Maverick's World

Inertial Measurement Units (IMUs)- Some key issues - IV

The future is hard to predict and the present can only be described partially because we are never fully informed about anything. With that caveat I can make the following guesses about where this is likely to proceed.

As seen in earlier posts, barring major advances in understanding of gravity, defining the accuracy of an IMU will remain a very challenging affair. The precision of an IMU is a relatively simpler affair and we are likely to see Atom optics based gyros used to "clock" the performance of other systems. This kind of bootstrapping will create a deeper understanding of the nature of the error in other systems.

As Atom Optics related technologies become better engineered, we will see a gradual shift in the mission critical side of IMU applications. On the commercial side we will likely see a growth in MEMS based applications. It is not entirely unlikely that these two branches may come to leverage off each other.

I feel we are likely to see the following happen too

1) Role of Sensor Fusion: Using sensors of different types to check on each other offers interesting avenues for reducing noise in measurements of gravity. Schemes involving magnetometers have already been demonstrated, but a number of other schemes may also be possible. Such schemes will improve the precision of any number of existing devices.

2) Clouds help reduce noise: In theory one could have the IMU transmit a signal to a cloud, process it on the cloud and then resend the filtered signal back to the guidance system. This would be too unwieldy to carry out in a military or strategic application, but it may be possible to use this approach with commercial devices.

3) Deep learning will help fish weak signals from noise: It may become possible to implement a deep learning to extract small signals out of the noisy data from a cheap IMU. In the event that a deep learning network is so trained, a version of this network could be deployed on an embedded system attached to the IMU. It is difficult to ascertain how "good" this could be in actual deployment but the idea is plausible.

When speaking about these issues in the context of an actual deployment (as opposed to a "hey check out my Github for my latest Python code" context) - we are looking at a lot of money and hours spent on developing high reliability code and hardware. Those problems easily add decades on to the simplest thing.

N.B.  In order to keep this simple I have left out two other sources of trouble in an IMU - offset and latency. The discussion of these topics is complicated for non-specialists and getting into that will not add anything to what I am attempting to do here.

posted by maverick @ 10:43 AM