A Practical Reliability Evaluation of Embedded Avionic Software

The implementation of multiple microprocessor based flight control systems has provided new highs of hardware reliability. Current design goals for system reliability are in the order of 10 to 50 thousand hours. The need to consider failure of software in the system MTBF calculations has arisen, even though software doesn't "fail" in the traditional hardware sense. The current state of the art in software reliability prediction badly lags that of the hardware. One current approach is to use multiple independently developed, software programs to eliminate the generic software fault in multi-processor systems. A side effect is to increase the cost of the software portion of a system. The General Electric Aerospace Control Systems Department has developed several flight control programs that have been in service for years, and some system failure data is examined in detail with respect to software faults. The results show that reliable software systems can be met with well tested and maintained single software designs. After one million flight hours in the field, one system exhibits continued reliability growth, with ten thousand hours MTBF for the software. The failures were all non-critical and include incorrect requirements as well as design and implementation errors. Over sixty percent of the failures were traced to incorrect requirements. The implementation of a reliable, well tested single software design system is a cost effective alternative to costly multiple "independent" program developments. Recent developments show unexpected correlation of errors in independently developed software designs, reducing expected improvements in reliability. This, coupled with the significant portion of errors imbedded in requirements, points toward well tested single software designs as a method of meeting software reliability requirements. Current FAA certification philosophy is open-minded on the single software design approach.