Application of artificial neural networks in simulating radon levels in soil gas

Abstract

Anomalies have been observed in radon content in soil gas from three boreholes at the Orlica fault in the Krško basin, Slovenia. To distinguish the anomalies caused by environmental parameters (air and soil temperature, barometric and soil air pressure, rainfall) from those resulting solely from seismic activity, the following approaches have been used. First, the seismic activity data were eliminated from the dataset and then an artificial neural network (ANN) with 5 inputs for environmental parameters and a single output (radon concentration) was trained with the standard backpropagation learning rule. Then the predictions of Rn concentrations ( C p) generated with this ANN for the whole dataset were compared to measurements ( C m) and three types of anomalies (CA — correct anomaly, FA — false anomaly and NA — no anomaly) have been detected in the signal | C m/ C p − 1| by varying five parameters describing an anomaly within predefined intervals. An exhaustive search among results was made to find the best ones and thus identifying the best set of parameters. Finally, an attempt was made to shorten the search procedure by training another ANN with numbers of anomalies of each type in the input and five anomaly detection parameters in the output. With these procedures we were able to correctly predict 10 seismic events out of 13 within the 2-year period.