Certifying Adaptive Flight Control Software

As aircraft designs become more complex, automationhas become an important factor in improving safety and reliability. Automated flight control systems can respond intelligently to faults when it is impractical for a human to take control quickly. In recent years neural networks have been proposed for fault identification and accommodation purposes within flight control schemes because they are well suited to non-linear, multi-variable systems. Because neural networks learn to associate various control actions with particular input data patterns, they avoid the need to explicitly program all the relevant fault situations. A major issue in the use of adaptive fault-tolerant flight control systems is certification. Current practice relies heavily on testing as a means of performing certification. This has two shortcomings, the first being high cost. The second shortcoming of certification by testing is that it only provides a fairly limited guarantee. Certification based on testing exclusively is meaningless for new software technologies, such as adaptive controllers and neural networks which ”learn” after deployment since the system tested in the lab is not the system that is being run in the field. In this paper, we formulate the research problems underlying the certification of adaptive systems. We also describe preliminary results of a certification case study, based on the methodology which combines formal methods and automated testing, and identify the areas of basic research needed to overcome the limitations observed in the proposed framework.