Conference paper

Cognitive Acoustic Analytics Service for Internet of Things

View publication


The rapid development of the Internet of Things (IoT) has brought great changes for non-contact and non-destructive sensing and diagnosis. For every inanimate object can tell us something by the sound it makes, acoustic sensor demonstrates great advantages comparing to conventional electronic and mechanic sensors in such cases: overcoming environmental obstacles, mapping to existing use cases of detecting problems with human ears, low cost for deployment, etc. It could be widely applied to various domains, such as predictive maintenance of machinery, robot sensory, elderly and baby care in smart home, etc. Whether we can use the acoustic sensor data to understand what is happening and to predict what will happen relies heavily on the analytics capabilities we apply to the acoustic data, which has to overcome the obstacles of noise, disturbance and errors, and has to meet the requirement of real-time processing of high volume signals with large number of sensors. In this paper, we propose a scalable cognitive acoustics analytics service for IoT that provides the user an incremental learning approach to evolve their analytics capability on non-intuitive and unstructured acoustic data through the combination of acoustic signal processing and machine learning technology. It first performs acoustic signal processing and denoising, enables acoustic signal based abnormal detection based on sound intensity, spectral centroid, etc. Then based on the accumulated abnormal data, a supervised learning method is performed as baseline and a neural network based classifier is used to recognize acoustic events in different scenarios with various volume of sample data and requirement of accuracy. In addition, acoustic sensor arrays processing is supported for localization of moving acoustic source in more complex scenario. In this paper, we designed a hybrid computing structure. Finally, we conduct experiments on acoustic event recognition for machinery diagnosis, and show that the proposed system can achieve high accuracy.