An investigation of detection limits for an early warning system for measurement of radioactivity in the atmosphere

An investigation of detection limits for an early

warning system for measurement of radioactivity

in the atmosphere

Ebba Ahlgren Cederl¨

of

Contents

3.2 Count rate . . . 10

3.3 Filter change . . . 12

3.4 Energy stabilisation . . . 13

3.5 Efficiency calibration . . . 14

4 Early warning system 16 4.1 Residuals . . . 17

4.1.1 Method . . . 17

4.1.2 Results . . . 18

4.2 Principal component analysis . . . 24

4.2.1 Logistic regression . . . 25 4.2.2 Method . . . 26 4.2.3 Results . . . 26 5 Discussion 30 6 Conclusions 32 7 Acknowledgements 33 A Code for programming the detectors 35 A.1 Run Filter measurements . . . 35

Abstract

This thesis investigates the detection limits for an early warning system detecting anthropogenic radioactivity in the atmosphere through air fil-tering. Two algorithms are discussed as potential warning systems, one based on residuals between spectra and one based on principal component analysis. Both methods are evaluated for two types of detectors, a NaI(Tl) detector and a HPGe detector, and with two sources, a137_{Cs source and}

a133Ba source. For both detectors, the highest sensitivity achieved is in the order of 1 kBq of accumulated activity on the filter. With an airflow of 1600 m3/h, this corresponds to an air concentration of 0.6 Bq/m3being detected after 1 hour.

1

Introduction

In the case of a nuclear accident, an early indication of abnormal levels of ra-dioactivity is important to be able to take the appropriate actions in time. Sweden, like most countries, has an active network for monitoring radioactiv-ity in the atmosphere. A high sensitivradioactiv-ity is achieved through high-resolution gamma-ray spectroscopy of particles in the atmosphere collected at six air filter stations located in Kiruna, Ume˚a, G¨avle, Kista, Visby and Ljungbyhed. The stations are managed by the Swedish Defence Research Agency (FOI) on behalf of the Swedish Radiation Safety Authority (SSM). Historically, studies of fall-out from nuclear tests have been carried fall-out with this network and it is part of Sweden’s monitoring of the Comprehensive Nuclear-Test-Ban Treaty (CTBT)1_.

At the air filter stations, particles are continuously collected by filtering the air. The filters are collected once a week and the radioactivity is measured with a shielded High Purity Germanium (HPGe) detector in the lab in Kista. The detected nuclides and their air concentrations are reported to SSM and a yearly summary is published by FOI (see Ref. [1] for the latest report). This method is very sensitive, but a result is obtained first roughly a week after the collec-tion has stopped. The implementacollec-tion of an early warning system would offer a faster indication of abnormal radiation levels in the atmosphere.

A type of early warning system with NaI(Tl) detectors has previously been in operation at the national air filter stations between 2004 and 2011. In 2012, FOI [2] evaluated the system’s stability and detection limits with respect to three different alarm criteria. The report found, from measurements in the lab, an absolute sensitivity in the order of magnitude 1 kBq accumulated activity on the filter or an air concentration of 1 Bq/m3 _{after 1 hour of collection with an} airflow of 1000 m3/h. For measurements at an air filter station, background ra-diation dominated by nuclides from the decay of222Rn complicated the analysis and reduced the sensitivity of the system. At low radon levels, the detection limit was around 10 times higher than in the lab. The contribution from the decay of222_{Rn is not constant with time and may display short term temporal} variations due to effects such as ventilation [3] and long term variations due to season [4]. In this thesis we aim to improve upon the previous results by taking into account the radon background in the designing of new early warning criteria.

Multiple methods for estimating and reducing the radon background have been proposed. Recent attempts to improve the detection limits of atmospheric ra-diation by background reduction, using sophisticated software and algorithms, have been successful. H´yˇza and Rul´ık [5] used a method based on principal component regression to identify the spectral components of the radon decay background and subtract it from the measured spectrum. They achieved a minimum detectable activity of 8 mBq/m3and 7 mBq/m3for131I and137Cs

spectively after 24 h of sampling. Kov´aˇr and ˇSolc [6] demonstrated a technique for subtracting Monte Carlo simulated natural background spectra, in addition to an estimate for the cosmic rays background, from the measured spectra to to decrease the decision threshold for an early warning system.

2

Background

2.1

Gamma-ray spectroscopy

In this thesis we use gamma-ray spectroscopy to identify and quantify radioac-tivity. Gamma-ray spectroscopy is the study of energy spectra from gamma-ray sources, obtained using a gamma-ray spectrometer. The gamma-ray spectrome-ters are based on a photon from the source transferring all, or part of its energy to an electron in an active absorbing material of the detector. The electron will, in turn, lose its energy through ionisation or excitation of atoms in the detec-tor material and through bremsstrahlung. From the energy that the electron deposits in the detector, the energy of the gamma photon can be inferred. A histogram, or a spectrum, is outputted by the spectrometer, showing the number of registered events as a function of deposited energy in discrete steps. Since all radioactive nuclides release specific energies when they decay, we can identify the nuclides present by analysing the spectrum.

Apart from the nuclides present, the appearance of the spectrum also depends on the type of detector and its geometry. Two important features in a gamma-ray spectrum is the photo peak and the Compton continuum. The photo peak, or full energy peak, is the peak in the spectrum corresponding to the full en-ergy of the incoming photon. If the size of the active absorbing material of the detector is increased, the probability of the photon being fully absorbed is increased, and consequently so is also the area of the photo peak. In addi-tion to full absorpaddi-tion, it is also possible that a photon only deposits part of its energy in the detector. The most simple case is when a photon enters the detector and Compton scatters, giving part of its energy to an electron (that is later absorbed by the detector) and then escapes. Since the energy this photon transfers to the electron depends on the scattering angle, this will give rise to an energy continuum in the spectrum up to a certain energy which is the maxi-mum energy a photon can transfer to an electron through Compton scattering. This corresponds to a scattering angle of 180◦_{. The energy continuum} associ-ated with each photo peak is known as the Compton continuum. In practice, the process is more complex than described here, since a photon might perform multiple Compton scatterings before escaping, or scatter by some other process.

Another important effect that will affect the appearance of the spectrum is the energy dependency of the efficiency with which the detector detect photons. A photon with very low energy has a high probability of being stopped by the

material surrounding the detector and fewer of them will be detected. As the energy of the photons is increased, they will be more likely to penetrate the ma-terial and be absorbed in the detector resulting in more counts in the detector. At even higher energies however the photons are too energetic and some will pass through the detector without being absorbed, lowering the efficiency again.

In this report we have used two common types of gamma spectrometers, a HPGe detector and a NaI(Tl) detector. They each come with their own advantages and disadvantages.

2.2

HPGe detectors

A HPGe detector is a type of semiconductor detector. A gamma photon entering the active material of the detector ionises the atoms in it and creates electron-and-hole pairs along its path. An applied electric field causes the electron-hole pair to drift in opposite directions. The motion of the electron and the hole constitutes a current that remains until they are collected at the end of the boundaries of the active volume. This electrical signal is proportional to the de-posited energy and is translated by the electronics of the detector to a spectrum.

