allfrey_htpd_2020_final

Abstract
In TAE Technologies’ current experimental device, C-2W (also called “Norman”)1, record breaking, advanced beam-driven field-reversed configuration (FRC) plasmas are produced and sustained in steady state utilizing variable energy neutral beams, expander divertors, end bias electrodes, and an active plasma control system. With a rapid shot-pace and an extensive number of data channels, the amount of data generated necessitates automated signal processing. To this end a machine learning algorithm, consisting of a multi-layered neural network as well as other data processing software, has been designed for signal feature identification, allowing for accurate and fast signal classification, anomalous condition detection, and providing for signal pre-processing. With a small set of training data the neural network can be “boot- strapped” to provide a robust classification system while minimizing human oversight. An example using data from the theta pinch plasma formation pulsed power system is described. However, this technique can be used for near-real-time preprocessing of any plasma diagnostic signal and has wide ranging application in fusion experiments for the varied data produced by plasma diagnostics.
Motivation/Formation Pulsed Power System
n The C-2W formation pulsed power2 consists of 190 capacitors, 280 switches and 380 Rogowski probes, with signals sampled at 25 MS/s.
n There are natural variations in the signals due to experimental settings, differences in mutual coupling between Formation coils, as well as the plasma formation process (see Fig. 2) which makes statistical analysis difficult. Instead a qualitative analytical method is necessary, motivating the use of a neural network for signal classification.
n The pulsed power system has occasional misfires, and though, the FRC is robust to these misfires, multiple misfires can decrease component lifetime.
n With the large number of signals, machine learning processes are optimal to identity the signal type according to a set of predefined classifications.
Neural Network Training
n 6774 human labeled signals are used to train the neural network, with a set of 11 classifications.
n The neural network code is written to allow for bootstrapping, where signals that have been automatically classified can be appended to the training data set if vetted by a human expert. This speeds up the training phase, as collecting the human labeled data is the bottleneck.
Post Shot Programs
n C-2W relies on programs that run after each experimental shot to process data.
n Signal classifications are stored in a database, along with calculated parameters, such as classification probability, the delay of the signal with respect to the trigger, etc.
Uses
n Acts as a pre-processor for further automated or human data analysis.
n Used to alert personnel to unusual system behavior which can be corrected.
Crowbar Delay Distribution
Neural Network
n Multi-layer Perceptron classifier from scikit-learn stochastic gradient descent for weight optimization.
n The neural network consists of four layers, with 626 neurons in the first layer, 500 in the second, 50 in the third and 11 in the output layer, corresponding to the 10 classifications. The number of neurons and layers are determined empirically.
n The neural network trains in a few minutes on a laptop. Good
Automated Signal Classification in the C-2W Experiment
Nathan Bolte, Ian Allfrey, Roberto Mendoza, and the TAE Team
TAE Technologies, Inc., 19631 Pauling, Foothill Ranch, CA 92610
3.0 2.5 2.0 1.5 1.0 0.5 0.0
Early Crowbar Good
Late Crowbar
Amplitude(A. U.)
0 2 4 6 8 10 Time(μs)
Late Crowbar Early crowbar
Late start Early Start
DAQ error No start
No crowbar Time offset Unlabeled error
Output layer (10)
Fig. 3 Using the neural network to classify the signals allows for statistical analysis on usable data. Without preprocessing the data, such analysis would be impossible.
Statistics
n More than 5 million pulsed-power signals have been automatically classified.
n There is a 0.3% misfire rate among the 280 switches, or less than one per shot.
n With the accuracy of the classification, as well as the near-real time processing, this implementation of automated classification can replace a full-time system expert.
Fig. 4 A detail of browser-based interaction with the automated classification system. Labels are available in near-real time, where improper classifications can be flagged by a user to incorporate poorly classified signals into the training data set for further improvement.
Input layer (626)
Hidden layers (500, 50)
n Identify nuanced features in signals. GOOD Mean (3489)
EARLY CROWBAR Mean (589)
Future work/Applications
n Feedback for inter-shot plant control, finetune machine configuration.
n Signal feature identification.
n Automated data pre-processing.
References
1. H. Gota et al., “Formation of hot, stable, long-lived field-reversed configuration plasmas on the c-2w device,” Nuclear Fusion 59, 112009 (2019)
2. I. Allfrey et al., (2019, June 22-28). Overview of the C-2W Formation Section Pulsed Power [Conference presentation]. PPPS 2019 Convention, Orlando, FL, United States
Visit the TAE Research Library at TAE.com Presented at HTPD 2020, December 15, 2020
Fig. 5 Graphic showing the classifications.
AA
LIso Rcharge
Lload
Fig. 1 Simplified circuit schematic showing a single formation coil with the location of the four Rogowski probes.
1.2 1.0 0.8 0.6 0.4 0.2 0.0
0.2 0.4
1.5
1.0
0.5
0.0
0.5
5 0 5 10 15 20 LATE CROWBAR Mean (332)
5 0 5 10 15 20
1.5
1.0
0.5
0.0
0.5
0.4
0.2
0.0
0.2 0.4
5 0 5 10 15 20 NO START Mean (186)
5 0 5 10 15 20
Rcharge
A
A
Fig. 2 Due to variations in the pulsed power circuits there is a wide range of signals within each classification, complicating the analysis. Classification allows for the rejection of unwanted data.