data analytics & machine learning€¦ · introduction • data analysis is key for bi systems...

22

Upload: others

Post on 22-Aug-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques
Page 2: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Data Analytics & Machine Learning

in BI

M. Gonzalez Berges, J.J. Gras, B. Salvachua

With input from

D. Alves, G. Azzopardi, E. Bravin, L. Coyle, M. Di Castro, L. Grech, A. Guerrero,

R. Jones, T. Levens, T. Pieloni, G. Valentino, M. Wendt, C. Zamantzas

Page 3: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Agenda

• Introduction

• OAF (Offline Analysis Framework)

• Current Use Cases

• Improve diagnostics with ML/DA

• BI wishes

• Conclusions

28th May 2019 – ML Workshop

Page 4: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Introduction

• Data Analysis is key for BI systems

• Going from instrument measurement to beam

parameters

• Machine Learning techniques

• Not much done so far compared with other

standard signal treatments

• Big interests in evaluating potential

28th May 2019 – ML Workshop BE-BI 4

Evaluate instrument status (predictive maintenance)

Asses instrument performance (aging)

Improve instrument response (calibration).

Page 5: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Offline Analysis Framework (OAF)

BI centralized tool that provides:

• Automatic daily reports based on analysis of

logging data

• About 50 reports generated per day

• Processing based on data set configuration files

(extension with python code is possible,

currently 5% of cases)

• CALS + Python

• Relies on “snapshots”, this functionality should

be kept in the new API with NXCALS

528th May 2019 – ML Workshop BE-BI

Page 6: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Offline Analysis Framework (OAF)

Example: BLM card temperature measurements, with statistical analysis to identify trends, outliers, etc.

6

Daily report (24 h of data), average and sigma distributions

28th May 2019 – ML Workshop BE-BI

Page 7: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

OAF Analysis (some examples)

728th May 2019 – ML Workshop BE-BI

Accelerator Instrument - Analysis

LINAC 4 Transmission efficiency / Line BCTs cross-calibration check

BLM crate humidity monitoring

LINAC 3 Transmission efficiency / Line BCTs cross-calibration check

PSB Overview of PSB beam Instrumentation

Wire scanner usage survey and analyze

PS Overview of PS beam Instrumentation

BLM: Comparison of the old and newly installed electronics results

SPS Monitoring of the BCT used for safety for the EA

BPM: MOPOS vs ALPS – evaluation of the new orbit system

Wire scanner usage survey and analyses

LHC BPM – Electronics Racks and acq card Temperature Survey

BLM – Acq Cards temp survey

DCCT BCT cross calibration check

Wire scanner usage survey and analyze

AD,LEIR… …

Page 8: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Other Analysis Tools

BLMs health system checks• Additional daily cron reports: connectivity-dac,

optical link errors, LSA BLM threshold changes, voltages status.

• CALS + LSA DB + Python and post processing

8

Reports are

produced with

summarized

information

28th May 2019 – ML Workshop BE-BI

Page 9: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Other Analysis ToolsExample: dedicated analysis tools for specific tasks such as BSRT calibration, BLM lifetime calibration and fill-by-fill monitoring, dBLM fill-by-fill analysis, etc.

928th May 2019 – ML Workshop BE-BI

Page 10: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Current ML use cases

• BLM radiation tests with TIM.

• Renovation of the LHC beam-based

feedback systems.

• LHC BLM spike classification applied to the

collimation alignment.

• LHC beam lifetime optimization at injection.

1028th May 2019 – ML Workshop BE-BI

Page 11: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Current ML use casesBLM radiation tests with TIM

• Radiation BLM tests done autonomously with the TIM train

• Faster-RCNN network for online 2D Beam Loss Monitors (BLM) localization

• Multiple RGB-D cameras used for 3D reconstruction of the environment

• 3D pose will be used by the robotic arm path planner to calculate a safe approach to the BLM in the reconstructed environment

• Image recognition

11

Collaboration with EN-SMM (M. Di Castro)

28th May 2019 – ML Workshop BE-BI

Page 12: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Current ML use cases

Renovation of the LHC beam-based feedback systems

• The LHC OFC is currently being upgraded to FESA3.

• As part of an academic exercise, we are investigating the use of Reinforcement Learning for orbit feedback control as opposed to the SVD beam response matrix.

• The objective is to respond more quickly to BPM or COD failures, and achieve equal, if not better performance in the orbit feedback.

• A simulation environment is being set up using OpenAI Gym.

• Anomaly detection of BPMs used for the orbit feedback is also being investigated using machine learning techniques such as Local Outlier Factor

12

Collaboration with OP-LHC and U. Malta (G. Valentino)

28th May 2019 – ML Workshop BE-BI

Page 13: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Current ML use cases

• LHC BLM spike classification applied to the

collimation alignment.

• Study the prediction of the LHC beam

lifetime at injection and the optimization of

the tune working point using ML algorithms.

