Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Increased warning times in JET APODIS disruption
predictor by using confidence qualifiers
J. Vega1, S. Dormido-Canto2, A. Murari3, G. A. Rattá1, R. Castro1 and JET Contributors*
EUROfusion Consortium, JET, Culham Science Centre, Abingdon, OX14 3DB, UK
1Laboratorio Nacional de Fusión, CIEMAT, Madrid, Spain 2Departamento de Informática y Automática, UNED, Madrid, Spain. 3Consorzio RFX, Padua, Italy.
*See the author list of “Overview of the JET results in support to ITER” by X. Litaudon et al. to be published in Nuclear Fusion Special issue: overview and summary reports from the 26th Fusion Energy Conference (Kyoto, Japan, 17-22 October 2016)
Acknowledgements
This work was partially funded by the Spanish
Ministry of Economy and Competitiveness under
the Projects No ENE2015-64914-C3-1-R and
ENE2015-64914-C3-2-R.
This work has been carried out within the framework of the EUROfusion Consortium and has received funding from the Euratom research and training programme 2014-2018 under grant agreement No 633053. The views and opinions expressed herein do not necessarily reflect those of the European Commission
Motivation
• Machine learning methods are used to learn something from past examples (model creation) in order to make predictions for new examples
• Machine learning methods have shown the potential to distinguish between disruptive and non-disruptive behaviours through the use of binary classifiers
• A training process with disruptive and non-disruptive examples allows determining a separation frontier between both plasma states to classify the disruptive/non-disruptive character of new examples
• Sometimes, the selection of disruptive examples is challenging
• When an alarm is triggered, only mitigation actions are possible
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
Disruption predictordecision function
non-disruptivebehaviour
disruptivebehaviour
For avoidance/mitigation purposes, three
plasma states have to be differentiated:
non-disruptive, abnormal and disruptive
Avoidance Mitigation
Motivation
• Machine learning methods could be used to find useful relationships between variables to discriminate among three plasma states: non-disruptive, abnormal and disruptive
• Abnormal predictions will trigger avoidance techniques
• Disruptive predictions will trigger mitigation techniques
• For machine protection purposes, in a first approach, the relationships
could not necessarily be related to physics quantities • The objective is recognizing the presence of dangerous behaviours for
the machine although the physics reasons are unknown
• From the machine learning perspective, an important goal would be to put into operation a single system able to decide between non-disruptive, abnormal and disruptive behaviours
• It should be emphasised that the predictor system has to be of general application (i.e. not focused on specific plasma events but able to identify any abnormal behaviour)
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
Disruption predictor
control loop
Non-disruptive
behaviour?
Completely disruptive
behaviour?
No
Yes
Yes
No
Avoidance
techniques
Mitigation
actions
The aim of this presentation is not the proposal of avoidance or mitigation methods but analysing
potential capabilities of machine learning to recognise abnormal plasma conditions and to trigger
alarms
Outline
• Disruption avoidance and mitigation
• Critical questions about machine learning and disruption avoidance
• A novel machine learning method to distinguish abnormal and disruptive behaviours
• Application to JET through the APODIS disruption predictor outputs
• Summary
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
Disruption avoidance and mitigation
• Existing techniques of avoidance/mitigation methods are • Injection of significant amount of gases through fast valves
• M. Lehnen et al. Nuclear Fusion 51 (2011) 123010 (12pp)
• M. Bakhtiari et al. Nuclear Fusion 51 (2011) 063007 (9pp)
• Killer pellets • G. Pautasso et al. Nuclear Fusion, 36, 10 (1996) 1291-1297
• N. Commaux et al. Nuclear Fusion 51 (2011) 103001 (9pp)
• ECRH injection • B. Esposito et al. Phys. Rev. Lett. 100, 045006 (2008)
• B. Esposito et al. Nuclear Fusion 51 (2011) 083051 (9pp)
• Depending on the disruption type and the time available between the alarm and the disruption, different strategies can be much more desirable than others
• Reliable disruption classifiers could be of big help • A. Murari et al Nuclear Fusion, 53 (2013) 033006 (9pp)
• B. Cannas et al Nuclear Fusion, 53, 9 (2013) 093023
• Disruption time predictors would be essential elements for proper selection of an avoidance/mitigation strategy
• G. Pautasso et al. Nuclear Fusion 42 (2002) 100-108
• F. C. Morabito et al. Nuclear Fusion 41, 11 (2001) 1715-1723
• B. Cannas et al. Nuclear Fusion 44 (2004) 68-76
• J. Vega et al. JET TFM May 8th, 2014
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
Disruption avoidance and mitigation • A pre-requisite to trigger avoidance/mitigation actions is the real-time identification of
abnormal/disruptive conditions through a corresponding alarm
• The reaction time can include • any computation time to select a specific A/M methodology
• the necessary time to fire technical systems
• the plasma response time to the A/M actions
• The warning time has to be greater than the reaction time • However, there are no reliable ‘disruption time predictors’ so far
• Therefore, in the presence of an alarm, the faster reaction the better
• Could machine learning help in making the decision about triggering alarms to start either avoidance or mitigation actions in a general context?