HPGe detectors have an excellent energy resolution when applied to gamma-ray spectroscopy, at the cost of a low efficiency in photon detection. Another drawback of the HPGe detector is that it must be cooled to below -150◦ C, often with liquid nitrogen, making it impractical for experiments outside the lab. However there exists electro-mechanically cooled HPGe detectors, like the one that we have used for the experiments in this thesis.

2.3

NaI detectors

A NaI detector is a type of inorganic scintillator. A scintillator converts the ki-netic energy of charged particles, produced by the ionising radiation of gamma photons, to measurable light. The intensity of the light produced is proportional to the amount of energy deposited, which makes it possible to deduce the energy of the photon.

A NaI(Tl) detector is a NaI crystal doped with Tl. Electrons in the crystal are either in the valence band or the conduction band. Between them is the forbid-den band, with energies that the electrons in the pure crystal cannot have. If an electron in the valence band absorbs enough energy, it can move freely in the crystal. The electron may de-excite back to the valence band through the emis-sion of a photon, but this process is ineffective and the typical band gap is of a size that would produce a photon outside the visible spectrum. To increase the probability of a photon in the visible spectrum, impurities can be introduced in the crystal. These are known as activators. At these locations in the crystal the band gap is modified and energy states in the forbidden band become available. The energy gap is decreased and a photon in the visible spectrum can now be

emitted. These locations are known as luminescence centres or recombination centres. The crystal is coupled to a photo multiplier (PM) tube, that converts the light pulse into electrons. This current is often too weak to measure so the electrons are multiplied by the PM tube into a measurable current. The relation between the number of electrons before and after amplification is known as the gain.

The high density of inorganic crystals and the high Z-value of iodine in NaI increases the probability of a gamma photon being absorbed by the detector, which results in both a high counting and photo peak efficiency, however the energy resolution is low compared to a HPGe detector. Another problem with the NaI detector is that the gain will shift with temperature, meaning the peaks will drift in the spectrum, further lowering the energy resolution and compli-cating the analysis of long term measurements. Nevertheless, the affordability of the NaI detector makes it favourable compared to the much more expensive HPGe detector.

2.4

Background radiation

During the period of measuring no abnormal incidents were reported, meaning we can treat the data collected as background measurements. The background radiation is dominated by gamma-energies originating from β− decay of radon decay products. In addition to this, there is also contribution from40_{K. In the} HPGe spectra a small contribution from cosmic rays can also be distinguished.

2.4.1 222Rn

Radon is a noble gas with three naturally occurring isotopes,219_Rn, 220_{Rn and} 222_Rn. 222_{Rn is the dominant isotope and is a part of the decay chain of}238_U. 238_{U is found in the ground and in rock. Since radon is a noble gas, it will} easily pass into the atmosphere once the mother nuclide has decayed. In the atmosphere the radioactive particles stick to dust particles and are carried by the winds and washed away by rain. The radon concentration in the air is thus sensitive to wind and weather.

The decay chain of222Rn is shown in Fig. 1. The chain ends with206Pb that is stable, however the progeny after210Pb are of less interest in our experiment, due to the long half-life of210Pb compared to the time-scale of our experiment, resulting in low detection rate of the progeny following.

2.4.2 40_K

40_{K is one of the primordial nuclides: with a half-life of 10}9 _{years it has existed} on Earth since its creation. 40_{K is present in all organic material, including the}

222_Rn 3.8235 d 218_Po 3.098 min 214_Pb 27.06 min 214_Bi 19.9 min 214_Po 163.6 μs 210_Pb 22.20 y 210_Bi 5.012 d 210_Po 138.376 d 206_Pb Stable 𝛼 𝛼 𝛽− 𝛽− 𝛼 𝛼 𝛽− 𝛽−

ϒ - Energy (keV) Intensity (%) 53.2284 [1.075] 214.9950 [7.251]

295.2228 [18.42] 351.9321 [35.60]

785.96 [1.06]

ϒ - Energy (keV) Intensity (%) ϒ - Energy (keV) Intensity (%)

609.320 [45.49] 1401.515 [1.330] 665.447 [1.531] 1407.988 [2.394] 768.36 [4.894] 1509.210 [2.130] 806.180 [1.264] 1661.274 [1.047] 934.056 [3.107] 1729.595 [2.878] 1120.294 [14.92] 1764.491 [15.30] 1155.210 [1.633] 1847.429 [2.025] 1238.122 [5.834] 2118.514 [1.160] 1280.976 [1.434] 2204.059 [4.924] 1377.669 [3.988] 2447.70 [1.548]

ϒ - Energy (keV) Intensity (%) 46.539 [4.25]

ϒ - Energy (keV) Intensity (%) 265.6 [51] 304.6 [28] 649.6 [3.4]

Figure 1: The decay chain of 222Rn, that is a part of the larger 238U decay chain. Gamma-energies and their probabilities (intensities) is given. Bold faced text indicates energies of interest in our gamma-spectra. Only energies with intensities ≥1% are given. Data from [7].

materials surrounding our detectors, and is characterised in the spectra by a photo peak at 1461 keV.

2.4.3 Cosmic rays

In a gamma-ray spectrum cosmic rays add an annihilation peak at 511 keV, resulting from pair production in the surrounding material, as well as a varying background to the continuum. There is also a small contribution from direct deposit in the detector. However, we expect this effect to be limited.

2.4.4 Background spectra

Fig. 2 (a) and (b) shows the typical background spectra from a 1 hour mea-surement of the filter using the NaI(Tl) and HPGe detector respectively. The measurements are taken simultaneously. The40_{K photo peak can be seen clearly} in both spectra, as well as the 295 keV and 351 keV photo peaks associated with the radon daughter214_{Pb and the 609 keV, 1120 keV and 1764 keV photo peaks} associated with the radon daughter214Bi. A small annihilation peak at 511 keV from cosmic rays can be seen in the HPGe spectrum.

The difference in energy resolution of the two detectors is clearly visible, illus-trating the advantage of using an HPGe detector for identification of nuclides. Looking at the intensity, we can also note the significant difference in counting efficiency between the two detectors.

214_𝑃𝑏 214_𝐵𝑖 214_𝐵𝑖 40_𝐾 214_𝐵𝑖 (a) 214_𝑃𝑏 214_𝐵𝑖 214 𝐵𝑖 40_𝐾 214_𝐵𝑖 (b)

Figure 2: Typical background spectrum from a 1 hour measurement of the filter by (a) the NaI(Tl) detector and (b) the HPGe detector. Some of the more prominent peaks are marked with the associated isotope. Their energies are listed as the bold-faced entries in Fig. 1.

To investigate the background in the room of the setup, a NaI(Tl) detector is placed roughly 5 metres from the filter. This is far enough to consider the surrounding materials as the dominant contributor to the spectrum, and not the filter. One of the ”No filter” spectra is plotted together with a ”Filter” spectrum, measured during the same time period, in Fig. 3. Most notably is the absence of the radon daughter peaks in the ”No filter” spectrum. However, the peak area of the 40_{K peak is of similar magnitude. This indicates firstly} that the radon daughter peaks are dominantly due to radioactive particles on

the filter and secondly, that the largest contribution to the 40K peak is the materials surrounding the detectors, not particles in the atmosphere.

0

200

400

600

800

1000

Channel

0

200

400

600

800

1000

1200

1400

1600

Intensity

Filter

No filter

Figure 3: NaI(Tl) spectrum measured with and without filter. Both spectra were recorded during the same time period, with a measurement time of 15 minutes.

