Use of convolutional neural network for sensor modeling and on-board diagnostic 2021-36-0102

The use of physical based mathematical models has been the standard when it comes to model development in the powertrain domain. More specifically, substitute models and diagnostic models (also known as virtual sensors) used to identify sensor fault are applied widely in the engine control in order to comply with on-board diagnostic (OBD) requirements. In the early years of calibration, the amount of data was limited: nowadays with extensive and detailed validations demanded by the sector, the vehicles are submitted to a bigger span of tests before reaching the final costumer. It is in this context that a neural network (NN) thermodynamic model of the engine coolant temperature (ECT) of an internal combustion engine (ICE) was created: the main drawback of any machine learning method is the amount of data necessary, but with data not being a problem, the NN can be quickly implemented and generalized to other projects, reducing time of development and costs for them. Since a hybrid-convolutional architecture was chosen in this work, the size of the final hybrid convolutional neural network (HCNN) is not a problem and it can be implemented in the electronic control units (ECUs) available in the market today without increasing substantially its costs (due to the higher computation and memory resource demanded). Once deployed, the model can then be used as protagonist in the plausibility diagnostic of the coolant temperature sensor. The following paper will explain the development of such HCNN, present its results, as well as the potential re-use as a second proof of concept.