Explosion Detection Using Smartphones: Ensemble Learning with the Smartphone High-Explosive Audio Recordings Dataset and the ESC-50 Dataset

Abstract

Explosion monitoring is performed by infrasound and seismoacoustic sensor networks that are distributed globally, regionally, and locally. However, these networks are unevenly and sparsely distributed, especially at the local scale, as maintaining and deploying networks is costly. With increasing interest in smaller-yield explosions, the need for more dense networks has increased. To address this issue, we propose using smartphone sensors for explosion detection as they are cost-effective and easy to deploy. Although there are studies using smartphone sensors for explosion detection, the field is still in its infancy and new technologies need to be developed. We applied a machine learning model for explosion detection using smartphone microphones. The data used were from the Smartphone High-explosive Audio Recordings Dataset (SHAReD), a collection of 326 waveforms from 70 high-explosive (HE) events recorded on smartphones, and the ESC-50 dataset, a benchmarking dataset commonly used for environmental sound classification. Two machine learning models were trained and combined into an ensemble model for explosion detection. The resulting ensemble model classified audio signals as either "explosion", "ambient", or "other" with true positive rates (recall) greater than 96% for all three categories.

Keywords: data; detection; explosion; infrasound; machine learning; smartphone.

Figure 1

The architecture of YAMNet and an example architecture of a transfer learning model using YAMNet.

Figure 2

The smartphone deployment configuration of (a) the vented encasement and (b) the aluminum foil tube.

Figure 3

Histograms of (a) the smartphone’s distance from the explosion source, (b) the effective yield of the explosion source, (c) the number of smartphone recordings per explosion event, and (d) the smartphone model used for data collection in SHAReD.

Figure 4

Flowchart for the ensemble model along with the construction of D-YAMNet and LFM models.

Figure 5

The confusion matrix of D-YAMNet on the test dataset. The percentages are calculated by rows (true labels) and the count for each cell is listed under these in parenthesis.

Figure 6

The (a) normalized amplitude and (b) power spectral density of an “explosion” waveform that was misclassified as “ambient” by D-YAMNet. The explosion was in the 10 kg yield category and recorded by a smartphone ~11 km away from the source at a sample rate of 800 Hz.

Figure 7

The confusion matrix of the LFM on the test dataset. The percentages are calculated by rows (true labels) and the count for each cell is listed under these in parenthesis.

Figure 8

The (a) normalized amplitude and the (b) power spectral density of an “other” waveform that was misclassified as “explosion” by the LFM. The “other” sound was from an ESC-50 waveform labeled “dog”.

Figure 9

The confusion matrix of the ensemble model on the test dataset. The percentages are calculated by rows (true labels) and the count for each cell is listed under these in parenthesis.

Figure 10

The precision-recall curves for (a) D-YAMNet and (b) LFM.

Figure 11

The average confusion matrix for (a) D-YAMNet and (b) LFM.

Figure 12

Expanded normalized microphone data from SHAReD for event INL_20220714_07 for smartphone ID (a) 1806169311 and (c) 2122963039 and the corresponding predicted labels from D-YAMNet, LFM, and the ensemble model for smartphone ID (b) 1806169311 and (d) 2122963039. The predicted labels were obtained on segmented section of the full waveform with 0.96 s duration and 50% overlap.

Figure 13

The (a) confusion matrix and (b) precision-recall curves for the YAMNet model. The average precision (AP) for VINEDA was 0.86.

Figure 14

The (a) confusion matrix and (b) precision-recall curves for the YAMNet model.