Predictive maintenance of centrifugal pumps: A neural network approach

Abstract

The efficiency of a production line is based on the reliability and correct functioning of the machinery that compose it. However, breakdowns can occur for various reasons and a single machine failure can cause severe production delays and related economic losses. Therefore, the ability to predict a possible fault and correct the cause in time is of paramount importance. This fault management approach is called predictive maintenance and aims to optimise the maintenance schedule and reduce the frequency of failures.

The objective of this research, conducted in collaboration with Neste Oyj, is to create an automated anomaly detection system, capable of predicting failures of centrifugal pumps operating in an industrial environment.

The limited availability of data has required the use of an approach based on a time series forecasting model, capable of predicting the normal behaviour of the machine. The anomalies are then detected calculating the prediction error between the forecasted and the actual values of various sensors. Several machine learning-based solutions were tested to fulfil this purpose, a Multilayer Perceptron (MLP), a Long Short-Term Memory (LSTM) and a Long Short-Term Memory Autoencoder. These models were compared with a statistical approach, a Vector autoregression (VAR), used as a baseline.

All proposed machine learning solutions outperformed the statistical model. The LSTM autoencoder proved to be the best model, achieving an F1 score of 0.986 on the test data.

Furthermore, this research investigated the possibility of using a generalised model capable of identifying faults on different pumps. The experimental results confirmed this hypothesis, allowing to reduce the necessary resources by not requiring a specific model for each pump.