Method for identifying a moving radiation source

1 . A method for identifying a moving radiation source by a radiation portal monitoring system, the radiation portal monitoring system comprising a radiation portal monitor comprising a plurality of radiation detectors configured to detect ionizing radiation of the moving radiation source and to generate a detection signal responsive to detection of the ionizing radiation, and at least one processor, the method comprising the at least one processor executing the steps of:

providing an identification machine learning model;

receiving labelled static identification training data generated by radiation detection of a plurality of known static radiation sources;

introducing to the static identification training data modifications representing detection signal alterations caused by radiation source movement through the radiation portal monitor to obtain pseudo-dynamic identification training data;

training the identification machine learning model using the pseudo-dynamic identification training data; and

identifying the moving radiation source from the detection signal using the trained identification machine learning model.

2 . The method according to claim 1 , wherein identifying the moving radiation source from the detection signal using the trained identification machine learning model further comprises a threshold-based sample selection step, the threshold-based sample selection step comprising the at least one processor executing the steps of:

sampling a count rate trace of the detection signal of a radiation detector of the radiation portal monitoring system at a sampling rate R;

selecting response spectrum samples of the sampled count rate trace with count rates above a count rate threshold; and

providing the selected response spectrum samples as an input to the trained identification machine learning model.

3 . The method according to claim 2 , wherein the at least one processor sets the count rate threshold as lying n times the fluctuations σ of a count rate background above a local minimum of the count rate trace.

4 . The method according to claim 2 , wherein the threshold-based sample selection step further comprises the at least one processor executing an energy selection step by selecting samples within one or more predetermined energy windows of interest.

5 . The method according to claim 1 , wherein identifying the moving radiation source from the detection signal using the trained identification machine learning model further comprises the at least one processor executing a detector selection step by selecting a detection signal of one or more radiation detectors of the radiation portal monitoring system with a largest count rate increase relative to a background count rate as an input to the trained identification machine learning model.

6 . The method according to claim 5 , wherein, when the at least one processor selects detection signals of a group of radiation detectors in the detector selection step, the detection signals of at least a portion of the group of radiation detectors are shifted in time by the at least one processor.

7 . The method according to claim 1 , wherein the identification machine learning model comprises an artificial neural network.

8 . The method according to one claim 1 , wherein the method further comprises the at least one processor executing the steps of:

providing an alarming machine learning model;

receiving alarming training data labelled as potential threat source worthy of further analysis or as benign source not worthy of further analysis;

training the alarming machine learning model using the alarming training data;

classifying the moving radiation source from the detection signal using the trained alarming machine learning model as potential threat source or benign source; and

in case of the moving radiation source being classified as potential threat source, identifying the moving radiation source from the detection signal using the trained identification machine learning model.

9 . The method according to claim 8 , wherein the alarming training data is generated by one or more of: count rate evolution data, count rate evolution data within one or more predetermined energy windows of interest, or spectral evolution data of the radiation portal monitor, labelled using secondary inspection data.

10 . The method according to claim 8 , wherein the alarming training data is generated by one or more of: count rate evolution data, count rate evolution data within one or more predetermined energy windows of interest, or spectral evolution data of the radiation portal monitor, conditioned by introducing modifications representing predetermined anomalies of response spectra of the detection signal caused by the presence of threat sources with varying shielding configurations.

11 . The method according to claim 8 , wherein the alarming training data is generated by one or more of count rate evolution data, count rate evolution data within one or more predetermined energy windows of interest, spectral evolution data, of another radiation portal monitor, labelled using secondary inspection data, wherein the count rate or spectral evolution data is conditioned by introducing modifications representing differences in detector configuration with respect to the other radiation portal monitor.

12 . The method according to claim 8 , wherein the method further comprises the at least one processor executing a conventional alarming algorithm using a dynamic threshold on an alarming function of one or more of: count rate evolution data, count rate evolution data within one or more predetermined energy windows of interest, or spectral evolution data of the radiation portal monitor, wherein the at least one processor classifies the moving radiation source as potential threat source if the alarming function exceeds the dynamic threshold.

13 . The method according to claim 12 , wherein prior to and/or while training the alarming machine learning model, the method further comprises the at least one processor classifying the moving radiation source from the detection signal as potential threat source or benign source using the conventional alarming algorithm.