2.5

Anthropogenic radioactivity

The aim of an early warning system is to give an early indication of anthro-pogenic levels of radioactivity in the atmosphere. There are different scenarios that could result in a large release of anthropogenic radioactivity in the atmo-sphere. One possible scenario is a nuclear power plant accident. A FOI report [8] published after the nuclear power plant accident in Fukushima Daiichi in 2011 reports that the radioactive nuclides that could be detected in Sweden after the accident were131I,129m,132Te and134,136,137Cs. Of these,131I was the most significant. It reached a maximum concentration in Sweden in the order of mBq/m3_{, while the}137_{Cs levels reached around a tenth of this. It is compared} to the nuclear power plant accident in Chernobyl 1986, when the concentrations of airborne131_{I in Sweden reached the order of 10 Bq/m}3_and137_{Cs reached the} order of 1 Bq/m3_. 137_{Cs has a half-life of around 30 years and is characterised} by a gamma emission at 661 keV. 131_{I has a half-life of about 8 days and is} characterised by a gamma emission at 364 keV.

3

Measurements

3.1

Air filter station and setup

The setup is located at the air filter station in Kista, Sweden. The filter station consists of a fan connected to a funnel with a filter mounted at the opening.

The fan pumps in air through the filter at a rate of roughly 1600 m3/h. The exposed area of the filter is 112 cm × 55 cm. The filters at Kista are automati-cally replaced every 28 hours. Each week the filters are collected and combined to one sample. 3-4 days after the final filter of the week has been collected the sample is measured by a shielded HPGe detector in the lab. This gives the radon daughters time to decay, which increases the sensitivity of detection of anthropogenic radiation.

For the early warning setup we introduce a NaI(Tl) detector 17 cm behind the filter and a HPGe detector 48 cm behind the filter, see Fig. 4.

Air-flow in Filter

NaI(TI) detector

HPGe detector

Inside

Outside

PC

Figure 4: A schematic picture of the early warning setup seen from the side. The NaI(Tl) detector is placed 17 cm behind the filter and the HPGe detector is placed 48 cm behind the filter. The detectors are both connected to a PC and operated via the software GammaVision R_.

We use an electro-mechanically cooled ORTEC Detective-100T 65 mm×50 mm HPGe detector and an ORTEC 905 series 2” × 2” NaI(Tl) scintillation radiation detector. Spectra collection is done with the software GammaVision R

(OR-TEC) and analyzed with self-written code.

3.2

Count rate

The count rate, measured in counts per second (cps), is an important quantity in this thesis. It is a measurement of how many photons are detected per

second by the detector in a chosen range. To calculate the total count rate in a spectrum we sum the counts in all channels and divide by the live time of the measurement. The live time is the effective time the detector spends able to record a new event. To calculate the count rate in a specific peak we first select a range in which the peak lies. We then perform a fit to the peak in the range with the following function:

f (x) = Ae−12(x−x0σ ) 2

+ Bx + C (1)

This is the sum of a Gaussian function, that accommodates for the peak, and a linear function, that accommodates for the background. The peak center, x0, and standard deviation, σ, are extracted from the fit and we define a new range x0− 3σ ≤ x ≤ x0+ 3σ where our peak lies. We calculate the mean count in the 4 channels directly to the left and the 4 channels directly to the right of the peak range. This is used as an estimate of the background and we subtract it from the counts in the channels in the peak range. We then sum all the counts in the background subtracted peak channels and divide by the live time of the measurement. In the case of individual uncertainties being much smaller than the trend in the data, error bars have been omitted in the plots.

The total count rate in the spectra varies over time, as demonstrated by Fig. 5. The difficulty thus lies in correctly identifying low-level abnormal radioactivity in a spectrum where the background radiation constantly varies.

15/02 2020 01/03 2020 15/03 2020 01/04 2020 15/04 2020 01/05 2020 80 100 120 140 160 Count rate (cps)

Figure 5: Count rate in the total spectrum over time. Each point correspond to a measurement with the NaI(Tl) detector for 1 hour.

Fig. 6 shows the variation in count rate over time for the total spectrum to-gether with the 609 keV, 1461 keV and 1764 keV peaks. The 609 keV and 1764 keV peaks originate from radon daughters and their variation seem to follow the variation in the total count rate. The 1461 keV, originating from40K however,

does not seem to have the same variation. This indicates that the variations in count rate in the total spectrum are mainly due to the variation of radon daughters on the filter. The small variations in the 1461 keV peak aligning with the variations in the radon daughter peaks is likely due to counts from the Compton continuum of the 1764 keV peak.

15/02 2020 01/03 2020 15/03 2020 01/04 2020 15/04 2020 01/05 2020 0 50 100 150 200 250 Count rate (cps) Total 609 keV peak (x10) 1461 keV peak (x10) 1764 keV peak (x10)

Figure 6: Count rate over time for the total spectrum (black), the 609 keV peak (green, ×10), the 1461 keV peak (red, ×10) and the 1764 keV peak (blue, ×10).

3.3

Filter change

In Fig. 7, the total count rate is plotted together with the time for filter change during a typical week. We see that the count rate drops immediately after a new filter is introduced, and builds up again as the radon daughters builds up on the filter. A typical NaI(Tl) spectrum during a day of relatively high levels of radon in the atmosphere is plotted right after a filter change, after 5 hours and after 10 hours in Fig. 8. The increase in the radon daughter peak areas over time at especially 295 keV, 351 keV and 609 keV can clearly be seen.

16/03 2020 17/03 2020 18/03 2020 19/03 2020 20/03 2020 21/03 2020 22/03 2020 80 100 120 140 160 180 Count rate (cps) 15 min measurement Start of new filter

Figure 7: Variation of the total count rate over time, each point represents a measurement of 15 minutes. Green vertical lines show the times a new filter is introduced. 0 500 1000 1500 2000 2500 Energy (keV) 0 200 400 600 800 1000 1200 Intensity Filter change + 5h + 10h

Figure 8: Typical NaI(Tl) spectrum right after a filter change (green) as well as 5 hours (blue) and 10 hours (red) after the filter change, during a day of relatively high levels of radon on the filter. The radon daughters peak areas grow with time as the radon builds up on the new filter again. The collection time for each spectra here is 15 minutes.

3.4

Energy stabilisation

The NaI(Tl) detector is sensitive to temperature. Variation in temperature will affect the gain of the detector, thus shifting the spectra. To counteract this we can use the naturally occurring40_{K peak for energy stabilisation of the}

mea-surements.

The variation of the 40K position over time is shown both before and after stabilisation in Fig. 9 together with the variation in temperature over the same time period. With no stabilisation large drops in temperature result in drops in the position. This is improved by the first stabilisation attempt. Tweaking the stabilisation parameters improves the stabilisation further and, although some oscillations still occur, there is no long term drift.

15/02 2020 01/03 2020 15/03 2020 01/04 2020 15/04 2020 01/05 2020 460 470 480 490 500 Position (channel)

No stab. Stab. Updated stab.

1461 keV peak 10 5 0 5 10 15 20 Te m pe ra tu re ( C) Air temperature

Figure 9: 40K peak position variation with temperature for the NaI(Tl) detector. Black points shows the channel of the40K peak centre over time. Red curve shows the temperature variation over the same time period.

