Improving Safety of AI Systems

This article explores the techniques to enhance the safety of AI systems, focusing on the predictive maintenance of aircraft engines. It addresses challenges such as lack of formal specification, implementation transparency, high dimensionality effects, differing probability distributions, and quality data. The proposed solutions include safe design, safe margin, and safe fail techniques. A case study on a prognostics and health management (PHM) system for aircraft engines demonstrates the application of these techniques.