You are on page 1of 3

Application: Mechanical Engineering

Predicting thermal runaway in boilers using ML

Every year, around 28000 deaths occur worldwide due to boiler explosions which in turn happen due
to the failure of the equipment due to a phenomenon called thermal runaway [1]. Inside the boiler,
the water initially undergoes nucleate pool boiling which means boiling with small bubble formation,
and as the temperature increases, the bubble sizes increase. A point in time comes when the boiling
becomes vigorous, with the highest heat transfer coefficient and the largest (critical) heat flux (CHF)
possible. This is known as boiling crisis. Just after this, suddenly, the whole heater surface becomes
dry and gets covered by the bubble. Since the mode of heat transfer shifts from convection to
conduction, and that too, through steam, the heat transfer becomes less effective and the
temperature of the equipment rises many orders of magnitudes in very less time. This phenomenon,
named thermal runaway, causes thermal failure of the equipment.

Fig 1: Nukiyama Pool Boiling Curve showing the different regimes of pool boiling for water with
bubble formation illustrations on a Heat Flux vs Wall Superheat (temperature difference between
the heater surface and the boiling point of the fluid) (log-log) plot [2].

One of the methods to mitigate the problem is to utilize the acoustics of bubble formation and
bursting [3]. The acoustic data has a clear change in the frequency of sound that has the peak
amplitude when the nature of the boiling changes.

To make the audio data compatible with ML algorithms, the first statistical procedure required is to
change the medium of conveyance of the information from audio to visual, in the form of
spectrograms or heat maps. Then, the ML algorithms such as Convolutional Neural Network (CNN)
or Recurrent Neural Network (RNN) can be used for image processing. At the fundamental level, the
experiments that have been conducted will need to be segregated into a training set and an
evaluation set for testing. Since there will be data from many boiling regimes, for example: nucleate
boiling regime, boiling crisis regime, and the regime of film boiling after the thermal runaway,
random sampling from the data from each regime is generally performed to make the evaluation
set, and the rest serves as the training set. The training set also is enriched by the data of the
experiments already conducted by others.

After the training and the testing process, the statistical metrics that signify the correctness of the
model are: Accuracy, Precision, Recall, and F1 Score. These are the metrics for the validation of the
model. Accuracy can be understood as how many of the cases where it was predicted positive, i.e.
thermal runaway was predicted, were actually data points of thermal runaway, while precision can
be thought about how consistent the predictions of the model were in the different regimes. The
CNN model scored high accuracy rates of 99%+ [4].

Recall, on the other hand, is a measure of how many of the real positive cases i.e. thermal runaway
really happening, were predicted as positive [5].

True Positives
Precision=
True Positive s+ False Positives
True Positives
Recall=
True Positives+ False Negatives
In a general case, we would try to minimize both False Positives as well as False Negatives, but in
reality, increasing precision generally leads to a decrease in recall. Therefore, the better metric for
comparison is F1 Score.

2
F 1 Score=
1 1
+
Precision Recall
Being the harmonic mean of Precision and Recall, the F1 score (ranging between 0 and 1) is lower for
cases where precision and recall have a large difference between them. The best possible model
would be one with F1 Score of 1, i.e., the closer the score is to 1, the better is the model.

References:

[1]: https://www.thestatesman.com/india/indian-scientist-working-technology-predict-control-
boiler-explosion-cause-28000-deaths-worldwide-last-one-decade-1503036084.html as accessed on
31 August 2022

[2]: https://www.nuclear-power.com/nuclear-engineering/heat-transfer/boiling-and-
condensation/pool-boiling-boiling-curve/ as accessed on 31 August 2022

[3]: Kumar Nishant Ranjan Sinha, Vijay Kumar, Nirbhay Kumar, Atul Thakur, Rishi Raj; “Deep learning
the sound of boiling for advance prediction of boiling crisis”; Cell Reports Physical Science 2021;
https://doi.org/10.1016/j.xcrp.2021.100382

[4]: Ibid

[5]: https://towardsdatascience.com/essential-things-you-need-to-know-about-f1-score-
dbd973bf1a3 as accessed on 31 August 2022
[6]: Ibid

You might also like