13

Collaboration OP-LHC and LHC Collimation

Collaboration OP-LHC and EPFL

Gabriella Azzopardi will present on the 4th June

Loic Coyle presented today

28th May 2019 – ML Workshop BE-BI

Page 14: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Improve of beam diagnostics

using ML/DA

14

Keeping the same goals:

Evaluate instrument status (predictive maintenance)

Asses instrument performance (aging)

Improve instrument response (calibration).

We have discussed within BI how beam diagnostics could

be improved if applying more sophisticated techniques.

We came out with a list of subjects where improving ML/DA

could have a direct impact

28th May 2019 – ML Workshop BE-BI

Page 15: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Improve of beam diagnostics

using ML/DA

15

Identify outliers in different instruments or measurements. Currently done

in most cases by comparison of simple thresholds (like BPM

temperatures):

• Extend the work started in ABP for identifying misbehaviors of BPMs.

• The next generation of acquisition cards are equipped with Ethernet

connection and higher computation power (FPGA) could envisage NN

algorithm to detect anomalies (example of BLM patterns).

• Study Wire-scanner distributions of power, position and profile.

Disentangle real beam effects vs instrumental problems.

Find the correct tune in a noisy spectrum:

• Feedback the tune finder with information on noise peaks.

Head-tail triggering too often with TBytes of data.

• Identify the type of instability like a 2nd level trigger and reduce the

data stored.

28th May 2019 – ML Workshop BE-BI

Page 16: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Improve of beam diagnostics

using ML/DA

16

Cross-calibration or online recalibration of instruments.

• BSRT-WS, explore analysis of images.• BLM lifetime-BCT / losses in IRs vs losses in other locations, improve

pattern recognition.

Development of direct e-cloud measurements/observation using BPMs.• Complex, requiring correlation with other data like cryogenics, bunch-by-

bunch patterns on beam size and charges.

Virtual instruments combining signal from different devices: schottky,

lifetime, luminosity prediction.

Asses performance of instruments or algorithms, example OFC by analysis

BPM signal and COD current, trained with fills data.

Beam size measurements using quadrupolar moment of BPMs.

28th May 2019 – ML Workshop BE-BI

Page 17: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Use Case AnalysisUse Case Data Source Impact

Identify Outliers for BLMs,

BPMs, Wire-scanners

CALS (NXCALS) Improve instrument

availability/performance

Tune measurement CALS (NXCALS) Improve tune signal

Head-tail triggering Files (TBytes)

CALS (NXCALS)

Reduce data volumen /

better analysis

Cross-calibration / online

calibration

CALS (NXCALS)

Images needed

Better measurements

BPMs quadrupolar moment CALS (NXCALS)

Online

Additional beam size

measurement

ecloud measurement with

BPMs

TBD Direct ecloud

measurement

Virtual instruments Several Additional mesurements

Algorithms assesment CALS (NXCALS) Performance monitoring

1728th May 2019 – ML Workshop BE-BI

Page 18: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

BI wishes

• Ideally to be provided centrally

18

• Centralized software

• Because python is intuitive many have started

here:

• Support of main data analysis libraries (numpy,

scipy, pandas, matplotlib, etc.)

• Support a (or several) machine learning

packages (scikit-Learn, pytorch, tensorflow,

keras)

28th May 2019 – ML Workshop BE-BI

Page 19: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

BI wishes

• Support on data collection and preparation:

• Data preparation: align time data, clean data,

etc.

• Logging Flexibility: ML relies in many cases on

the analysis of “big data” samples. Flexibility on

increasing the logging rate for certain periods,

like MD or commissioning is desirable. Example:

G.Azzopardi: training of BLM spike using dedicated 100Hz

BLM stream data, stored in csv files. This was crucial in order

to be able to measure the shape of the signal. Similar cases

might apply to UFO studies

1928th May 2019 – ML Workshop BE-BI

Page 20: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

BI wishes

• BI will heavily rely on NXCALS

• Guidelines on performance for data

insertion/extraction avoiding custom setups

• Currently files used in some cases, some ad-hoc

infrastructure (servers + net links)

• Backwards compatibility API to keep our

tools running

• Evaluation of online analysis

2028th May 2019 – ML Workshop BE-BI

Page 21: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

Conclusion

• Data Analysis is part of BI core activities

• ML has only been started

• Rely as much as possible in NXCALS

provided features

2128th May 2019 – ML Workshop BE-BI

Page 22: Data Analytics & Machine Learning€¦ · Introduction • Data Analysis is key for BI systems • Going from instrument measurement to beam parameters • Machine Learning techniques

22

Workshop on

“Data Science and Machine Learning”

• Sunday 6th of October• Morning: tutorials

• Afternoon: presentations / demonstrations

• Full Details• https://icalepcs2019.bnl.gov/workshops.html#11

Contributions

are welcome!

28th May 2019 – ML Workshop BE-BI