Predictive Maintenance of Pumps at ‘Den Magiske Fabrikken’, Using Machine Learning Techniques
Keywords:Machine learning, Predictive maintenance, Long short-term memory, Support vector machine, Naïve bayes, Principle component analysis, Progressive cavity pump
In this work, we investigate machine learning methods to predict the failures of progressive cavity pumps (PCP). The PCPs are located in a biogas plant, Den Magiske Fabrikken, in Norway, which is transforming food waste and animal manure to biogas and biofertilizer. Available measurements were pump on-signal, speed, current, torque and control signal, inlet flow, inlet pressure and outlet pressure, and several vibrations derived signals.
Five categories were defined to categorize the operation of the pumps as: “stopped”, “normal running”, “7 days from failure”, “1 day from failure” and “1 hour from failure”. The objective was to train a Machine Learning model to predict these categories. The data was pre-processed to clean gross outliers and scale the signals using different techniques.
This paper presents results from the same Long Short-Term Memory (LSTM) model using two different approaches for scaling the data. The results are evaluated using confusion matrices where one scaling method clearly improves the results when testing on new data points. Further work is presently being carried out to implement the selected methods in real-time and to generalize the model (Holm, 2022).
Copyright (c) 2022 Martin Holm, Ozgur Yalcin, Carlos Pfeiffer, Håkon Viumdal
This work is licensed under a Creative Commons Attribution 4.0 International License.