Bayesian Analysis of Time-intervals for Radiation Monitoring Peng Luo a T. A. DeVol a and J. L. Sharp b a.Environmental Engineering and Earth Sciences b.Applied Economics and Statistics
2011 SRC-HPS Technical Seminar Time-Interval ? Time t2t1t3t4t5 … Background
2011 SRC-HPS Technical Seminar Time-Interval ? Time … Presence of a source t2’t1’t3’t4’t5’t6’t7’t8’t9’
2011 SRC-HPS Technical Seminar Bayesian Statistics Bayes’ theorem x – observation, counts or time-interval; r – count rate P(r): gives the prior prob. of observing r before data is collected, P(x|r): gives the prob. to obtain the observation x given r P(r|x): summarizes the knowledge of r given the prior and data PriorLikelihood Posterior Bayesian statistics give direct probability statements about the parameter based on prior information and actual data. Subjective
2011 SRC-HPS Technical Seminar Conjugate Prior: Gamma distribution Poisson Distribution Time-Interval Distribution Gamma Distribution
2011 SRC-HPS Technical Seminar Example 1:Background Data Gamma Prior Posterior
2011 SRC-HPS Technical Seminar Example 2: A Source is Present
2011 SRC-HPS Technical Seminar Data Acquisition & Analysis Time-interval data acquisition: experiments (~10 5 pulses) and simulation (~10 6 pulses). Three methods are used to analyze the data: classical (1.65σ), Bayesian with counts (cnt), and Bayesian with time-intervals (ti). Fixed count time is 1 second. Compared the three methods in terms of average run length (time to detect the source) and detection probability (1- ) where is false negative (FN) rate.
2011 SRC-HPS Technical Seminar Bayesian Analysis Methodology If i=1 Posterior (i) Likelihood Prior Prior(1) If r > r 0 Prior(i+1)= Posterior(i) i=i+1 Yes No Obs.(i) …… Observations
2011 SRC-HPS Technical Seminar Average Run Length Average Run Length: the average time needed to make a decision whether an alarm is issued. Time-interval can make a quick decision. (s) a b Experimental result Mean count rate (cps) Average run length (s) a b Simulated result Mean count rate (cps) Average run length (s)
2011 SRC-HPS Technical Seminar Detection Probability Detection probability is defined as 1- 5s bkg.+20s source +5s bkg. 10 4 trials 95% detection limit (DL) Bayesian methods give lower false positive (FP) rates Also lower detection probabilities at low levels 5 s 20 s 5 s Detection probability Mean count rate (cps) FP ( ) FN ( )
2011 SRC-HPS Technical Seminar Detection Probability 5 s 20 s 5 s Detection probability Mean count rate (cps) 95 % 60% DL ARL – 60% DL Average run length Mean count rate (cps)
2011 SRC-HPS Technical Seminar Source Time Effect 2s, 5s, 20s and 50s source time Bayesian analysis has the ability to reduce both false positive and false negative rates. ab cd 2s 5s 20s 50s
2011 SRC-HPS Technical Seminar If source time < count time 0.5 s source 1 s count time Time-interval may result in a higher detection probability when the source time is shorter than the fixed count time. Detection probability Mean count rate (cps)
2011 SRC-HPS Technical Seminar Effect of Change Point s 5 s Change point --- a point at which the radiation level changes to an elevated level. Change point determines the amount of background data that are included in Bayesian inferences. The detection of radioactive sources may be delayed.
2011 SRC-HPS Technical Seminar Enhanced Reset
2011 SRC-HPS Technical Seminar Moving Prior t Data Prior The moving prior relies on the latest information to calculate the posterior probability by updating the prior probability with each new data point.
2011 SRC-HPS Technical Seminar Result of Modified Methods 10 (pulses)/20 (pulses) for ‘Enhanced Reset’. 10 (pulses) is set for ‘moving prior’. Both modified Bayesian methods result in a higher detection probability than the typical Bayesian analyses. The performances of two modified methods are independent of the change point. Time-Interval Data
2011 SRC-HPS Technical Seminar Summary ARL indicates that Bayesian(ti) could more rapidly detect the source than other two methods. When source time < count time, Bayesian(ti) results in a higher detection probability than Bayesian(cnt), which are both less than the frequentist method. For no source present, both Bayesian(ti) and Bayesian(cnt) result in lower false positive rates than the frequentist method. When source time > count time, Bayesian analysis results in a lower detection probability relative to the frequentist method, which increases with source time. Modified methods can improve the performance of Bayesian analysis by reducing the effect of the background.
2011 SRC-HPS Technical Seminar Acknowledgement Funded by DOE Environmental Management Science Program. (Contract number: DE-FG02-07ER64411) Dr. Fjeld and Dr. Powell (EE&ES)