Equation-based domain knowledge utilization into neural network structure and learning

Abstract

The use of neural networks as black boxes, though useful for modeling complicated industrial systems, has some limitations. No physical interpretation can be given to sections of the trained network. Incorporation of domain knowledge into neural network attempts to address this lacuna. Most of the attempts in this direction have been in the area of data classification in which sub-classes created with the help of domain experts have led to better neural networks. This work attempts to incorporate domain knowledge into the structure of a neural network for solving a regression problem—that of a piston pump leakage prediction. It shows a way in which prior knowledge about subsystems, in the form of equations, can be used to create a neural network for modeling the entire system. This approach significantly outperforms a traditional feed forward neural network. As a key contribution, this approach allows physical interpretation of the neurons which can aid in troubleshooting and anomaly detection.