14 . The method according to claim 12 , wherein the at least one processor uses the detection signal of the radiation portal monitor together with output data of the conventional alarming algorithm as alarming training data to train the alarming machine learning model.

15 . The method according to claim 12 , wherein while training the alarming machine learning model, the method further comprises the at least one processor determining and comparing the accuracies of the classification of the moving radiation source from the detection signal as potential threat source or benign source using the conventional alarming algorithm and using the alarming machine learning model, and terminating the using of the conventional alarming algorithm if the accuracy of the classification using the alarming machine learning model exceeds the accuracy of the classification using the conventional alarming algorithm.

16 . The method according to claim 8 , wherein the method further comprises the at least one processor training the alarm machine learning model for occupancy classification using one or more of count rate evolution data, count rate evolution data within one or more predetermined energy windows of interest, or spectral evolution data of the radiation portal monitor or another radiation portal monitor, labelled by true or false occupancy, as occupancy training data for the alarm machine learning model.

17 . The method according to claim 1 , wherein the method further comprises the at least one processor executing the steps of:

providing an occupancy machine learning model;

receiving occupancy training data labelled by true occupancy or false occupancy;

training the occupancy machine learning model using the occupancy training data; and

classifying a detection signal as being associated with true occupancy or with false occupancy using the trained occupancy machine learning model;

in case of a detection signal being classified as being associated with true occupancy, identifying a moving radiation source from the detection signal using the trained identification machine learning model.

18 . The method according to claim 17 , wherein the occupancy training data is generated by one or more of: count rate evolution data, count rate evolution data within one or more predetermined energy windows of interest, or spectral evolution data of the radiation portal monitor or another radiation portal monitor.

19 . The method according to claim 17 , wherein the at least one processor uses the detection signal of the radiation portal monitor together with occupancy data of conventional occupancy sensors and/or occupancy data from time delayed detection signals of an array of radiation detectors arranged along a direction of passage through the radiation portal monitor as occupancy training data.

20 . The method according to claim 17 , wherein the at least one processor uses occupancy sensing data from another radiation portal monitor, conditioned by introducing modifications representing differences in detector configuration with respect to the other radiation portal monitor, as occupancy training data.

21 . The method according to claim 1 , wherein the identification machine learning model is validated using a validation dataset of detection signals labelled by using secondary inspection data.

22 . The method according to claim 1 , wherein the labelled static identification training data is generated by a plurality of measurements per known static radiation source at several distances between the known static radiation source and radiation detectors, wherein the radiation detectors are part or not part of the radiation portal monitor.

23 . The method according to claim 1 , wherein the modifications introduced to the labelled static identification training data represent detection signal alterations caused by one or more of: alteration of response spectrum of the detection signal due to different relative positions of the radiation detectors, broadening of peak structures in response spectrum of the detection signal due to shielding of the radiation source by a container containing the radiation source, alteration of the response spectrum due to variation of angle of incidence of the ionizing radiation, changes in the detection signal background due to traffic passing through the radiation portal monitor and at least partially shielding background radiation, shifting of spectra of the radiation source to higher or lower energies due to imperfect radiation detector calibration, response spectrum alteration due to varying speed of the radiation source movement, response spectrum alteration due to detector under-performance, and/or response spectrum alteration due to statistical nature of the radiation detection.

24 . A radiation portal monitoring system for detecting and identifying a moving radiation source, comprising a radiation portal monitor comprising a plurality of radiation detectors configured to detect ionizing radiation of the moving radiation source and to generate a detection signal responsive to detection of the ionizing radiation, and at least one processor configured to:

provide an identification machine learning model;

receive labelled static identification training data generated by a plurality of measurements of static radiation sources;

condition the static identification training data by introducing modifications representing detection signal alterations caused by radiation source movement through the radiation portal monitor to obtain pseudo-dynamic identification training data;

train the identification machine learning model using the pseudo-dynamic identification training data; and

identify the moving radiation source from the detection signal using the trained identification machine learning model.

