TL;DR I think that it would be almost impossible to acquire the data you are interested in, because it will be very difficult to collect and there are a lot of inherent biases that you cannot get rid off.
First of all, you have to know that there are four general ways to obtain a paper:
- Via a subscription to a journal.
- Via individual purchase from the journal.
- Via freely available preprints.
- Via the author.
You only have a hope to collect any useful amount of data from the first two sources and preprint servers, such as the Arχiv.
The major paper consumer by far is academia itself. As most academians get access to papers via university-wide subscriptions via the universities’ internet connections, there is no easy way for publishers to say who actually downloads a paper, because the journal will only know, which university the downloader belonged to. Something similar holds for big research societies such as the Max Planck Society.
Now, going by your question, you are not interested in this consumer, but this is important to consider as to what other paper consumers do, when they want to obtain a particular paper. As the vast majority of non-academic paper consumers were once in academia, they are likely to have some ties to their alma mater, by which they can easily (though perhaps legally questionably) get them access to a desired paper. Apart from reducing the number of downloads by other ways (which could generate the data you desire), this introduces a strong bias, as the prevalence of such downloads is likely to depend on the sector and field.
Moreover, as academia towers other sectors in paper consumption, it may be a considerable source of noise, even if you are not interested in it. Though academia¹ only downloads only a very small relative amount of its papers via non-university subscription or direct purchase, the total number of such downloads may be relatively high in comparison to what other sectors download. For example it could be that the student who does some surveillance job it an institution (which allows him to study while working) downloads more papers via that institution’s subscription for academic purposes than the institution downloads for its own purposes.
Another problem is the field dependence in the availability of online preprints (and similar): Making papers freely available is much more common in physics than medicine. This will give you an inherent bias on downloads from preprint servers and thus render this potential data source useless. Moreover, this will cause a strong bias in data collected by other means. If the papers of some disciplines are much more likely to be available freely, they are much less likely to be downloaded on ways that allow for collecting data on the paper consumers. Combining the two sources might reduce this bias, but won’t eliminate it, as a large number of freely available preprints and similar is available via private hosting.
Moreover, the data available on the total number of individually purchased articles is already rather scarce (see this question) and difficult to come by and it seems to be rather low. Breaking down this data as to the source of the purchases certainly won’t be easier. Also, collecting or using this data may violate privacy agreements.
Due to the immense price of such subscriptions, I assume that most non-academic subscribers are rather large organisations and companies, which in turn may cover many sectors and thus won’t generate useful data.
¹ by which I mean people doing research and similar to further their academic career