3.5

Efficiency calibration

We perform a calibration of the efficiency in order to be able to convert the count rate in a peak in a spectrum to an accumulated activity on the filter, and by extension the concentration in the air. In addition, we can also use the spectra obtained from the efficiency calibration to simulate spectra containing anthropogenic radioactivity. We thus want to perform the efficiency calibration with both a 137_{Cs source and a} 131_{I source, two of the nuclides we are most} interested in detecting. Due to the short half life of131_{I we did not have access} to such a source, instead we used a133Ba source. 133Ba has a half-life of around 10 years and one of the gamma emissions of 133Ba is at 356 keV, close to the 364 keV peak of131I.

The filter is divided into a 8 × 4 grid of squares. A collection of three 137Cs point sources, with a combined activity of 526 kBq at the time of the measure-ment, are placed in one square, directly onto the filter, and are measured for 5 minutes by the NaI(Tl) and HPGe detectors. The measurement is repeated for

all 32 squares. The count rate in the 661 keV peak of137Cs, normalised to the maximum value measured, is plotted for each square for the NaI(Tl) detector in Fig. 10 (a) and for the HPGe detector in Fig. 10 (b). For the NaI(Tl) detector at 100% relative efficiency the count rate in the 661 keV peak is 645 cps. With a source activity of 526 kBq and 0.85 percent probability of a disin-tegration producing a 661 keV photon, 100% relative efficiency corresponds to 645/(0.85 · 526000) = 0.1% detection probability. For the HPGe detector, 100% relative efficiency corresponds to 0.005% detection probability.

0

20

40

60

80

100

(%)

0

20

40

60

80

100

(%)

Figure 10: Heat map of relative efficiency. Detector response to137_{Cs for (a)} NaI(Tl) detector and (b) HPGe detector. In both figures the intensity have been normalised to the maximum value measured. 100% relative efficiency for the NaI(Tl) detector corresponds to 0.1% detection probability of a 661 keV 137_{Cs photon and for the HPGe detector it corresponds to 0.005% detection} probability.

Summing all 32 spectra and dividing by the total live time gives us a standard spectrum, S. We want to calculate how much activity on the filter that this spectrum corresponds to. For simplicity let us assume that the activity on the filter is constant during the measurement. Our standard spectrum then corre-sponds to a measurement of 1 s with an activity of 32 · 526 kBq= 16.8 MBq 137_{Cs evenly distributed over the filter. To simulate a spectrum with} anthro-pogenic radioactivity, we add c · T · S to a background spectrum, where T is the measurement time in seconds of the background spectrum and c is a factor determining how much we add. This spiked spectrum then corresponds to c·16.8 MBq accumulated activity on the filter.

measure-ment at 4 squares. The combined activities of the133Ba sources at the time of the measurement is 376 kBq and a 100 % relative efficiency corresponds to a detection probability of a 356 keV photon of 0.06% for the NaI(Tl) detector and 0.008% for the HPGe detector. Repeating the same steps as we did for137Cs, we get a133Ba standard spectrum corresponding to c · 1.5 MBq.

To convert the accumulated activity on the filter to an air concentration, C, we use the following formula:

C = A

F · T (2)

where F is the airflow, T is the measurement time and A is the accumulated activity on the filter. Here we do not take into consideration the decrease in the activity A over time, due to the long half-life of the sources.

4

Early warning system

To function well as an early warning system we expect the system to be sensitive enough to detect anthropogenic radioactivity in a spectrum as early as possible, at the same time we do not want the system to generate more than around 1 false positive per month. As mentioned earlier, after the Chernobyl accident the maximum air concentration of 137_{Cs reached levels of 1 Bq/m}3 _and 131_I reached levels of 10 Bq/m3_{in Sweden. We would want to be able to detect this} level of concentration, preferably even lower. The filters at the filter stations are exchanged and collected every 3-4 days2 and measured 1-2 days later with high sensitivity (detection limit <25 μBq/m3) using HPGe detectors in laboratory. We would thus need our system to give an alarm within around 4 days to not be redundant.

A warning system can be designed to be general, being able to signal abnormal levels of radioactivity of all energies, or it can be specific, only signalling certain energies. In hope of increasing the sensitivity of the system we have chosen to design a specific warning system, specifically looking for radioactivity in the 661 keV and 364 keV area, where we would expect photo peaks of137_{Cs and} 131_I. The methods can be generalised to look for radioactivity at other energies too.

In this thesis we have investigated two possible algorithms for the design of the early warning system. The first one is based on the residuals between spectra, and the second is based on principal component analysis with logistic regression for classification. To evaluate the performance of the methods we calculate the false alarm rate and the miss rate defined as:

2_{This does not apply to the filter station in Kista where the filters are exchanged every 28} hours.

False alarm rate = False positives

False positives + True negatives (3)

Miss rate = False negatives

False negatives + True positives (4)

The false alarm rate is the proportion of the unspiked spectra for which the test prediction is spiked (false alarm), and the miss rate is the proportion of the spiked spectra for which the test prediction is unspiked (missed alarm).

4.1

Residuals

Our first approach is based on looking at the residuals between spectra. The shape of the background spectra changes with the radon concentration in the air, making it hard to detect anthropogenic radioactivity simply by subtracting a mean background spectrum from the suspect spectrum. However, it might be possible to find, for each suspect spectrum, a background spectrum with a similar level of radon daughters. We could then perform a fit of our background spectrum to our suspect spectrum, calculate the residual and look for anomalies. If a spectrum contains anthropogenic radioactivity, after subtracting the fitted spectrum, we should only have the counts from the anthropogenic radioactivity left.

4.1.1 Method

We begin by dividing our data of N = 1741 spectra into 90% training data and 10% test data. We then start with the training of the model. For every spectrum, xi, in the training set we calculate its residual with spectrum xj, also in the training set, in the following way: first, p1xj+ p2 is fitted to xi using least squares. Secondly, the sum of the residual, yij = P

all channels

(xi− x0j), is calculated, where x0_j is the fitted spectrum. This is then repeated for each xj, with j 6= i, so that the sum of the residuals between xi and all other training spectra has been calculated. For each i we save the yij with a value closest to zero, and call this the minimum sum of residuals (MSR) for spectrum i. This is repeated for each spectrum xi in the training set, until we have a vector Y containing one MSR per spectrum. Ideally we would expect all of these values to be zero if there indeed are recurring shapes within the background spectra. We calculate the mean value, m, and standard deviation, σ, of Y . We will use these values to design the warning criterion.

To test the method, we spike 50% (87 spectra) of the test set with varying ac-cumulated activities of137_{Cs and}133_{Ba separately. We calculate the MSR like} we did before, now letting xi be a spectrum in the test set and xj a spectrum in the training set. A spectrum xi is marked as unspiked if its MSR lies in the range m ± kσ, and spiked if it lies outside, where the value for k must be

chosen optimally. A large k will lower the false alarm rate, but also increase the miss-rate and the opposite is true for a small k.

