Because INS/IRS/FMS is still a fairly recent discipline, there is no totally consistent answer that holds good for everyone. Each company develops it's own solution, and as with VCRs (remember the competing claims of Betamax, Phillips, and VHS?) eventually one solution will dominate the market (like Microsoft, for instance), and everyone will do it that way.
At present, there are only 2 serious contenders - Honeywell, who build the systems for the Boeing, and the Airbus consortium. Whilst I think that Airbus produce excellent airframes and engines, I think that the Honeywell approach will eventually become the industry norm (at least, for a while - until it all changes).
With the Boeing 737-400 system (which is the one that the JAA use for examination), you have a left IRS and a right IRS, both independently fed into a Kalman filter. The IRS can only have a present position fed into it during initialisation, ie, on start-up. From that moment on, all of the IRS input is treated as sacrosanct within the IRS itself. If it's going to get mangled, that happens downstream, within the FMS's Kalman filter.
As you advance the TO/GA levers on the take-off roll, the FMS is automatically updated with the data-base present position of the piano keys of the take-off runway. The raw IRS positions are unchanged.
Once airborne, DME/DME positions are computed. The FMS does NOT automatically move the IRS positions to the DME/DME. What is does is, over a period of time take out a small amount of the apparent fix error on each fix. In broad terms, it finds out the history of the actual DME fixes compared with each separate IRS. The computer knows that each IRS will generate a combination of bounded and unbounded errors. The bounded errors will be a sinusoidal Schuler with a period of 84.4 minutes and the unbounded errors will be a combination of a ramp and a parabola.
Knowing all this, if the Kalman filter gets enough fixes (which it will over a reasonable period of time) it can then do a curve-fitting exercise during which it can build up the history of the fixes compared with the raw IRS positions. It knows the format of the differential equations linking the differences - one will be sinusoidal and the other will be a ramp - all it has to do is solve the coefficients in the formulae - it's a form of regression analysis.
Using this, it can build a mathematical model of the error growth rate in both the left and the right IRS, to predict where the next DME/DME fix will come. The more good DME/DME fixes it gets over time, the more refined this model becomes.
If, at any stage, it loses DME/DME, it uses the last model it had of the error growth rate to predict the probable error of the raw left and right IRS positions. If there is no other information to suggest which is the more reliable one (which there usually isn't, once DME has been lost), then it takes the left one.
However, at the end of the day, it is only the FMS position which gets changed. The raw left and the raw right IRS positions continue to be computed until you switch the thing off.
Simple, really!!!
Last edited by oxford blue; 21st May 2002 at 21:50.