Publication: Solución IoT para detección de sonidos medioambientales mediante aprendizaje profundo
Loading...
Official URL
Full text at PDC
Publication Date
2020
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
El trabajo plantea una solución conceptual para detectar e identificar sonidos mediante la aplicación de técnicas inteligentes basadas en Redes Neuronales Convolucionales (aprendizaje profundo) bajo el paradigma de Internet de las Cosas (IoT). El objetivo viene motivado por la necesidad de detectar sonidos de motosierras para prevenir la tala indiscriminada en zonas forestales como fin último. No obstante, la propuesta es aplicable a cualquier entorno donde la monitorización del ruido sea el objetivo principal.
Para ello se proporciona una solución integrada, formada por una serie de módulos para captura, procesamiento y transmisión de datos e información mediante los dispositivos físicos, repositorios y algoritmos correspondientes. Más específicamente, se dispone de un dispositivo móvil de captura de audios, simulado en este caso por audio clips de una duración determinada. Se dispone de un computador donde se procesan los audios, bien en local o en remoto, generando los correspondientes espectrogramas a partir de los audios, que se pasan a un modelo específico de red neuronal convolucional con sus correspondientes capas para el entrenamiento y posterior clasificación en las tres categorías de clases utilizadas, una de ellas la de los sonidos de motosierra.
Al servidor remoto tienen acceso tanto el dispositivo móvil como el computador, para llevar a cabo los procesamientos requeridos; si bien, el primero carece de capacidad para procesos computacionalmente costosos, tal como ocurre con el entrenamiento de la red, por lo que esta fase sólo es posible realizarle en el computador, eso sí, bien en remoto o en local. El modelo de red y los parámetros obtenidos durante el entrenamiento se almacena convenientemente en remoto, con acceso del dispositivo móvil, que realiza el proceso de clasificación. Los resultados de la clasificación se envían a la nube, donde se almacenan para su posterior visualización con fines de monitorización, a la vez que se levantan las correspondientes alarmas de aviso cuando se detectan las acciones de tala no controlada a través de una cuenta Twitter y, por tanto, accesible desde cualquier dispositivo con esta capacidad.
Los resultados obtenidos permiten validar el modelo desde el punto de vista conceptual previsto, tanto desde el punto de vista de los módulos de procesamiento inteligente como del planteamiento dentro del paradigma IoT.
The work proposes a conceptual solution to detect and identify sounds by applying intelligent techniques based on Convolutional Neural Networks (deep learning) under the Internet of Things (IoT) paradigm. The objective is motivated by the need to detect the sounds of chainsaws to prevent indiscriminate felling in forest areas as the goal. However, the proposal is applicable to any environment where noise monitoring is the main objective. To do that, an integrated solution is provided, consisting of several modules, including capturing, processing and data transmission and information through the corresponding physical devices, repositories and algorithms. More specifically, a mobile device captures audios, simulated in this case by audio clips of a specified duration. There is a computer where the audios are processed, either locally or remotely, generating the corresponding spectrograms from the audios, which are fed to a specific model of convolutional neural network with its corresponding layers for training and subsequent classification in the three categories of classes used, one of which is chainsaw sounds. Both, the mobile device and the computer can get access to the remote server, to carry out the required processing; although the former lacks the ability for computationally expensive processes, such as the ones occurred during the network training, so this phase can only be performed on the computer, however, either remotely or locally. The network model and the parameters obtained during the training phase are conveniently stored remotely, with access from the mobile device, which performs the classification process. The classification results are sent to the cloud, where they are stored for later visualization and for monitoring purposes, while the corresponding warning alarms are raised when uncontrolled deforestation actions are detected through a Twitter account and therefore, accessible from any device with this capacity. The results derived from the previous processes allow validating the conceptual model, both from the point of view of the intelligent process, based on the convolutional neural network and the proposed approach within the IoT paradigm.
The work proposes a conceptual solution to detect and identify sounds by applying intelligent techniques based on Convolutional Neural Networks (deep learning) under the Internet of Things (IoT) paradigm. The objective is motivated by the need to detect the sounds of chainsaws to prevent indiscriminate felling in forest areas as the goal. However, the proposal is applicable to any environment where noise monitoring is the main objective. To do that, an integrated solution is provided, consisting of several modules, including capturing, processing and data transmission and information through the corresponding physical devices, repositories and algorithms. More specifically, a mobile device captures audios, simulated in this case by audio clips of a specified duration. There is a computer where the audios are processed, either locally or remotely, generating the corresponding spectrograms from the audios, which are fed to a specific model of convolutional neural network with its corresponding layers for training and subsequent classification in the three categories of classes used, one of which is chainsaw sounds. Both, the mobile device and the computer can get access to the remote server, to carry out the required processing; although the former lacks the ability for computationally expensive processes, such as the ones occurred during the network training, so this phase can only be performed on the computer, however, either remotely or locally. The network model and the parameters obtained during the training phase are conveniently stored remotely, with access from the mobile device, which performs the classification process. The classification results are sent to the cloud, where they are stored for later visualization and for monitoring purposes, while the corresponding warning alarms are raised when uncontrolled deforestation actions are detected through a Twitter account and therefore, accessible from any device with this capacity. The results derived from the previous processes allow validating the conceptual model, both from the point of view of the intelligent process, based on the convolutional neural network and the proposed approach within the IoT paradigm.
Description
Máster en Internet de la Cosas, Facultad de Informática UCM, Departamento de Ingeniería del Software e Inteligencia Artificial, Curso 2019/2020.