Angelos
Key Veteran
Location: nr Oxford, OX11, UK
| I don’t know much about neural networks but I do know their operating principle. For some applications they are the ideal solution like predicting which way the stock market is going etc. However I see a problem with Autonomous Flight applications.
Neural Networks are trained not programmed. Neurons may be simulated by software but their overall response is kind of unpredictable. No matter how many hundred of hours you test the neural network it will not guarantee that the response will always be what you expect.
The issue I see here is that you can’t debug it and you can’t really calculate a failure probability matrix. If you can’t at least do that, how do you go about proving the system is safe and justify your case to the insurance companies.
At work for our personnel safety system we use relays to build a single point of failure redundant system with fault reporting. Yes I know it relays sounds outdated but is it simple to calculate the failure rates of the system and get permission to operate the machine (an electron accelerator that produces strong x-rays). If we had be using PLCs then we would have to prove that the PLC code is perfect and that the PLC interpreter code is perfect and that the microprocessor inside the PLC does not have any hardware bugs and take into account failure rates for several types of electronic components. At the very minimum we would have to build the system twice based in different makes of PLCs based on different processors and with the code written by two different people who shouldn’t really talk to each other about how they are doing it.
Anyway, I think I have pushed it too far with the above example, but before I am convinced that an autonomous UAV is safe for commercial work I would like to see some properly done calculations regarding failure rates. I bet no UAV manufacturer have thought about this before! |
