van Dinter, Raymon
[VerfasserIn];
Chao, Ke
[VerfasserIn];
Leduc, Philippe
[VerfasserIn];
Tekinerdogan, Bedir
[VerfasserIn];
Catal, Cagatay
[VerfasserIn];
Yiping, Sun
[VerfasserIn]
Predicting the Remaining Useful Life of Manufacturing Systems Using Controller Data
Beschreibung:
Context: Predictive maintenance typically relies on large datasets of run-to-failure data from sensors designed for predictive maintenance purposes, like accelerometers and microphones. However, many industrial companies don't keep track of maintenance logs and motion control data, so they often lack records of run-to-failure data or effective sensors. On the other hand, manufacturing companies often have access to controller data from their Programmable Logic Controller (PLC) systems.Objective: This study aims to develop an approach for predicting the Remaining Useful Life (RUL) of manufacturing systems using PLC data.Method: A model pipeline, which uses stationary PLC data to detect degradation was developed in this research. The models evaluated for degradation modelling are based on: Singular Value Decomposition, Dynamic Time Warping, and Long Short-Term Memory Neural Networks (LSTM). The unsupervised models use a single ground truth time window as input and aim to reconstruct the time window using prior knowledge. The hypothesis is that when the models are trained on healthy, normal operations, the model performs well in reconstructing any normal operation data. However, when degradation increases, the deviation will increase as well. As such, the degradation can be measured through the deviation of the model. The degradation measure can also be used for modeling the RUL using an exponential model for our case study.Results: We tested our model pipeline’s capabilities on an industry case study: the of bearing degradation in a water-cooled direct drive rotary motor. As a result, our model pipeline successfully predicted failure up to 11 hours in advance using information from the torque applied to the motor shaft.Conclusion: The experimental results demonstrated that (1) controller data can contain degradation information, (2) it is possible to process the controller data into a health indicator, and (3) the RUL of a motor can be estimated based on the health degradation data. We can estimate the RUL of a motor based on the health degradation data, and our model provides a robust and fast method for doing so. Our model can estimate the RUL up to 11 hours in advance, which is sufficient for manufacturers to organize a servicing event while avoiding serious impact on the production plan