One of the most prominent applications of fiber optic Distributed Acoustic Sensing (DAS) is Perimeter Security via fence monitoring, which is possible when we attach a fiber to the fence. In this study, we aim to detect and classify events occurring around said fence, such as climbing, cutting, and bending. For this, we investigate Deep Learning algorithms, more specifically Convolutional Neural Networks (CNN), as a mean to detect anomalies and classify them. We recorded 48,445 samples of the mentioned events, which were carefully processed and labeled. From each record, we exploited multiple data instances, resulting in a large enough training dataset to produce a robust classifier. We report the optimum network architecture that suited our study for both the anomaly detection and classification task. The optimal model is tested before and after deployment on-site, we report the quantified performance on a test set via a confusion matrix, and observations about the model’s behaviour on the field. Furthermore, we compare our trials and results on two types of fences, namely rigid and loose, to show how it affects the performance of the trained CNN models, as the signal propagates differently between rigid and loose clotures. We report an overall accuracy of 96.15% for the optimal anomaly detection model, and a lower 52.9% for the 3-class classification model. These results are explained and commented on. Finally, we conclude by providing an educated proposal for future improvements.
KEYWORDS: Acoustics, Sensors, Data modeling, Wavelets, Signal processing, Electronic filtering, Denoising, Signal detection, Signal to noise ratio, Interference (communication)
We aim to classify acoustic events recorded by a fiber optic distributed acoustic sensor (DAS). We derived the information from probing the fiber with light pulses and analyzing the Rayleigh backscatter. Then, we processed this data by a pipeline of processing algorithms to form the input for our machine learning classification model. We put random matrix theory to the test to distinguish the acoustic event of interest from the noise. We conditioned the raw trace using moving average and wavelet-based filtering algorithms to improve the signal-to-noise ratio. For raw, low pass, and wavelet-based filtered signals that we inject into a convolutional neural network, we rely on the magnitude of their complex coefficients to categorize the nature of the event. We also investigate Mel-Frequency Cepstral coefficients specific to the event as an input for the classifier and compare their performance to other signal representations. We run the experiments on the CNN for two-class and three-class classification using datasets from a DAS that is deployed for perimeter security and pipeline monitoring. We obtained the best results when using the MFCCs paired with wavelet denoising, achieving accuracies of 96.4% for the “event” class and 99.7% for the “no event” class when it comes to the two-class process. The three-class process yielded optimal accuracies of 83.3%, 81.3%, and 96.7% for the “digging,” “walking,” and “excavation” classes, respectively. Finally, the training execution time is exceptionally long because the extensive dataset and the model’s architecture are complex. As a result, we make efficient use of the CPU and GPU to maximize our machine’s power using the Keras API’s sequence data generator. Compared with the serial implementation, we report an improvement of up to 4.87 times.
This work aims at the detection and classification of Distributed Acoustic Sensor (DAS) acquired acoustic signals. We obtained the data by probing an optical fiber with light pulses and gauging the Rayleigh backscatter. Said data contains four different classes; Walking, Shovel and Pick digging as well as Hammer hitting. We first proceed by detecting the event and its location along the fiber and extracting it from the random noise using Spiked Random Matrix Theory (RMT) models, namely Marchenko-Pastur (MP) and Tracy-Widom (TW) distributions. We then label the datasets accordingly and proceed with the classification process using machine learning algorithms. For this, we test and evaluate Convolutional Neural Networks (CNN), which has been proven to provide high accuracies in similar studies, taking the spectrograms of the signals as our network’s input. We conclude by providing the performance of our CNN architecture and propose a few options to further improve the performance of the model.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.