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
time
alarm avoidance/mitigation start
disruption
reaction time
warning time
Critical questions about machine learning and
disruption avoidance
• Drawback: machine learning techniques require training (the larger dataset the
better).
• This is a problem for ITER and DEMO
• There are alternatives to classical approaches of machine learning in disruption
prediction (valid for mitigation actions)
• Predictors from scratch: at least, one disruption is necessary to start the learning process. New
learning is added after each missed alarm
– S. Dormido-Canto et al. Nuclear Fusion 53 (2013) 113001 (8pp)
– J. Vega et al. Nuclear Fusion. 54 (2014) 123001 (17pp)
• Predictors based on anomaly detections: no data from past discharges are needed. Under test in
the JET real-time network
– J. Vega et al. 1st EPS Conference on Plasma Diagnostics. April 14-17, 2015. Frascati, Italy
– J. Vega et al. Proc. of the 26th Symposium on Fusion Engineering (SOFE 2015). May 31st-June 4th, 2015. Austin
(TX), USA
– S. Esquembri et al. Conference Record of the 20th IEEE Real-Time Conference. Jun 5th-10th, 2016. Padova, Italy
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
Critical questions about machine learning and
disruption avoidance • Can we qualify the prediction of avoidance actions in a running discharge?
• How can be false alarms discriminated?
• How sure are we to trigger an avoidance alarm and not a mitigation alarm?
• Some kind of measure is necessary (probability, index, error bars …)
• By assuming a reasonable level of confidence to trigger avoidance methods,
what specific actions have to be put into operation?
• Magnetic field, plasma current, injected power, gas fuelling, modifying pulse
schedule, changing operation scenario, …
• In contrast to many existing fusion devices, ITER pulses will primarily operate with a pulse
schedule managed by logic based on plasma conditions rather than being strictly driven by
a time schedule
• Additional predictors (machine learning, theoretical based or thresholds based) can
be necessary
• Type of abnormal condition, potential riskiness, technical system failure, human error, not
enough information, …
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
Machine learning and avoidance alarms
• Is really possible to develop a machine learning predictor to identify
abnormal plasma behaviours that have to be corrected to avoid
disruptions?
• Training is an issue: no reliable ways of classifying examples with
the label ‘abnormal behaviour’
• It does not seem easy the creation of a three-class classifier to
distinguish three different behaviours: safe, abnormal and
disruptive
• Could machine learning methods be used to find useful relationships
between variables to discriminate among three plasma states with only
two types of examples (disruptive/non-disruptive) in the training
process?
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
• Avoidance actions can only be carried out after the recognition of abnormal
behaviours
non-disruptive disruptive
abnormal
non-disruptive disruptive
Safe, abnormal, disruptive?
Conceptual view of classifiers and ‘abnormal’
examples
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
• Usually, training datasets are not linearly separable
• Low reliability predictions are ‘strange’ examples
• The ‘strangeness’ has to be quantified
• Low reliability predictions can be used to recognise ‘abnormal’ examples
• How?
• Is there any mathematical support for this assumption?