For the NaI(Tl) measurements with137Cs we find that the method works best if the fit of spectrum xj to xi is done with respect to the counts in all channels except for the ones where we expect anthropogenic radioactivity and that the residual is then calculated in this excluded area. We choose a region of 25 channels around the 661 keV peak. For the measurements with 133_{Ba we find} that the method works best if the fit is performed with respect to the counts in all channels above the 356 keV peak and the residual is calculated in a 25 channel region around the peak. For the HPGe detector, we exclude an area of 15 channels in both the137_{Cs and} 133_{Ba measurements from the fit, and then} calculate the residual in this excluded area.

4.1.2 Results NaI(Tl)

An example of the137_{Cs training data is presented for the NaI(Tl) detector in} Fig. 11 (a) and the result for a 3.2 kBq spiked test set is presented in Fig. 11 (b). The alarm limit in these plots is set to k = 1, which results in a false alarm rate of 0% in the test data. Choosing k = 0.25 results in a false alarm rate of 1.1%, or 100 false alarms every year. Tab. 1 summarises the miss rates for spectra spiked with varying activities for the different false alarm rate scenar-ios. We notice that the miss-rate does not decrease consistently with increasing activity. At activities below 3 kBq the MSR for the spiked spectra can be both negative or positive, depending on the background. As the activity is increased the MSR is more likely to be positive. As the MSR transitions from negative to positive, it will at some point not give an alarm. The method is thus only stable when all MSR are positive. We mark the activity when this is achieved in the table as ”stable”. This is achieved for activities above 3 kBq.

0 200 400 600 800 1000 1200 1400 1600 Spectrum number 200 100 0 100 200 300 400 500

Minimum sum of residuals

Residuals (Training set)

(a) 0 25 50 75 100 125 150 175 Spectrum number 0 250 500 750 1000 1250 1500 1750 2000

Minimum sum of residuals

Residuals (3.4 kBq)

Unspiked Spiked

(b)

Figure 11: Minimum sum of residuals for 137Cs detection with the NaI(Tl) detector. (a) Red dots show the training set. Light red area marks an area where a point lies within ±1σ from the mean value. This area is used as an alarm criterion to find spiked spectra. (b) Example of test data, the red dots are unspiked and the green dots are spiked with 3.4 kBq accumulated activity of137_{Cs on the filter. The x axis is an arbitrary indexing of the spectra.}

Table 1: Miss rate for 137Cs detection with the NaI(Tl) detector for different false alarm rate scenarios in the test data. A is the accumulated activity on the filter with which the spectra are spiked.

A [kBq] Miss rate [%]

false alarm rate 0% (k = 1)

false alarm rate 1.1% (k= 0.25) 0.6 44 9 1.0 1 1 1.4 12 6 1.8 1 1 2.2 1 0 2.6 0 0 3.0 (stable) 1 0 3.4 0 0

An example of the133Ba training data is presented in Fig. 12 (a) and the result for 3.2 kBq spiked test set is presented in Fig. 12 (b). The alarm limit in these plots is set to k = 1, which result in a false alarm rate of 0% in the test data. The miss rate is presented in Tab. 2. Stability is achieved first above 7 kBq.

0 200 400 600 800 1000 1200 1400 1600 Spectrum number 200 100 0 100 200

Minimum sum of residuals

Residuals (Training set)

(a) 0 25 50 75 100 125 150 175 Spectrum number 0 2000 4000 6000 8000 10000 12000

Minimum sum of residuals

Residuals (7.0 kBq)

Unspiked Spiked

(b)

Figure 12: Minimum sum of residuals for 133_{Ba detection with the NaI(Tl)} detector. (a) Red dots show the training set. Light red area marks an area where a point lies within ±3σ from the mean value. This area is used as an alarm criterion to find spiked spectra. (b) Example of test data, the red dots are unspiked and the green dots are spiked with 7 kBq accumulated activity of 133_{Ba on the filter. The x axis is an arbitrary indexing of the spectra.}

Table 2: Miss rate for133_{Ba detection with the NaI(Tl) detector for different} false alarm rate scenarios in the test data. A is the accumulated activity on the filter with which the spectra are spiked.

A [kBq] Miss rate [%]

False alarm rate 0% (k = 3)

False alarm rate 1.1% (k= 2) 1.0 76 70 2.0 11 6 3.0 2 2 4.0 2 2 5.0 1 0 6.0 0 0 7.0 (stable) 0 0 HPGe

An example of the137_{Cs training data for the HPGe detector is presented in} Fig. 13 (a). Fig. 13 (b) shows the test data spiked with 1 kBq of137_{Cs. Here} the value for k is chosen as 2, which results in a false alarm rate in the test data of 0%. The miss rate for the different false alarm rate scenarios are presented in Tab. 3 for varying amounts of accumulated activities. Here, we achieve stability with an accumulated activity of above 1 kBq.

0 200 400 600 800 1000 1200 1400 1600 Spectrum number 0 2 4 6

Minimum sum of residuals

Residuals (Training set)

(a) 0 25 50 75 100 125 150 175 Spectrum number 0 10 20 30 40 50 60 70

Minimum sum of residuals

Residuals (1.2 kBq)

Unspiked Spiked

(b)

Figure 13: Minimum sum of residuals for137_{Cs detection with the HPGe} detec-tor. (a) Red dots show the training set. Light red area marks an area where a point lies within ±2σ from the mean value. This area is used as an alarm crite-rion to find spiked spectra. (b) Example of test data, the red dots are unspiked and the green dots are spiked with 1.2 kBq accumulated activity of 137_{Cs on} the filter. The x axis is an arbitrary indexing of the spectra.

Table 3: Miss rate for137_{Cs detection with the HPGe detector for different false} alarm rate scenarios in the test data. A is the accumulated activity on the filter with which the spectra are spiked.

A [kBq] Miss rate [%]

False alarm rate 0% (k = 2)

False alarm rate 1.1% (k= 1) 0.4 70 56 0.6 20 12 0.8 2 1 1.0 0 0 1.2 (stable) 0 0

An example of the133Ba training data is presented in Fig. 14 (a) and the result for 3.2 kBq spiked test set is presented in Fig. 14 (b). The alarm limit in these plots is set to k = 2, which result in a false alarm rate of 0% in the test data. The miss rate is presented in Tab. 4. Stability is achieved above 1 kBq.

0 200 400 600 800 1000 1200 1400 1600 Spectrum number 8 6 4 2 0 2 4

Minimum sum of residuals

Residuals (Training set)

(a) 0 25 50 75 100 125 150 175 Spectrum number 0 20 40 60 80 100 120 140

Minimum sum of residuals

Residuals (1.2 kBq)

Unspiked Spiked

(b)

Figure 14: Minimum sum of residuals for133_{Ba detection with the HPGe} detec-tor. (a) Red dots show the training set. Light red area marks an area where a point lies within ±2σ from the mean value. This area is used as an alarm crite-rion to find spiked spectra. (b) Example of test data, the red dots are unspiked and the green dots are spiked with 1.2 kBq accumulated activity of 133Ba on the filter. The x axis is an arbitrary indexing of the spectra.

Table 4: Miss rate for133_{Ba detection with the HPGe detector for different false} alarm rate scenarios in the test data. A is the accumulated activity on the filter with which the spectra are spiked.

