Real-time (RT) disruption prediction (DP) is essential in fusion to avoid irreversible damage to the devices. Nowadays, physics models for DP are under development and, usually, machine learning methods (with success rate above 90%) are used [1]. However, large training datasets are required but the generation of large databases for DP will not be possible in ITER/DEMO. To overcome this issue, the approach of prediction from scratch can be used [2, 3]: adaptive predictors are built from only 1 disruptive and 1 non-disruptive discharge and they are re-trained after each missed alarm. This paper deals with disruption mitigation and performs the RT implementation of the Venn predictor developed in [3], which provides a probability error bar for each prediction. Results with three ITER-like wall campaigns of JET (1036 non-disruptive and 201 disruptive discharges) show that only 12 re-trainings are necessary, the success rate is 94% and the false alarm rate is 4%. The RT implementation has been carried out with a fast controller with DAQ FPGA-based data acquisition devices corresponding to ITER catalogue (in particular, a reconfigurable Input/Output platform has been used). Three input signals (sampled at 1 ms) have been used: plasma current, locked mode and internal inductance. These signals are read from JET database. Then D/A conversions are carried out and used as inputs to the data acquisition card. In this way, the whole process of digitization, data analysis and prediction is performed. Predictions are made by using the mean values and the standard deviations of the Fourier spectrum (removing the DC component) of the latest 32 samples of the above signals every 2 ms (prediction period). The Venn predictor uses the nearest centroid taxonomy. The computation time for each prediction takes hundreds of us.
Real-time implementation with FPGA-based DAQ system of a probabilistic disruption predictor from scratch
Murari A;
2017
Abstract
Real-time (RT) disruption prediction (DP) is essential in fusion to avoid irreversible damage to the devices. Nowadays, physics models for DP are under development and, usually, machine learning methods (with success rate above 90%) are used [1]. However, large training datasets are required but the generation of large databases for DP will not be possible in ITER/DEMO. To overcome this issue, the approach of prediction from scratch can be used [2, 3]: adaptive predictors are built from only 1 disruptive and 1 non-disruptive discharge and they are re-trained after each missed alarm. This paper deals with disruption mitigation and performs the RT implementation of the Venn predictor developed in [3], which provides a probability error bar for each prediction. Results with three ITER-like wall campaigns of JET (1036 non-disruptive and 201 disruptive discharges) show that only 12 re-trainings are necessary, the success rate is 94% and the false alarm rate is 4%. The RT implementation has been carried out with a fast controller with DAQ FPGA-based data acquisition devices corresponding to ITER catalogue (in particular, a reconfigurable Input/Output platform has been used). Three input signals (sampled at 1 ms) have been used: plasma current, locked mode and internal inductance. These signals are read from JET database. Then D/A conversions are carried out and used as inputs to the data acquisition card. In this way, the whole process of digitization, data analysis and prediction is performed. Predictions are made by using the mean values and the standard deviations of the Fourier spectrum (removing the DC component) of the latest 32 samples of the above signals every 2 ms (prediction period). The Venn predictor uses the nearest centroid taxonomy. The computation time for each prediction takes hundreds of us.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
