Ok so there is no concern on the convergence of the resulting density function when I simulate previous deleted values with the estimated density to which I add the new batch of values in order to compute a new estimate of the density. I should say that I'm a little concerned by this step. Speed is not my concern since the sensor issues values quite slowly.

Thanks for your detailed answer. But you haven't given your opinion of the method I originally planned to use (that is estimating the density function)

The problem lies in the fact that I need to process the output of ten thousands of sensors. Thanks for the threads you hinted at, I didn't know what was the English term for this specific problem