25 . The radiation portal monitoring system according to claim 24 , wherein the at least one processor is further configured to:

sample a count rate trace of the detection signal of a radiation detector of the radiation portal monitoring system at a sampling rate R;

select response spectrum samples of the sampled count rate trace with count rates above a count rate threshold; and

provide the selected response spectrum samples as an input to the trained identification machine learning model.

26 . The radiation portal monitoring system according to claim 24 , wherein the radiation portal monitor comprises one or more panel radiation detectors comprising a plurality of adjoining plastic scintillator slabs, a plurality of silicon photomultiplier sensors arranged at an edge of at least one of the plastic scintillator slabs and configured to detect scintillation light generated in the scintillator slabs responsive to the radiation events, a plurality of signal processing units each connected to one of the silicon photomultiplier sensors, and a joint analyzing circuit configured to perform signal analysis to determine the energy of the detected radiation events.

27 . A non-transitory computer-readable medium comprising computer program code for identifying a radiation source moving through a radiation portal monitor of a radiation portal monitoring system, the radiation portal monitor comprising a plurality of radiation detectors configured to detect ionizing radiation of the radiation source and to generate a detection signal responsive to detection of the ionizing radiation, the computer program code configured to control at least one processor of the radiation portal monitoring system such that the at least one processor executes the steps of:

providing an identification machine learning model;

receiving labelled static identification training data generated by a plurality of measurements of static radiation sources;

conditioning the static identification training data by introducing modifications representing detection signal alterations caused by radiation source movement through the radiation portal monitor to obtain pseudo-dynamic identification training data;

training the identification machine learning model using the pseudo-dynamic identification training data; and

identifying the moving radiation source from the detection signal using the trained identification machine learning model.

28 . A method for identifying a moving radiation source by a radiation portal monitoring system, the radiation portal monitoring system comprising a radiation portal monitor comprising a plurality of radiation detectors configured to detect ionizing radiation of the moving radiation source and to generate a detection signal responsive to detection of the ionizing radiation, and at least one processor, the method comprising the at least one processor executing the steps of:

providing an identification machine learning model; and

identifying the moving radiation source from the detection signal using the identification machine learning model, wherein

the identification machine learning model is trained using pseudo-dynamic identification training data, wherein

the pseudo-dynamic identification training data is obtained by introducing to labelled static identification training data modifications representing detection signal alterations caused by radiation source movement through the radiation portal monitor, and wherein

the labelled static identification training data is generated by radiation detection of a plurality of known static radiation sources.

29 . A non-transitory computer-readable medium comprising computer program code configured to control at least one processor such that the at least one processor executes the steps of the method according to claim 28 .

30 . A method for generating pseudo-dynamic identification training data for an identification machine learning model configured to identify a moving radiation source from a detection signal of a radiation portal monitoring system, the radiation portal monitoring system comprising a radiation portal monitor comprising a plurality of radiation detectors configured to detect ionizing radiation of the moving radiation source and to generate a detection signal responsive to detection of the ionizing radiation, the method comprising the steps of:

generating static identification training data from radiation detection of a plurality of known static radiation sources; and

generating pseudo-dynamic identification training data by introducing to the static identification training data modifications representing detection signal alterations caused by radiation source movement through the radiation portal monitor.

31 . The method according to claim 30 , wherein radiation detection of a plurality of known static radiation sources comprises one or more radiation detectors executing a plurality of measurements per isotope at several distances between the isotopes and the one or more radiation detectors.

32 . The method according to claim 30 , wherein static identification training data is further generated by simulation of spectra of known isotopes or by accessing a database with spectra of known isotopes.

33 . A computer-implemented method of training an identification machine learning model to obtain a trained identification machine learning model configured to identify a moving radiation source from a detection signal of a radiation portal monitoring system, the radiation portal monitoring system comprising a radiation portal monitor with a plurality of radiation detectors configured to detect ionizing radiation of the moving radiation source and to generate a detection signal responsive to detection of the ionizing radiation, the method comprising at least one processor executing the steps of:

providing the identification machine learning model;

receiving labelled static identification training data generated by radiation detection of a plurality of known static radiation sources;

introducing to the static identification training data modifications representing detection signal alterations caused by radiation source movement through the radiation portal monitor to obtain pseudo-dynamic identification training data;

training the identification machine learning model using the pseudo-dynamic identification training data; and

storing the trained identification machine learning model.