A [kBq] Miss rate [%]

False alarm rate 0% (k = 2)

False alarm rate 1.1% (k= 1.75)

0.6 37 36

0.8 8 8

1.0 2 2

1.2 (stable) 0 0

4.2

Principal component analysis

Each NaI(Tl) spectrum collected consists of 1024 channels with a varying amount of counts in each. We could then choose a 1024 dimensional basis, consisting of base vectors ab with the counts of channel b in the bth row and 0 in the rest and describe each spectrum as a linear combination xi =P1024_b=1 ab. However, describing the data with this many dimensions may not be the most efficient way. If we decide to just drop some dimensions, we will lose the information from that vector and gain no benefit from ever having known it. However, if we can find a new set of vectors as a combination of the old ones and order them by how well they describe the data, we will know which vectors are important. We can then throw away as many of the least important vectors as we want and, since these new vectors are combinations of the old, still keep the most important parts of all of our old vectors. This is the idea of principal compo-nent analysis (PCA): to reduce the dimensions of the data while loosing as little information as possible.

Assume we have a set D of spectra, each with d channels. Let D = {x1, x2, ..., xn}, where xiis a d × 1 dimensional vector containing the counts in each bin in spec-trum i, be the set of all measurements. Further assume that we have calculated the mean in each channel and subtracted it from all measurements so that the counts in each channel is centred around 0. This does not change how the data points are positioned relative to each other. We also assume the variance in each channel of the set is 1.

Let B be the d × n matrix with the spectra vectors as columns.

B =x1 ... xn 

(5) We wish to find a d × m projection matrix W such that

is a projection of our data on a m dimensional space (m < d), spanned by the column vectors w1, ..., wm of W, without loosing too much information. Intu-itively, if the variance in one direction of the data is very low, removing that dimension will not cause a significant data loss. On the other hand, ignoring dimensions with very high variance will cause the compressed data to be very different from the original data. Motivated by this, when reducing the dimen-sions we try to keep the directions with the maximum variance. We define the d × d data covariance matrix S as:

S = 1 nBB

T ₍₇₎

It can be shown that Sii is the variance of the i th channel and Sij is the covari-ance between the i th and j th channels. Since this implies that Sij = Sji, the matrix S is symmetric.

Solving the optimisation problem of finding the dimensions, or directions, of the data in which the variance is maximal, gives the result that it is the eigenvectors of S corresponding to the largest eigenvalues (see derivation in e.g. Ref. [9]). These eigenvectors are known as the principal components.

The matrix W is thus formed by letting the columns of W be the m eigenvectors corresponding to the m largest eigenvalues of the data covariance matrix S. By projecting our spectrum onto the subspace spanned by the columns in W, as in Eq. 6, we can describe every data point as the m-dimensional coordinate vector zi.

We can take advantage of this in the design of a warning system in the following way. If a spectrum exhibits features very different from the background the base vectors wiwould need to be combined in a way that differs from the background measurements to explain the feature and thus the projection zi would differ for a spiked and an unspiked spectrum. We use this to identify a spectrum with features deviating from that of the background.

To implement PCA we use the decomposition.PCA class from the Python pack-age sci-kit learn [10].

4.2.1 Logistic regression

From the PCA, each spectrum xiis represented as the m-dimensional vector zi, where m is the number of principal components we decide to save. We now want to classify the data points based on their coordinates into spiked or unspiked. This can be done with logistic regression, which is a method for classifying data. It fits a Sigmoid function to the data, defined as:

h(zi) = 1 1 + e−βT_z

where βT = [β0, β1, ..., βm−1] is a vector containing the fitting parameters. h(zi) is a measure of the probability that ziis positive (or ”spiked”). To fit h we train it on data containing both spiked (h = 1) and unspiked (h = 0) spectra.

To implement logistic regression we use the linear model.LogisticRegression class from the Python package sci-kit learn [10].

4.2.2 Method

We divide the data of N = 1741 spectra into 90% training data and 10% test data, where each data point is a measurement of 1 hour. We spike 50% (783 spectra) of the training data with 1 kBq 137Cs for the NaI(Tl) detector and 0.8 kBq for the HPGe detector. To minimise influence from other channels, we crop the data to only look at a region of interest (ROI) around137Cs. For the NaI(Tl) detector we have chosen a ROI of 25 channels, and for the HPGe 15 channels. The principal components are found using the training data. For easy visualisation we have chosen 2 principal components (m = 2). We then perform a logistic regression on the training data to fit the Sigmoid function, h(zi). We set the limit that a spectrum is classified as spiked if h(zi) ≥ 0.5. We evaluate the sensitivity of the algorithm by testing it with the test data. 50% (87 spectra) of the test data are spiked with different amounts of137_{Cs for each} test. The testing is carried out by projecting the test data on the previously found principal components and let the fitted Sigmoid function predict whether the point contained spiked data or not. We evaluate the method by looking at the false alarm rate and miss rate in the test data.

The training and testing is then repeated in the same way using 133Ba as a source. For the NaI(Tl) detector we spike the training data with 2 kBq of133Ba and choose a range of 70 channels. For the HPGe detector we train with 0.7 kBq133Ba and calculate the PCA in a range of 15 channels.

4.2.3 Results

NaI(Tl)

The coordinates of the projected data is presented in Fig. 15(a) for the training data spiked with137_{Cs. An example of test data spiked with 1.2 kBq of}137_Cs is presented in figure 15 (b).

For the 133Ba data, the PCA is presented in Fig. 16 and the miss rate is presented in Tab. 6.

5 0 5 10 15 20 25 30 PC1 2 0 2 4 6 PC2

NaI detector (Training set)

Unspiked Spiked (a) 5 0 5 10 15 20 25 30 PC1 2 0 2 4 6 PC2

NaI detector (Test set, 1.2 kBq)

Unspiked Spiked

(b)

Figure 15: 2-component PCA with logistic regression for137Cs detection with NaI(Tl) detector. Each plot is showing the coordinate for data points projected onto the principal components. Logistic regression have been performed on the training data and is used to predict the nature of the data point. A data point in the light red area is predicted to be an unspiked spectrum, while a data point in the green area is predicted to be a spiked spectrum. (a) Training data spiked with 1 kBq of137_{Cs. (b) Test data spiked with 1.2 kBq of}137_{Cs. The spectra} were cropped before performing the PCA to only look at the ROI.

Table 5: Miss rate for NaI(Tl) detector test data with varying accumulated activities of137_{Cs on the filter.}

A [kBq] Miss rate [%] False alarm rate 0%

0.6 20

0.8 10

1.0 1

1.2 0

Table 6: Miss rate for NaI(Tl) detector test data with varying accumulated activities of133_{Ba on the filter.}

A [kBq] Miss rate [%] False alarm rate 0%

1.0 44

1.4 9

1.8 1

0 10 20 30 40 50 60 PC1 3 2 1 0 1 2 3 4 PC2

NaI detector (Training set)

Unspiked Spiked (a) 0 10 20 30 40 50 60 PC1 3 2 1 0 1 2 3 4 PC2

NaI detector (Test set, 2.2 kBq)

Unspiked Spiked

(b)

Figure 16: 2-component PCA with logistic regression for133Ba detection with NaI(Tl) detector. Each plot is showing the coordinate for data points projected onto the principal components. Logistic regression have been performed on the training data and is used to predict the nature of the data point. A data point in the light red area is predicted to be an unspiked spectrum, while a data point in the green area is predicted to be a spiked spectrum. (a) Training data spiked with 2 kBq of133_{Ba. (b) Test data spiked with 2.2 kBq of}133_{Ba. The spectra} were cropped before performing the PCA to only look at the ROI.

