Recent advances in ANN (Artificial Neural Network) technology enable new methods to be developed in sensor technology. There are a large number of cases where there exists a causal relationship between one or more inputs and a physical quantity, but where an easily implemented analytical relationship between the inputs and the output can not easily be found. In such cases, machine learning techniques, such as artificial neural networks, are able to model that functional relationship. However, using conventional computing hardware, these methods, while theoretically attractive, are too computationally intensive for field deployment in real-time systems.
Using a hardware implementation of an artificial neural network architecture, these computational restrictions can be eliminated. The authors have been working on the implementation of neural network based virtual sensors, specifically for use within automotive powertrain applications, using multiple inputs to estimate multi-dimensional non-linear relationships between existing sensor data and desired engine control parameters.
This paper will discuss the process of developing a hardware ANN based function estimator, with special focus on virtual sensing applications. The example used in this paper is a virtual mass airflow sensor (MAF), where the MAF estimate has been derived from other sensor inputs which are readily available to the engine controller.