Neural Networks in the automotive industry: Huber Group is working on substituting a NOx sensor with a neural network/ Learning network to replace physical model

The Huber Group, TIER-1 supplier and development provider, is currently researching the possible applications of neural networks in the field of selective catalytic reduction (SCR) and is already noting the first signs of success. The aim is to replace a NOx sensor in SCR systems by a neural network or moreover to integrate a trained neural network into an electronic control unit.

Current SCR systems use two NOx sensors. The task of the NOx sensor, which is mounted in front of the SCR catalytic converter, is to detect the nitrous oxides emitted on combustion. This measurement is used by the electronic control unit as a calculation basis to determine the quantity of AdBlue® injected for the SCR reaction. Considering the cost optimisation aspect of the SCR system technology, one possible option could be to dispense with the PreCat NOx sensor. When substituting such components, an attempt is usually made to extract real physics by constructing a corresponding simulation model. It is however extremely difficult to determine the formation of nitrous oxides on the basis of such a physical model, as the procedures are complex and also specific to each vehicle. However a neural network can learn the behaviour of the engine concerning NOx formation, without the added timeconsuming undertaking of physical modelling.

A neural network describes the interconnection of artificial neurons, whereby each neuron represents a simple CPU. The neural network forms a connectionist system that is a system based on the interaction of many interconnected single units. The neurons fit together in a specific network structure via weighted connections. A neural network is able to learn by example, whereby learning is to be understood as optimising the weighted connections between the single neurons. Considered more simply, the „knowledge“ of a neural network is stored in its weighted connections. The aim is to use the association ability of the neural networks to be able to determine a target variable, using the „knowledge“ from real and not trained input variables.

Current tests have shown very promising results. There is a high correlation between the comparison of existing measurements of the NOx mass flow of a vehicle and the calculated measurements of a trained neural network. The greater the significance of the input variables made available to the neural networks, the higher the precision with which the emissions performance can be learned with respect to NOx.

„We have been working on the application of neural networks within an SCR system for some time now. Generally speaking I think there is enormous potential regarding the employment of neural networks within automotive applications, especially where it is particularly complex or uneconomic to set up a model due to scarcely extractable connections. I assume that in the future we will be faced with an increasing number of such cases, with ever increasing complexity and  at the same time there will be a rise in the claim on the controllability of the systems“ says Daniel Geiger, Function Development Engineer at Huber Group. „There are naturally still numerous questions which have to be answered concerning the current application. Overall the neural network does not only have to withstand a comparison with real measurement results. Topics such as fault tolerance or the degree of association ability also have to be examined. The validation process is thus extensive. The actual extent of adapting the procedure to various vehicles will have to be examined in more detail. For this reason we are currently also working on creating a profile to generate specific training data which will prospectively be able to define a driving cycle to optimise data collection.“

This function of a „virtual NOx sensor“ and other sensors is planned as a software function in the very near future and can be implemented as „object code“ in existing software architecture.

This post is also available in: Englisch