EEVblog Electronics Community Forum
Electronics => Projects, Designs, and Technical Stuff => Topic started by: cdev on October 06, 2018, 04:11:31 pm
-
Hey people, have you ever considered using the RPI and external reference data
to implement an improvement to the usual GPSDO control loop which integrates input from other variables that influence quartz crystals that areant usually considered? -
It seems that the more information thats brought to the table to learn how to discipline the GPSDO most effectively- and also learn the quirks of the case, its environs, including latitude, longitude, seasons, etc, plus whatever sensors are on there - using the whole bundle of data to discipline its oscillator in the most appropriate way- the more the better.
This task might be perfect for a neural network.
This seems like it would be a fun project for learning a bit more on NNs.
A neural network might be able to improve the model semi-automatically over time.
You would need a good reference GPSDO, to provide the point of reference for frequency - (and the quality of the reference GPSDO would be important while training) and good sensors to provide training data and you just keep refining the model to make it more and more accurate.
I'm sure this is what newer commercial GPSDOs must do.
Any of a number of inexpensive SOCs (in addition to the RPI) could be used but the RPI and a few others may stand out in the pack if their GPUs could be utilized because they support openCL.
One likely would want to make a multi-partitioned case with a thermal barrier to separate the XCXO and GPS area from the CPU/GPU (and the power supply).
The neural network might actually end up simplifying the design.
-
This might be an interesting alternative the traditional Kalman Filter where a specific model for changes is gradually improved in the light of measurements but this is more used during hold-over periods. With a good antenna there is very little need for hold-over especially for hobbyists - it is more important if you're running a telecoms system.
The disciplined side of GPSDO would only benefit if the system could be used to allow longer time constants - essentially extending how long the OCXO (or other oscillator) maintained a better frequency accuracy than that of the input from the 1pps from the GPS. But it would be difficult to do this as the learning examples would only be as accurate as the disciplined oscillator - it would be difficult to improve on it. What I mean is that the goal might be to predict what the next Vc setting should be in the light of temperature, and other inputs but you only know what Vc value was set by the GPSDO which may not be the optimal value - you're only teaching the system to emulate itself.
So for hold-over this is very feasible (you have the GPSDO in non-holdover mode to provide a good source of training data) but for general operation I don't think applying machine learning will actually achieve much unless you have a better oscillator (hydrogen maser?) to compare it with.
-
This kind of approach (which is basically a way of temperature compensating an oscillator a bit better) might be useful in helping make a slightly more accurate, non GPSDO local time server. Basically just training the system to "be more like the GPSDO" given the inputs available?
But as you say, you're unlikely to be able to make it any more accurate than your reference standard using this method. Unless there is something I'm not understanding about it. The lack of transparency into what the NN is actually calculating also has an element of disturbing to it.