Deployment of a Smart and Predictive Maintenance System in an Industrial Case Study

Abstract

Industrial manufacturing environments are often characterized as being stochastic, dynamic and chaotic, being crucial the implementation of proper maintenance strategies to ensure the production efficiency, since the machines' breakdown leads to a degradation of the system performance, causing the loss of productivity and business opportunities. In this context, the use of emergent ICT technologies, such as Internet of Things (IoT), machine learning and augmented reality, allows to develop smart and predictive maintenance systems, contributing for the reduction of unplanned machines' downtime by predicting possible failures and recovering faster when they occur. This paper describes the deployment of a smart and predictive maintenance system in an industrial case study, that considers IoT and machine learning technologies to support the online and real-time data collection and analysis for the earlier detection of machine failures, allowing the visualization, monitoring and schedule of maintenance interventions to mitigate the occurrence of such failures. The deployed system also integrates machine learning and augmented reality technologies to support the technicians during the execution of maintenance interventions.

Smart retrofitting in maintenance: a systematic literature review

“…This fault estimation includes detecting the presence of failures, isolating the failed component, and identifying the specific failure mode. • Prognosis Assessment (PA): incipient failures can be detected before they can affect the device performance (Aqueveque et al, 2021;Bucci et al, 2020;Alves et al, 2020). This is possible through the estimation of the health state and the prediction of Remaining Useful Life (RUL).…”

“…Improving reliability is often cited as a result of an improved ability to schedule maintenance actions (Priller et al, 2014;Yiu et al, 2019;Catenazzo et al, 2018), and in some cases, as a result of implementing prognostics and health management through retrofitting (Ranjbar et al, 2019;Cattaneo & Macchi, 2019;Vogl et al, 2015). When failures occur, SRM can also reduce the Mean Down Time (MDT), effectively improving maintainability (Sezer et al, 2018;Lesjak et al, 2014;Alves et al, 2020). Besides an improved maintenance schedule, maintainability might be augmented by retrofitting through faster response times (Wang et al, 2020;Liang et al, 2020;Gayathri & Vasudevan, 2018), which enable the use of alarms, earlier detection of faults, and remote notifications.…”

“…When performing CBM, sensor data must be collected before being able to perform analyses. Run-to-failure data is necessary to train CBM models so they can detect the current state of a device (Calabrese et al, 2020;Alves et al, 2020). However, such data is not always available to plants that perform CBM on a device for the first time (Lei et al, 2018).…”

Experiences in Developing Predictive Maintenance Solutions

“…Since the majority of the events are not related to fails, i.e. almost 98% of the original machine events are warnings, resulting in an imbalanced dataset, the model was designed to group events in 5 minutes blocks and thus predict the type of event that may arise in the next 5 minutes (failure or not) [2]. designed to group events in 5 minutes blocks and thus predict the type of event that may arise in the next 5 minutes (failure or not) [2].…”

“…7 represents the results for the training and validation accuracy and loss for the 150 neuron configuration and considering the range up to 30 epochs. designed to group events in 5 minutes blocks and thus predict the type of event that may arise in the next 5 minutes (failure or not) [2]. Fig.…”