The Leading eBooks Store Online
for Kindle Fire, Apple, Android, Nook, Kobo, PC, Mac, BlackBerry ...
Sensor Fusion And New Materials
An easy example is the ‘‘electronic nose’’. These typically comprise an array of 20 or more chemical sensors, each with its own sensitivity to a particular gas or combination of gases. Each sensor is pretty crude and lacks selectivity, however it only needs to have a repeatable set of characteristics for it to be a useful addition to the nose.
It is the combination of various characteristics that enables the nose to discriminate between aromas. And as I cannot believe that evolution provided us with a specific sensor for Chateau Lafitte 1992, I can only presume that our own noses operate in much the same way.
Another even simpler example is the determination of acceleration from a distance sensor and an accurate time signal. This is very easy and straightforward because all the data are instantaneously available and it all relates to the same instant in time. Difficulties arise if one sensor has a delayed response compared with those of the rest of the system; and these difficulties are compounded by additional non-linearities. For example, if you wish to determine the acceleration of a model train, but your only inputs are time and a ‘‘distance’’ encoder which is attached to a carriage which is connected to the train in front by a rubber band; then life suddenly becomes a lot more complicated.
Sensor fusion has a great deal of potential, but getting a valid result can be far from straightforward. Anyone who thinks that they can simply connect a large number of sensors to a neural network and press the ‘‘learn’’ button, is likely to be disappointed by the result.
‘‘Garbage in garbage out’’ is especially valid when multiple sensors are involved. Each sensor input must first be cleaned up, delayed by some amount which may also be a variable, and then weighted by the network in accordance with its significance to the end result. Finally you need to test the whole system with a large variety of real data to make sure that it actually works, and then cross your fingers.
Fundamental to all the above is that you must understand how your system works. If you do not have this knowledge and instead abdicate responsibility to an artificial neural network that has no idea whether it is monitoring the acceleration of a train or the temperature in a furnace, then you will deserve the end result. This is a shame because sensor fusion does have a lot to offer, but it needs to be used with great care.
Previously published in: Sensor Review, Volume 22, Number 4, 2002
99 pages; ISBN 9781845447229
, or download in