Classified as non-disruptive However, the closest examples are disruptive
Low reliability in the prediction
non-disruptive disruptive Separating hyper-plane
Classified as disruptive However, the closest examples are non-disruptive
Low reliability in the prediction
Training examples of class ‘non-disruptive’
Training examples of class ‘disruptive’
Examples are feature vectors: nx
Examples to classify
Conformal predictors
• Conformal predictors allow qualifying a prediction through a nonconformity measure • Intuitively, this is a way of measuring how different a new example is from
old examples
• Given a nonconformity measure and a bag of examples, the nonconformity score for each example in the bag is
• Because a nonconformity measure can be scaled, the numerical value , does not, by itself, tell how unusual finds to be. Therefore, a comparison of to the other is necessary
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
nA
1,...,
nz z
iz
1 1 1: ,..., , ,..., ,
i n i i n iA z z z z z
i
nA i
z
i j
Conformal predictors
• A convenient way of making the comparison is to compute the fraction
• The p-value is the fraction of the examples in the bag as nonconforming as
• It should be noted that
• If it is small (close to its lower bound 1/n for a large n), then is very nonconforming (an
outlier, i.e. very strange)
• If it is large (close to its upper bound 1), then is very conforming
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
# 1,..., :
:j i
i
j np value z
n
iz
1
1i
p value zn
iz
iz
Recipe to classify with a conformal predictor
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
1 1 1bag underlying classifier
nonconfor
Given a of examples , ..., , , and , ..., , an ,
a and a = , to classimi fyty m inteas o 1 of classes ( isur
unknown), th
e new ex
e confor
ample
n i i i i M
n n n
z z z y y L L
A z x y M y
x
1
1,
Provisionally set ,
1,
Set , ..., ,
# 1, ..., |Set the p-value:
:= lab
mal evaluation can be performed with the following algorithm:
k
n n k
i n n i
i n
y
FOR k M
z y
FOR i n
A z z z
END FOR i
i np
n
END FOR k
Label prediction
x
el of largest p-value
: largest p-value. : 1 - 2 p-valuend
Credibility Confidence
It is assumed that xn belongs to class k
The p-value is the fraction of nonconformity scores that are equal or greater than n
Nonconformity scores
Conformal prediction applied to recognise non-disruptive, abnormal and
disruptive behaviours from disruptive/non-disruptive training examples
1 1bag
underlying classifier
Given a of training examples , ..., , , and , ,
an , noncormity measure new exampla and a = , to classify into 1
of 3 classes
e
n i i i i
n n n
non disrup disruptivez z z y y tive
no
A z y
n
x
x
1
1, 2
Provisionally set ,
1,
Set , ..., ,
# 1, ..., |Set the p-value:
, , , the following evaluation can be carried out:
k
n n k
i n n i
i n
y
FOR k
z y
FOR i n
A z z z
END FOR i
i np
n
END FOR k
abnorm disrupdisrupti
Label pr
v tivee
e
al
x
1 ,
,
,
disruptive
disruptive
disruptive
non disruptive
non disruptive
non disruptive
n
abnif p pn
dict on di
disruptiv
ion sruptiif p
ormal
p p
vep
if e
The nonconformity scores can be seen as the ‘prediction reliabilities’
Nonconformity scores
,
,
n n
n n
non disruptivez
z disruptive
x
x
11
11
disruptive
non disruptivep
n
pn
Preliminary results in JET
• Bag of examples
• Training dataset of the APODIS 2nd layer classifier
• Underlying classifier
• Support Vector Machine classifier
• Nonconformity measure
• Samples to classify
• Outputs of APODIS 1st layer ( ) every 32 ms (from plasma start to extinction)
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
3
n x
distance to the separating hyperplane well classified
distance to the separating hyperplane bad classifiedn
ifA
if
APODIS review
• Input signals: plasma current, locked mode amplitude, total input power, plasma internal inductance, plasma density, stored diamagnetic energy time derivative, and radiated power
• As a discharge is in execution, the three most recent 32 ms temporal segments are classified as disruptive or non-disruptive
• The three models may disagree about the discharge behaviour 2nd layer
t t - 32 t - 64 t - 96
M1 M2 M3
t + 32 t t - 32 t - 64 t - 96
M1 M2 M3
t + 64 t + 32 t t - 32 t - 64 t - 96
M1 M2 M3
t + 96 t + 64 t + 32 t t - 32 t - 64 t - 96
M1 M2 M3
The classifiers operate in parallel on consecutive time windows
PREDICTOR
First
layer
Second
layer
Decision Function: SVM classifier
[-64, -32] [-96, -64] [-128, -96]
M1 (SVM)
M2 (SVM)
M3 (SVM) M1, M2 and M3 are 3 independent models
