An Automated Sensor Fusion Approach for the RUL Prediction of Electromagnetic Pumps

Abstract

The remaining useful life (RUL) prediction of industrial cyber-physical system components demands the use of reliable prognostics parameters and frameworks. Against the traditional use of a single measure of degradation, data from multiple sensors provide abundant characteristic information for modeling, assessing, and extracting useful parameters via appropriate signal processing and sensor fusion methods. This study introduces a multi-sensor prognostics approach which merges highly prognosible statistical features from vibrational and pressure sensor measurements after a multi-level wavelet decomposition of the signals. The prognostic algorithm presented in this work for solenoid pump RUL prediction is a multi-objective genetic algorithm-optimized long short-term memory (MOGA-LSTM) which accepts the fused sensor features as input and returns the RUL of the pump as output. The framework was tested on a run-to-failure experiment on a VSC63A5 Solenoid pump following a significant pump malfunction caused by a clogged suction filter after the test. Using standard prognostic performance evaluation metrics, the performance of the prognostics framework was compared with other reliable state-of-the-art methods with a remarkable comparative advantage in addition to better automation potentials for real-time condition monitoring and RUL prediction.

A CEEMDAN-Assisted Deep Learning Model for the RUL Estimation of Solenoid Pumps

“…Ongoing research on reliability studies suggests that hydraulic components are some of the most complex systems to accurately monitor and has motivated diverse prognostic and health management (PHM) methodologies for these components-pumps, valves, actuators, etc. This complexity is a result of the thermodynamic, fluid dynamic, mechanical, and electrical processes (not to mention uncertainties) which simultaneously occur as the components are in operation [13]; nonetheless, vibration monitoring, acoustic emission monitoring, and pressure monitoring are among the most popular methods for pumps [4,14].…”

“…In previous studies on failure diagnostics of electromagnetic pumps [4,27,28], filter clogging has shown to be one of the high-ranking failure modes. In this study, a natural running condition (powered by 220 V, 60 Hz) was simulated for a VSC63A5 solenoid Pump whereby the working fluid-5 Litres of diesel was contaminated with ten(10) grams of Iron(III) oxide (Fe 2 O 3 ).…”

“…The complexity and predictive performance of the developed model was evaluated with the earlier listed prognostics evaluation metrics in comparison with other ML-based estimators-multi-objective genetic algorithm-optimized long short term memory (MOGA-LSTM) [4], deep belief network (DBN) [22], and a 3-layer deep neural network (DNN) which had been earlier deployed/developed for the same purpose in our past study. It is worth noting that these algorithms shared a similar architecture like the proposed method for fair comparison since it would be futile to explore/optimize different variants/architectures for the respective models.…”

“…This presents ample opportunities for developing (and improving) high-performing generalized models for accurate failure diagnostics and prognostics, even for unseen equipment working conditions [3]. Particularly for the remaining useful life estimation of equipment/components, several articles have recorded various successes in the use of DL methods regardless of the types and architecture of the artificial neural network (ANN) employed [4].…”

“…Generally, data-driven predictive maintenance relies significantly on the availability and prognosibility of sensor data. This usually demands the use of multiple sensors and/or multiple sensor-extracted features (in the case of a single-sensor situation) [4]. Because most sensor measurements are non-stationary (with some background noise), signal denoising is almost a necessary pre-processing step to ensure that noise and/or insignificant spectral components of the signals are excluded from the feature extraction and dynamic modelling process.…”