Using Fuzzy Self-Organising Maps for Safety Critical Systems

This paper defines a type of constrained Artificial Neural Network (ANN) that enables analytical certification arguments whilst retaining valuable performance characteristics. Previous work has defined a safety lifecycle for ANNs without detailing a specific neural model. Building on this previous work, the underpinning of the devised model is based upon an existing neurofuzzy system called the Fuzzy Self-Organising Map (FSOM). The FSOM is type of ‘hybrid’ ANN which allows behaviour to be described qualitatively and quantitatively using meaningful expressions. Safety of the FSOM is argued through adherence to safety requirements – derived from hazard analysis and expressed using safety constraints. The approach enables the construction of compelling (product-based) arguments for mitigation of potential failure modes associated with the FSOM. The constrained FSOM has been termed a ‘Safety Critical Artificial Neural Network’ (SCANN). The SCANN can be used for nonlinear function approximation and allows certified learning and generalisation. A discussion of benefits for real-world applications is also presented within the paper.