Train temporal segments (ms) w.r.t. disruption
J. Vega et al. Fus. Eng. Des. 88 (2013) 1228-1231
Preliminary results in JET
• Bag of examples
• Training dataset of the APODIS 2nd layer classifier
• Underlying classifier
• Support Vector Machine classifier
• Nonconformity measure
• Samples to classify
• Outputs of APODIS 1st layer ( ) every 32 ms (from plasma start to extinction)
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
3
n x
distance to the separating hyperplane well classified
distance to the separating hyperplane bad classifiedn
ifA
if
Support Vector Machines
• SVM gives the distance (with sign) of the examples to the separating hyper-plane
• d1= +1.5
• d2= - 2.1
• The label of an example is determined through the sign(distance to the separating hyper-plane)
• Sample 1 belongs to class {+1}: sign(d1) > 0
• Sample 2 belongs to class {-1}: sign(d1) < 0
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
Class {+1} Class {-1} Separating hyper-plane
1 2 d1
d2
Preliminary results in JET
• Bag of examples
• Training dataset of the APODIS 2nd layer classifier
• Underlying classifier
• Support Vector Machine classifier
• Nonconformity measure
• Samples to classify
• Outputs of APODIS 1st layer ( ) every 32 ms (from plasma start to extinction)
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
3
n x
distance to the separating hyperplane well classified
distance to the separating hyperplane bad classifiedn
ifA
if
Preliminary results in JET
• Bag of examples • Training dataset of the APODIS 2nd layer classifier
• Underlying classifier • Support Vector Machine classifier
• Nonconformity measure
• Examples to classify • Outputs of the APODIS 1st layer ( ) every 32 ms (from plasma start to extinction)
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
3
nx
distance to the separating hyperplane well classified
distance to the separating hyperplane bad classifiedn
if
ifA
Avoidance/mitigation predictor
M1 (SVM)
M2 (SVM)
M3 (SVM)
and non dis disruptivp v eru ti ep t p t : prediction reliabilities
Predictions
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
1 ,
,
,
disruptive
disruptive
disruptive
non disruptive
non disruptive
non disruptive
non disrupt
abnorma
disruptive
if p pn
Label prediction if p p
if p p
e
l
iv
The temporal evolution of the ‘prediction reliabilities’ is analysed
, non disrupti disruptivevet p p
The first time
a mitigation alarm is triggered
non disrupdisrupti vee tivp p
Two consecutive
predictions in which
1
an avoidance alarm is issued
disruptive non disruptivep pn
Very preliminary results with JET discharges
• This avoidance/mitigation predictor (AMP) has been
applied to 789 JET discharges in the range 82429 –
83793
• 81 unintentional disruptions successfully predicted by
APODIS
• 708 non-disruptive discharges
• All disruptions are predicted
• 94% with more than 10 ms of warning time
• 37% of the alarms correspond to avoidance alarms
• Average(warning time) ± standard deviation
• APODIS: 428 ± 1166 ms
• AMP: 606 ± 1874 ms
• False alarms rates
• APODIS: 0.99%
• AMP: 5.08%
• For APODIS warning times between 150 ms and 2 s,
on average, the AMP recognises avoidance alarms 66
ms earlier than APODIS (only mitigation)
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.
Summary
• Machine learning tools can be used to recognise in a single predictor the need of triggering either avoidance or mitigation alarms
• The discrimination between avoidance and mitigation alarms is crucial
• The development of a three class predictor with only two types of training examples has been accomplished through the theory of conformal predictors
• Only disruptive and non-disruptive examples are necessary
• Preliminary application to JET data to distinguish between avoidance or mitigation is promising • More than 1/3 of the alarms are avoidance ones
• The triggering of specific avoidance or mitigation techniques requires additional inputs to decide
• The application provided is an initial approach to be analysed in more detail and not depending on APODIS at all • Other types of confidence classifiers (for example, probabilistic predictors) have to be developed
• Machine learning methods can be used in a variety of approaches for combining avoidance and mitigation predictions • The essential point is to find proper features to differentiate possible states
• Classical approaches (large datasets for training are required)
• From scratch approaches
• Anomaly detection approaches
2nd IAEA TM 2017. MIT PSFC. Cambridge, MA, USA.