Implementation of a Virtual Sensor on an ECU Using Recurrent Neural Networks

Advanced technologies such as artificial intelligence offer new opportunities to improve existing software development processes in a modern vehicle. Oftentimes such improvements can be accomplished through exact knowledge of critical vehicle states and inputs. Using physical sensors for such tasks can be expensive or even impossible, and the implementation of an alternative virtual sensor using artificial intelligence offers significant advantages. However, many times the deployment to embedded hardware can be challenging. Typically, memory footprint is crucial on embedded hardware and production code needs to be optimized. In this session, see how a virtual pressure sensor, developed with a recurrent neural network, can be implemented in a production code generation tool chain using only fixed-point datatype. The newly implemented workflows are fully automated and were developed in a joint project between MathWorks and Mercedes-Benz.