Hybrid Neural Network and Physics-Based Digital Twins for Condition-Based Maintenance

This paper presents a case study of the challenges faced in developing a reliable, robust, and accurate digital twin system for an automotive shock absorber. Specifically, this digital twin system’s role is to estimate the current gas pressure in the reservoir chamber and compare it with the expected pressure. It is quantitatively demonstrated that design choices of sensors and algorithms have a significant effect on the accuracy of the system which is not proportionate to the hardware costs of the digital twin system. The evaluated sensor suites cost is significant, with an overall cost ranging from A$297 to A$4,292, representing a 14-fold difference in costs. The study shows that the use of an expansive and costly sensor suite does not necessarily reflect proportionately in the accuracy of the system. The algorithms and sensors utilised in the digital twin architecture have a significant effect on the accuracy of the system with the RMSE ranging from 3.83 Bar to 0.85 Bar, a four-fold variation in accuracy. The digital twin approach showed significant benefit in accuracy highlighted by the most accurate sensor only approach achieving a RMSE of 2.27 compared to the 0.84 of the full digital twin approach. The lowest cost system which maximally utilised Bayesian methods and physical modelling generated the second most accurate estimate with a RMSE of 1.4 Bar, 165% of the most accurate system, which is still effective for the task, but at 7% of the cost. This demonstrates that by leveraging algorithmic development in a hybrid architecture, performance can be significantly improved and both dataset sizes and training times for the neural network components can be significantly reduced.