V&V of Fixed-Structure Neural Networks for Flight Critical Systems

Onboard nonlinear models are a key enabling technology for virtual sensors, model-based control, reconfigurable control and model-based diagnostic algorithms. Before such models can be used in safety-critical applications, such as civilian aircraft, they must undergo extensive testing to verify that there is no combination of inputs that will generate an undesirable output. While this can be done relatively easily for lookup tables and some nonlinear decision logic, it is more difficult for more powerful models such as multi-layer perceptrons (MLPs) or polynomial neural netowrks (PNNs).

BAI is working with Goodrich Aerospace to develop analysis techniques and a software tool to verify and validate fixed structure neural networks that are replacing lookup tables for near term upgrades of current generation aviation systems. Specific model types include single- and multiple-layer perceptrons with Sigmoidal basis functions, polynomial neural networks, Pi Sigma networks, and polynomial models resulting from orthogonal basis function modeling. The verification technique combines exhaustive testing over a uniform grid of test points with analysis techniques that bound the output and error of the network between test points. The software tool, which will automate this process, will ultimately be qualified under RTCA DO-178B and any other applicable FAA directives for automated V&V tools.