HPGe

The results for PCA with the HPGe detector is presented in Fig. 17 (a) for the 137_{Cs spiked training data and an example for the test data is presented in (b)} with data spiked with 0.7 kBq of137Cs. The false alarm rate is 0% in the test data, and the miss rate for varying amounts of activity is presented in Tab. 7.

Table 7: Miss rate for HPGe detector test data with varying accumulated ac-tivities of137_{Cs on the filter.}

A [kBq] Miss rate [%] False alarm rate 0%

0.4 62

0.5 20

0.6 1

4 2 0 2 4 PC1 4 2 0 2 4 6 PC2

HPGe detector (Training set)

Unspiked Spiked (a) 4 2 0 2 4 PC1 4 2 0 2 4 6 PC2

HPGe detector (Test set, 0.7 kBq)

Unspiked Spiked

(b)

Figure 17: 2-component PCA with logistic regression for137Cs detection with HPGe detector. Each plot is showing the coordinate for data points projected onto the principal components. Logistic regression have been performed on the training data and is used to predict the nature of the data point. A data point in the light red area is predicted to be an unspiked spectrum, while a data point in the green area is predicted to be a spiked spectrum. (a) Training data spiked with 0.8 kBq of137_{Cs. (b) Test data spiked with 0.7 kBq of}137_{Cs. The spectra} were cropped before performing the PCA to only look at the ROI.

We repeat the measurements with 133_{Ba. We train the method with 0.7 kBq} spiked data. The training data is presented in Fig. 18 (a) and an example of the test data, spiked with 0.6 kBq is presented in Fig. 18 (b). The false alarm rate in the test data is 0%, and the miss rate for varying activities is presented in Tab. 8.

Table 8: Miss rate for HPGe detector test data with varying accumulated ac-tivities of133_{Ba on the filter.}

A [kBq] Miss rate [%] False alarm rate 0%

0.3 60

0.4 29

0.5 7

4 2 0 2 4 6 8 10 PC1 4 2 0 2 4 6 8 10 PC2

HPGe detector (Training set)

Unspiked Spiked (a) 4 2 0 2 4 6 8 10 PC1 4 2 0 2 4 6 8 10 PC2

HPGe detector (Test set, 0.6 kBq)

Unspiked Spiked

(b)

Figure 18: 2-component PCA with logistic regression for133Ba detection with HPGe detector. Each plot is showing the coordinate for data points projected onto the principal components. Logistic regression have been performed on the training data and is used to predict the nature of the data point. A data point in the light red area is predicted to be an unspiked spectrum, while a data point in the green area is predicted to be a spiked spectrum. (a) Training data spiked with 0.7 kBq of133_{Ba. (b) Test data spiked with 0.6 kBq of}133_{Ba. The spectra} were cropped before performing the PCA to only look at the ROI.

5

Discussion

With these methods we managed to achieve the same order of sensitivity at the air filter station, as the previous report achieved in the lab. In this thesis we did not perform a measurement of the absolute sensitivity in the lab, but it would be interesting to be able to compare with the previous results. We now discuss a few suggestions for developing the methods further:

The predictions of the residuals method is not consistent for spectra spiked with equal amounts of activity, especially for the NaI(Tl) detector. More data would be needed to be confident on the false alarm rate and miss rates presented in this thesis. We suspect that this method depends much on the background at the time of the measurement. For the NaI(Tl) the overlap between the nuclides of interest and the radon daughter peaks is considerable and complicates the anal-ysis. To develop the residual method further we could try plotting the residual in different areas against each other, like the area around the137_{Cs peak against} the area around the 1764 keV peak. An increase in the count rate in the area around137_{Cs could be due to an increase in the 609 keV peak interfering, or} from the Compton continuum of higher energy radon daughter peaks. Plotting the residual in the 661 keV peak against the residual in the radon daughter peak at 1764 keV could then give us more information on whether the increase is due to increased levels of radon or if it is due to a 137Cs source. A higher sensitivity could possibly also be achieved by tweaking some of the parameters in the model. We believe the most important parameters to be the range in

which the residual is calculated, the range in which the fit of the spectra to each other is performed and the number of standard deviations (k ) chosen as the alarm limit.

The PCA method performs well for both the NaI(Tl) and the HPGe detectors. To improve the sensitivity of the PCA method we believe some of the more im-portant parameters to tweak are the weights assigned to each class in the fitting of the logistic regression, the activity of the spiked training set and the range in which PCA is performed. By changing the weights, it is possible to favour detecting spiked spectra at a lower activity, in exchange for a higher probability for false positives. The activity with which we spike the training set will also affect the fitting of the logistic regression. One could try to spike the training data with lower activities to see if that improves the sensitivity. Finally, the range in which the PCA is performed is important to be able to separate spiked and unspiked spectra. Ideally, the PCA would be performed on the full spec-trum and be able to detect anthropogenic radioactivity of any energy. Trying this resulted in no separation of the spiked and unspiked spectra in our case. We suspect that this might be due to the large variance in the radon daughter peaks overshadowing the large variance in the spiked channels. This would lead the method to select principal components in directions in which the background maximally varies. Two ideas to solve this would be to either train the method on data spiked with high levels of activity or add more principal components.

Another method we could try is to look at the relation between peaks. For the NaI(Tl) detector, we could try to calculate the ratio of the peak area for the 356 keV peak and the 1764 keV as well as the 609keV/1764 keV peak ratio. At normal background we would expect this to be close to constant if there is equilibrium between the radon daughter. However if there is131_{I in the} spec-trum its peak at 364 keV will interfere with the 356 keV peak and cause the counts to deviate from the constant relation. The same thing applies to the 609 keV/ 1764 keV peak ratio if there is137Cs in the spectrum, which has a peak at 661 keV interfering with the 609 keV peak. For the HPGe detector the peaks would not interfere, due to its high energy resolution, but we could define a ROI for131I and137Cs and look at the variation in time of the count rate in these channels.

To gain a higher sensitivity with the system a higher false alarm rate could be accepted. Too many false alarms, however, will drown out the true alarms. We suggest that a maximum of 1 false alarm per month is tolerable. This corre-sponds to a false alarm rate of 0.1%. The alarm criteria for both methods would need to be adjusted to allow this false alarm rate. An attempt was made in this thesis to investigate the gain in sensitivity from a high false alarm rate with the residuals method, but due to insufficient data no conclusion can be drawn. To counteract the increase in false alarm rate it would perhaps be possible to com-bine different alarm criteria and request them all to trigger to send an alarm. It is important to note that the sensitivity of the alarm systems we combine

should be similar, since the sensitivity would only be as good as the lowest one of the systems.

In addition to improving the algorithms we also suggest three analyses to be performed to better understand the data. Firstly, to investigate the impact of the background on the sensitivity of the methods, we suggest comparing the methods for a spiked spectrum with a high radon background with one with a low radon background. Secondly, we suggest comparing different measurement times. In this thesis all spectra analysed were collected for 1 hour. Finally, we suggest investigating the seasonal fluctuations of the data. It would be inter-esting to see how well the background spectra in summer can be used to detect anthropogenic radioactivity in the winter, and vice versa. Furthermore, a long term test of the two methods at the filter station would be necessary to fully evaluate the methods performances. It is important to note, before implement-ing the results in this thesis at any of the other filter stations in Sweden, that the setup in Kista differs from the others. The exposed filter area, for example, is roughly twice the size in comparison with the ones at the other stations. It could therefore be good to test the performance of the system at the other stations too.

