In spectrum sensing, threshold is the most important term for estimating performance measure. In order to carry out particular type of CFAR detection, threshold is determined by reverse calculation from false alarm probability. But most of the time it doesn't match with predefined detection. So I want to know is their any other method by which we can find out threshold so that in determining performance measure it will perfectly match.
Dear sir
You can determine threshold by monte-carlo approach, you consider the H0 hypothesis and try to meet the required false alarm probability.
BR
The Neyman Pearson criteria gives a good approach. Is it possible the reason it does not match your computed results is that the assumed null distribution is not representative of the true/empirical null? I've attached a figure that shows how a "naive" null distribution can differ substantially from an observed null distribution; in which case, the threshold for event declaration (a detection) will be much to low, and you'll receive many false alarms.
CFAR is looks straightforward in theory but is problematic in practice.
The false-alarm (FA) rate is the prior probability p (relative frequency) of facing FAs times the conditional probability q of registering an FA given that an occasion for an FA arises; FAR = p q. Theory assumes that you know p and q.
In practice, however, you first need to define all of the classes of events that constitute FAs. These presumably include classes of signal noise, classes of target-like objects or events, and so forth. Then you need to predict both p and q for each class of FA under the particular conditions of operation. Little prediction of FAR is possible if you are not first very clear about the classes of non-target “things” that cause false alarms for your detector, and also know the p and q for each class.
You may furthermore imagine that the classes of FA are independent, in which case the overall FAR is the sum of pXq for all classes. But they may not be independent because FA classes are not defined by the way you see the world, but by the way your detector sees the world -- through the feature set (target characteristics) that your detector is using to identify genuine targets. Hence different detectors generally see different classes of FA, and people generally see different classes of FA than detectors do.
What you may find, then, is that an in-depth applied analysis of FA classes is fraught with ambiguity. If so, performance prediction for your detector must simply assume (1) that the detector sensitivity S has been adjusted to whatever value happens to deliver the operationally prescribed CFAR; and (2) that the probability of detection D therefore an uncertainty +/-dD owing the fact that you could not predict what that sensitivity setting S may be. The size of the uncertainty dD depends on the size of the range of sensitivity adjustment S1 to S2 that can be expected in practice. Unfortunately this uncertainty +/-dD can be rather large for many detectors in realistic applications, in which case performance prediction is very problematic indeed.
There are ways of rolling uncertainties +/-dD about probabilities D into a net probability D’(D,dD), in which case D’ would be your probability of detection given that you are uncertain about the detector sensitivity setting required for a given CFAR during operations, but care is required in its interpretation and use.
Monte Carlo simulations are a very good method for estimating the exact threshold value for guaranteeing a given false alarm probability. I have used this method for the Weibull distribution in the following paper.
Good Luck!
Article Estimation of the Relation between Weibull Distributed Sea C...
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing