Intelligent Low-level Signal Detection and Zero- Suppession in Raw LArTPC Waveforms through Deep Learning Techniques Lorenzo Uboldi Supervisor Michael Wang Final report 24 September 2019
DUNE Deep Underground Neutrino Experiment is a LArTPC detector that aims to study: ● Neutrinos from accelerator ● Rare events neutrinos (Supernovae burst, proton decay) The latter have intrinsic problems: ● lower energy, close to the limit of the detector ● to have higher chances to be detected need 100% live time Solutions: ● Save all data and analyse offline (throughput ~ TB/s) ● Online data discrimination Lorenzo Uboldi | Final report 2 24/09/19
Classical approach For low energy events: ● Too much data from DAQ ● Signal and noise are almost indistinguishable ● Zero suppression method ● Events with charge collection below a fixed threshold are discarded Very hard to detect low energy neutrinos (e.g.: from supernova burst) Lorenzo Uboldi | Final report 3 24/09/19
New approach Use machine learning. Benefits: ● We don’t have to study and develop a method to discrimante signal ● The machine will learn itself what a signal is and what is not ● Sometimes the machine is able to undestand underlying features and correlations that we are not Cons: ● We need labelled data (Monte Carlo must be accurate) Lorenzo Uboldi | Final report 4 24/09/19
What is a “classical” neural network? y ∈[ 0 , 1 ] 1 − w i x i 1 + e Input vector max ( 0 ,w i X i ) X to each neuron Lorenzo Uboldi | Final report 5 24/09/19
Advanced NN: Convolutional Neural Networks ● Convolutional NNs use convolution to take advantage of the space invariance and to extract features from data. ● Neurons are replaced by filters that are fitted to find features ● Each layer is made of many filters 2-dimensional CNN is the cutting edge technology for image recognition, object detection, self- driving cars, ecc... We use 1D Deep Waveforms are one Convolutional Neural dimensional Network Lorenzo Uboldi | Final report 6 24/09/19
Learning process Supervised learning is the process where labelled data is given to a neural network and its weights are fitted to give the most accurate predictions as possible Data Labels Predictions 0.02 0 Fit and 0.36 0 adjust weights 0.94 1 The training process requires a lot of data to be sensitive. I used ~400000 waveforms for training. Lorenzo Uboldi | Final report 7 24/09/19
Inference Trained Predictions Unseen data Network For testing a model the labelled data is split: ● 80% for training; ● 20% for testing predictions performances Lorenzo Uboldi | Final report 8 24/09/19
Results for monoenergetic 5 MeV neutrinos Fast, light and robust CNN has been developed 1 . 00 Training accuracy Test set Validation accuracy Accuracy 0 . 95 U Plane 0.955 0 . 90 Accuracy V Plane 0.976 0 . 85 Z Plane 0.974 0 . 80 0 . 75 0 2 4 6 8 Epochs Simulated waveforms have a minimum of 2000 electrons productions: Lower than actual zero suppression threshold Lorenzo Uboldi | Final report 9 24/09/19
Next step: realistic dataset Energy deposited Argon 3.0 Krypton ● Data generated from Radon CC 10,000 events 2.5 ● Minimum collection of 2000 electrons 2.0 ● Neutrinos from the A.U. Supernova “marley” 1.5 generator (Charge Current Neutrino) 1.0 ● Radiological background is a mixture of Radon 222, 0.5 Argon 39 and Krypton 85 0.0 0 1 2 3 4 5 MeV Lorenzo Uboldi | Final report 10 24/09/19
Dataset 222 Rn 778979 waveforms picked randomly from 10000 events: 14.3% ● 80% used for training and 85 Kr Noise validation 14.3% 42.9% ● 20% for testing 14.3% 14.3% 39 Ar CC Lorenzo Uboldi | Final report 11 24/09/19
Strategy ● First neural network To discriminate any signal from noise ● Second neural network To discriminate neutrinos from background Each NN has six convolutional layers developed to exploit the spatial invariance of the peak Neutrino Background Noise Lorenzo Uboldi | Final report 12 24/09/19
Results Second NN First NN 0.998 0.002 1.000 0.000 Radio Noise True label True label 0.125 0.875 0.058 0.942 Signal Neutrinos Noise Signal Radio Neutrinos Predicted label Predicted label False negative: False positive: ● Radiological = 95.6 % ● Radon = 26.1 % ● Neutrinos ● Argon = 0.4 % = 17.0 % ● Krypton = 56.8 % At the end 92.21 % of the neutrinos are saved Lorenzo Uboldi | Final report 13 24/09/19
Results Energy deposited 2500 Saved CC Lost CC 2000 Confused radio 1500 # 1000 500 0 0 1 2 3 4 5 6 MeV Saved CC 10 3 Lost CC Confused radio 10 2 # 10 1 10 0 0 1 2 3 4 5 6 MeV Lorenzo Uboldi | Final report 14 14 14 24/09/19
Peak finding, data preparation 15 10 5 440 windows large 0 100 bins, moving 10 5 10 bins at time 15 20 25 0 1000 2000 3000 4000 6 15 6 4 4 10 4 2 5 2 2 0 0 0 0 5 2 2 2 10 4 15 4 4 20 6 6 6 25 0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100 Out = 0 Out = 0 Out = 0 Out = 0 Inside = 1 Lorenzo Uboldi | Final report 15 15 15 24/09/19
Results Another CNN was trained on labelled “windows” data Accuracy of more than 99% 10 5 5 ADC ADC 0 0 5 5 10 0 1000 2000 3000 4000 0 1000 2000 3000 4000 1.0 0.6 0.5 0.8 Peak probability Peak probability 0.4 0.6 0.3 0.4 0.2 0.2 0.1 0.0 0.0 0 100 200 300 400 0 100 200 300 400 Red line is the position of the peak given by the Monte Carlo Lorenzo Uboldi | Final report 16 16 16 24/09/19
Last step: ICARUS data ICARUS is a LArTPC detector part of SBN at Fermilab. It aims to study neutrino oscillations, sterile neutrino, ecc... Problem: due to “hot” electronic the noise is very high, in particular middle induction panel has a really poor signal-noise ratio 7.5 5.0 2.5 0.0 ADC 2.5 5.0 7.5 10.0 0 1000 2000 3000 4000 Since ICARUS is a LArTPC detector as DUNE will be, why not to apply the developed method to it? Lorenzo Uboldi | Final report 17 17 17 24/09/19
CNN method applied to ICARUS ● Actual algorithm stores all waveforms and performs a heavy computing de-convolution to extract signal from noise. Must run offline and needs huge storage space and a lot of time. ● Performance for events with less than 15000 collected electrons drops below 75% of accuracy ● The CNN trained on ICARUS data keeps an accuracy between 99.2% and 99.8% for events with collected electrons as low as 2000 ● The NN for peak finding has an accuracy of 98%: the overall accuracy is higher than 97%! ● The inference speed is so high that there is no need to make it offline: a Nvidia Tesla K80 is actually faster than the detector throughput Lorenzo Uboldi | Final report 18 18 18 24/09/19
Further Developement For DUNE: ● Use a training set derived from a much bigger number of events in order to generalize better and being able to avoid sensitivity on statistical fluctuations ● Improve accuracy ● Design a specific hardware (FPGA or ASIC) for our CNN able to handle a huge throughput of data for live application (EdgeTPU by Google has been tested with no satisfying results) For ICARUS: ● Include our method in LarSoft and perform more accurate tests and comparisons ● If it turns out to be robust and stable, include hardware and software in the DAQ for the CNNs predictions Lorenzo Uboldi | Final report 19 24/09/19
Recommend
More recommend
