [chapter]

Stéphane Bonnevay, Jairo Cugliari, Victoria Granger
2019 Springer Reference Sozialwissenschaften  
In industrial context, event logging is a widely accepted concept supported by most applications, services, network devices, and other IT systems. Event logs usually provide important information about security incidents, system faults or performance issues. In this way, the analysis of data from event logs is essential to extract key informations in order to highlight features and patterns to understand and identify reasons of failures or faults. The objective is to help anticipate equipment
more » ... ilures to allow for advance scheduling of corrective maintenance. In this paper, we address the problem of fault detection from event logs in the electrical industry. We propose a supervised approach to predict faults from an event log data using wavelets features as input of a random forest which is an ensemble learning method. This work was carried out in collaboration with ENEDIS, the distribution operator of the electrical system in France. 14 We believe that event logs could be processed and analysed to unveil useful information, in 15 addition to devices' primary data. More precisely, we assume that these data can be useful to inform 16 about the device's operative state and eventually to predict device failure. However, event logs 17 concern a wide range of uses and the diculty comes from the volume and variety of logs received. 18 Log events are continuously recorded composing a data streamow related with high volumes, as 19 being generated not only for irregular functional conditions, but also for normal operative states. 20 The main challenge is to analyse this data and extract useful knowledge from the unremitting ow 21 of notications. The issue therefore is to identify appropriate events containing helpful information. 22 Furthermore, it is essential to detect a shift or an alteration in the patterns of these specic events 23 which could alert users about a fault occurrence. 24 In literature, patterns from event logs are dened in various ways, for example as partial orders 25 of a process 1,2,3 , or considered as Petri nets 4 . Also as repeated sequences that capture process 26 models from event logs in order to improve their detection 5 . From these denitions, authors develop 27 some specic pattern detection approaches mainly based on unsupervised or supervised learning 28 1 techniques. Unsupervised pattern detection approaches take an event log as input and generate 29 patterns based on statistical properties 2,3,5 . In unsupervised learning, clustering techniques are 30 widely used 6,7 . Supervised pattern detection approaches take patterns and logs as input and detect 31 pattern instances as results 4 . Combination of these two approaches into semi-supervised techniques 32 have been also studied 8 . From another point of view, visualization and interactive tools have been 33 developed to help user observe and analyse both patterns and event sequences, as EventFlow 9 . 34 Event logs are frequently composed of event codes and their associated text messages. In that 35 case, the use of text parsing or natural language processing techniques is necessary 6,10 . 36
doi:10.1007/978-3-030-20055-8_13 dblp:conf/softcomp/BonnevayCG19 fatcat:gqjyi5zfanerjklrmckhslmtpi