Intelligent Low-level Signal Detection and Zero- Suppession in Raw - - PowerPoint PPT Presentation

▶
intelligent low level signal detection and zero
SMART_READER_LITE
LIVE PREVIEW

Intelligent Low-level Signal Detection and Zero- Suppession in Raw - - PowerPoint PPT Presentation

Feb 26, 2024 293 likes •494 views

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

slide-1
SLIDE 1

Lorenzo Uboldi Supervisor Michael Wang Final report 24 September 2019

Intelligent Low-level Signal Detection and Zero- Suppession in Raw LArTPC Waveforms through Deep Learning Techniques

slide-2
SLIDE 2

2

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

24/09/19 Lorenzo Uboldi | Final report

slide-3
SLIDE 3

For low energy events:

  • Too much data from DAQ
  • Signal and noise are almost indistinguishable

3

  • Zero suppression method
  • Events with charge collection below a fixed threshold are

discarded

Classical approach

Very hard to detect low energy neutrinos (e.g.: from supernova burst)

Lorenzo Uboldi | Final report 24/09/19

slide-4
SLIDE 4

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)

4

New approach

Lorenzo Uboldi | Final report 24/09/19

slide-5
SLIDE 5

5

What is a “classical” neural network?

y∈[0,1] X max(0,wi Xi)

Input vector to each neuron

1 1+e

−wi xi

Lorenzo Uboldi | Final report 24/09/19

slide-6
SLIDE 6

6

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... Waveforms are one dimensional We use 1D Deep Convolutional Neural Network

Lorenzo Uboldi | Final report 24/09/19

slide-7
SLIDE 7

7

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 1 Labels Fit and adjust weights Predictions 0.02 0.36 0.94 The training process requires a lot of data to be sensitive. I used ~400000 waveforms for training.

Lorenzo Uboldi | Final report 24/09/19

slide-8
SLIDE 8

8

Unseen data

Trained Network Predictions For testing a model the labelled data is split:

  • 80% for training;
  • 20% for testing predictions performances

Inference

Lorenzo Uboldi | Final report 24/09/19

slide-9
SLIDE 9

9

Results for monoenergetic 5 MeV neutrinos

2 4 6 8

Epochs

0.75 0.80 0.85 0.90 0.95 1.00

Accuracy

Training accuracy Validation accuracy

Test set Accuracy U Plane 0.955 V Plane 0.976 Z Plane 0.974

Lorenzo Uboldi | Final report

Simulated waveforms have a minimum of 2000 electrons productions: Lower than actual zero suppression threshold

24/09/19

Fast, light and robust CNN has been developed

slide-10
SLIDE 10

10

1 2 3 4 5 MeV 0.0 0.5 1.0 1.5 2.0 2.5 3.0 A.U.

Energy deposited

Argon Krypton Radon CC

Next step: realistic dataset

  • Data generated from

10,000 events

  • Minimum collection of 2000

electrons

  • Neutrinos from the

Supernova “marley” generator (Charge Current Neutrino)

  • Radiological background is

a mixture of Radon 222, Argon 39 and Krypton 85

Lorenzo Uboldi | Final report 24/09/19

slide-11
SLIDE 11

11

Dataset

222Rn

14.3%

85Kr

14.3%

39Ar

14.3% CC 14.3% Noise 42.9% 778979 waveforms picked randomly from 10000 events:

  • 80% used for training and

validation

  • 20% for testing

Lorenzo Uboldi | Final report 24/09/19

slide-12
SLIDE 12

12

Strategy

  • First neural network
  • Second neural network

To discriminate any signal from noise To discriminate neutrinos from background Noise Neutrino Background Each NN has six convolutional layers developed to exploit the spatial invariance of the peak

Lorenzo Uboldi | Final report 24/09/19

slide-13
SLIDE 13

False negative:

  • Radiological = 95.6 %
  • Neutrinos

= 0.4 %

Radio Neutrinos Predicted label Radio Neutrinos True label 0.998 0.002 0.058 0.942

Second NN

Noise Signal Predicted label Noise Signal True label 1.000 0.000 0.125 0.875

First NN

Results

False positive:

  • Radon

= 26.1 %

  • Argon

= 17.0 %

  • Krypton = 56.8 %

At the end 92.21 % of the neutrinos are saved

13 Lorenzo Uboldi | Final report 24/09/19

slide-14
SLIDE 14

14 14 14

1 2 3 4 5 6 MeV 500 1000 1500 2000 2500 #

Energy deposited

Saved CC Lost CC Confused radio 1 2 3 4 5 6 MeV 100 101 102 103 # Saved CC Lost CC Confused radio

Results

Lorenzo Uboldi | Final report 24/09/19

slide-15
SLIDE 15

Peak finding, data preparation

15 15 15

1000 2000 3000 4000 25 20 15 10 5 5 10 15

20 40 60 80 100 6 4 2 2 4 20 40 60 80 100 6 4 2 2 4 6 20 40 60 80 100 6 4 2 2 4 6 20 40 60 80 100 25 20 15 10 5 5 10 15

440 windows large 100 bins, moving 10 bins at time Out = 0 Out = 0 Out = 0 Out = 0 Inside = 1

Lorenzo Uboldi | Final report 24/09/19

slide-16
SLIDE 16

Results

16 16 16

1000 2000 3000 4000 10 5 5 10 ADC 1000 2000 3000 4000 5 5 ADC 100 200 300 400 0.0 0.2 0.4 0.6 0.8 1.0 Peak probability 100 200 300 400 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Peak probability

Another CNN was trained on labelled “windows” data Accuracy of more than 99% Red line is the position of the peak given by the Monte Carlo

Lorenzo Uboldi | Final report 24/09/19

slide-17
SLIDE 17

Last step: ICARUS data

17 17 17

1000 2000 3000 4000 10.0 7.5 5.0 2.5 0.0 2.5 5.0 7.5 ADC

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

Lorenzo Uboldi | Final report 24/09/19

Since ICARUS is a LArTPC detector as DUNE will be, why not to apply the developed method to it?

slide-18
SLIDE 18

CNN method applied to ICARUS

18 18 18 Lorenzo Uboldi | Final report 24/09/19

  • 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

slide-19
SLIDE 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

19 Lorenzo Uboldi | Final report 24/09/19