Unattended sensor using deep machine learning techniques for rapid response applications

摘要

The ability for sensing platforms to collect data intermittently in various settings has been explored extensively. However, many existing solutions are not intelligent and cannot be implemented in real-time. This paper addresses the need for a near real-time, low-cost intelligent autonomous unattended sensors (AAUS) integrating an interchangeable a mobile radiation sensor, with the ability to transmit actionable information to a base station. We address this through discussion of current technologies, our implementations, and experiments as well as a complete pipeline for future frameworks. Our method continuously listens for specific frequencies with the ability measure radiation counts, implements onboard audio classification via machine learning methods, and transmits the results requested. This technique utilizes existing hardware for data management and machine learning algorithms for classification, such as an inexpensive single board computer, a Artificial Neural Network (ANN) and a bgeigie Nano radiation sensor. Our approach performs a real-time Fast Fourier Transform (FFT) continuously in an environment and calculates whether the frequency is within the range of interest. If correct, the sound is recorded, and a pre-trained ANN, fine-tuned on specific data will classify the recorded sound. Depending on the requested information the node will either transmit radiation counts or the classification of the audio input. However, the transmission of audio will only occur if the degree of certainty is above a threshold value. The onboard shallow ANN implentation in this paper experiences an overall classification of 64%.