I am wondering if over time MEMS accelerometer produces any worse output than when it is new? I found this research paper which acknowledges that temperature affect sensor output but there are no conclusions about reading stability over time.
The short answer is yes; they do change characteristics with time - and I have measured them to do so. On advice from the manufacturer of the device I work with (in this case, VTI, now a division of Murata), most of the ageing occurs in the first year.
As far as what changes, the most obvious is bias. I have seen drift from 3mg to 10mg on the SCA3100, though the maximum specification is 30mg over its lifetime (since the max bias is specified to be 70mg, and bias over temperature is specified to be 40mg). These are 6-sigma specifications so few will ever reach that magnitude.
There is no obvious change to the amount of noise. Take this claim with a grain of salt - this statement is "by eye" and not properly based on data or statistical testing.
As far as other specifications is concerned (e.g. scale factor, cross-axis sensitivity), I gave no data, measurement or intuitive feel as to the magnitude of drift, if indeed there is any appreciable drift.
Other devices, of course, will have different characteristics, but I have no data to say what they may be. I suspect the cheaper the accelerometer, the more it will age with time, but I have no evidence of that.
There are also short-term ("in-run") changes to the bias, even if the temperature is perfectly constant. If this is of interest to you, look for allan variance accelerometers on Google for a start on what drift characteristics to look for.
Stability wise, there are very few MEMS accelerometer designs that would have their mechanical stability affected, until of course the cantilever beams break. But that is pretty noticeable.
Parametrically you will expect some shifts over time. You can see that as the beam flexes that dislocations will tend to migrate to higher stress areas, so that will affect conduction in those areas. While they are carefully designed and tested and aged to minimize failure accumulation, nothing is ideal so eventually you should expect to see devices start to have shifts in spring constants on the beams and perhaps even some nano-fracturing. They are designed to last millions or billion of cycles but they will fail eventually and the lead up to failure certainly will have parametric shifts is perhaps a better way of putting it.
A lot depends on the design of the MEMS part and of the electrical part. And while one can broadly say that there are certain failure mechanisms, unless you have looked at a specific design in detail you can't be sure. But broadly speaking yes, expect shifts.
Things that I would broadly expect is shifts in:
- output to g level coefficient change
- linearity at maximum loading
- increased noise floor
- shift in resonant frequency (if it ever had one)