Comparing the two detectors, the HPGe detector is over all performing better than the NaI(Tl) in sensitivity. But, in practice, the gain in sensitivity is not worth the high cost of a HPGe detector. It could be interesting to look at other detectors. LaBr3 is an inorganic scintillator with improved energy resolution compared to NaI(Tl). It is more expensive than the NaI(Tl) detector, but less expensive than the HPGe detector.

6

Conclusions

In this thesis we have investigated two algorithms for the detection of anthro-pogenic radioactivity in the atmosphere and the minimum detectable activity by a HPGe and a NaI(Tl) detector. We have chosen to focus on the detection of 137_{Cs and} 131_{I, which are two common nuclides released during a nuclear} power accident. A summary of the sensitivity is given in Tab. 9. The highest sensitivity achieved for detecting137_{C is with the PCA method, where we} man-aged to achieve a minimum detectable activity of around 1 kBq of activity on the filter for both the HPGe and NaI(Tl) detectors. With an airflow of 1600 m3/h and a measurement time of 1h this corresponds to an air concentration of 0.6 Bq/m3. This would have been enough to be able to detect the release from Chernobyl when it reached its maximum, but not the release from the Fukushima accident. For133Ba we managed to achieve a sensitivity in the or-der of 2 kBq for the NaI(Tl) detector and 1 kBq for the HPGe detector with the PCA method. This corresponds to 1.2 Bq/m3_{respectively 0.6 Bq/m}3_{in air} concentration. The residuals method performed, at best, equally good as the PCA method, but performed much worse for the NaI(Tl) detector.

Table 9: Sensitivity of the residuals and PCA methods for the NaI(Tl) and HPGe detectors. Numbers are given in kBq of accumulated activity on the filter. Method 137_Cs [kBq] 133_Ba [kBq]

NaI(Tl) HPGe NaI(Tl) HPGe

Residuals ∼3 ∼1 ∼7 ∼ 1

PCA ∼1 ∼1 ∼2 ∼1

With a sensitivity of 1 kBq of accumulated activity, an air concentration below 0.02 Bq/m3 will not accumulate enough activity on the filter before the filter is changed.

7

Acknowledgements

I would very much like to thank my supervisors Johan Kastlander and Torbj¨orn B¨ack for guiding me and helping me complete this thesis. I would also like to thank the workshop personnel at FOI for adapting the filter station so that the HPGe detector could be introduced to the set up.

References

[1] M. Goliath, C. Hellesen, L. Karlkvist, J. Kastlander, H. Olsson, and C. S¨oderstr¨om, “Radionuclide particles in ground level air in sweden during 2018,” Tech. Rep. FOI-R–4778–SE, FOI, 2019.

[2] J. Kastlander and C. S¨oderstr¨om, “Early warning system at the national filter stations,” Tech. Rep. FOI-R–3581–SE, FOI, 2012.

[3] A. Mauring, T. G¨afvert, and T. B. Aleksandersen, “Implications for anal-ysis of 226Ra in a low-level gamma spectrometry laboratory due to varia-tions in radon background levels,” Applied Radiation and Isotopes, vol. 94, pp. 54–59, 2014.

[4] P. Bossew, “A very long-term HPGe-background gamma spectrum,” Ap-plied Radiation and Isotopes, vol. 62, no. 4, pp. 635–644, 2005.

[5] M. H´yˇza and P. Rul´ık, “Low-level atmospheric radioactivity measurement using a NaI(TI) spectrometer during aerosol sampling,” Applied Radiation and Isotopes, vol. 126, pp. 225–227, 2017.

[6] P. Kov´aˇr and J. ˇSolc, “Subtraction of natural radiation contribution from gamma-ray spectra measured by HPGe detector,” Applied Radiation and Isotopes, vol. 134, pp. 167–171, 2018.

[7] National Nuclear Data Center, “NuDat 2 database.” https://www.nndc. bnl.gov/nudat2/. [Online; accessed 12-March-2020].

[8] C. S¨oderstr¨om, J. Kastlander, M. Meister, N. Tooloutalaie, S. Ban, N.-O. Bergkvist, L.-E. De Geer, C. During, K. Elmgren, T. Fritioff, H. Rameb¨ack, and A. Tovedal, “Radioaktiva utsl¨app fr˚an k¨arnkraftsolyckan i Fukushima Daiichi,” Tech. Rep. FOI-R–3458–SE, FOI, 2012.

[9] M. P. Deisenroth, A. A. Faisal, and C. S. Ong, “Dimensionality reduction with principal component analysis,” in Mathematics for Machine Learning, Cambridge University Press, 2020.

[10] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Pas-sos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay, “Scikit-learn: Machine learning in Python,” Journal of Machine Learning Re-search, vol. 12, pp. 2825–2830, 2011.

A

Code for programming the detectors

A.1

Run Filter measurements

REM R u n F i l t e r M e a s u r e m e n t s : S t a r t s two d e t e c t o r s REM w a i t s t h e t i m e i n s e c o n d s s p e c i f i e d i n LiveTime , s a v e s REM t h e f i l e s a s ASCII . REM T h i s i s i t e r a t e d t h e number o f t i m e s s p e c i f i e d by N l o o p s REM ” S e t t i n g s − Change t h e s e ” SET VARIABLE ” N a I D e t e c t o r ” , 2 SET VARIABLE ” HpGeDetector ” , 1 SET VARIABLE ” N l o o ps ” , 999 SET VARIABLE ” LiveTime ” , 900

REM ” Program − Do n o t change a n y t h i n g h e r e ” SET DETECTOR $ ( N a I D e t e c t o r )

SET PRESET CLEAR

SET PRESET LIVE $ ( LiveTime )

SET DETECTOR $ ( HpGeDetector ) SET PRESET CLEAR

SET PRESET LIVE $ ( LiveTime )

SET DETECTOR $ ( N a I D e t e c t o r ) LOOP 999

SET DETECTOR $ ( N a I D e t e c t o r ) CLEAR

START

SET DETECTOR $ ( HpGeDetector ) CLEAR START SET DETECTOR $ ( N a I D e t e c t o r ) WAIT FILL BUFFER SET DETECTOR 0

SAVE ”C: \ User \ Exjobb−T i d i g V a r n i n g \ c u r r e n t \ NaI ? ? ? . SPE”

SET DETECTOR $ ( HpGeDetector ) WAIT

FILL BUFFER SET DETECTOR 0

SET DETECTOR $ ( N a I D e t e c t o r ) END LOOP

A.2

Execute

REM E x e c u t e : C a l l s R u n F i l t e r M e a s u r e m e n t s REM t h e number o f t i m e s s p e c i f i e d a f t e r LOOP.

REM E n a b l e s more i t e r a t i o n s o f R u n F i l t e r M e a s u r e m e n t s REM than t h e maximum 999 o f t h e LOOP

SET DETECTOR 1 LOOP 999

CALL ”C: \ User \ Exjobb−T i d i g V a r n i n g \ R u n F i l t e r M e a s u r e m e n t s . JOB”

SET DETECTOR 1 END LOOP