Coming Soon: Fast and Accurate Modelling of Hydraulic Control Systems Using Machine Learning

While our usual schedule of conferences across the world is reviewed, our research and development programmes continue. Ongoing research, due for release at CTI Symposium USA, includes the application of machine learning for fast and accurate modelling in a variety of applications, including for hydraulic system development.

You can receive regular updates from us (such as our progress on this research) by signing up to our ‘Research Reviews’ here.



The research presents a methodology which saves time and expenditure in modelling automotive systems. The objective of the research is to use this model of the system in the development phase of prototype hardware. A hydraulic actuation system was chosen for this study due to the inherent non-linearities in the dynamic behaviour of such systems. Traditional physics-based simulations of hydraulic systems are typically limited with respect to their accuracy and computational cost.

The research proposes a fast and accurate method using machine learning concepts to train a correlated model of an automotive valve body. With experience in hydraulic design and development, a technique is devised to categorise the different behaviours of the system, and the model is trained using gaussian process regression (GPR).

For different behaviour, labels are generated using support vector classifier (SVC) and then individual GP sub-models are trained for each label. The algorithm was trained and tested using data that we’ve collected during the testing of a DSD-designed valve body on an in-house valve body test stand.

The proposed methodology generated an efficient and accurate prediction model. The method achieves an efficient and accurate plant model that is capable of predicting the hydro-mechanical hysteresis behaviour in the automatic transmission of a commercial vehicle. In contrast with the traditional physics-based modelling we have devised a data driven methodology to predict the accurate behaviour of a dynamic system. Instead of doing a blind implementation of the artificial intelligence approach, we have built on extensive experience in the development of hydro-mechanical systems to get robust and accurate results, which we refer to as ‘physics driven artificial intelligence’.

This is a new horizon where we can make virtual validation processes more robust, and test across a wider range of test cases without the use of physical hardware and time on a test rig. To achieve this, we have categorised the behaviour of the system using a classification technique and then learned a regression model for each behaviour independently. We have also compared the performance of different kernel functions, and then selected the kernel function with the best performance to make the predictions. We have investigated the non-linear behaviour of the hydro-mechanical hysteresis and latch point in the valve body to devise inputs that take these into account in the learning and prediction model.

This work presents an effective and elegant way of tackling the common issues with a non-linear model, that could be implemented on other systems with inherent non-linearities.


Drive System Design Controls Engineer Yash Bagla is due to present this research at CTI USA, rescheduled to 12-14th October 2020.

You can receive regular updates from us (such as our progress on this research) by signing up to our ‘Research Reviews’ here.