A noise monitoring system with domain adaptation based on standard parameters measured by sound analyzers

Continuous and unattended noise monitoring systems are currently used in airports, railways and roads to actively monitor the noise sources present in these scenarios. The noise monitoring process involves, among other tasks, the detection and classification of the captured acoustic events, using the different procedures to identify the events of interest allowed by international standards. These procedures ranging from those relying on a human operator carefully trained to recognize the nature of each event (by evaluating the time profile of an unique parameter, mainly the LA(eq)), to those based on automatic processing and machine learning. However, most existing automatic detection approaches for acoustic events: 1) rely on complex features beyond those provided by standard measurement systems, or even use raw audio as input; and 2) need to be trained in multiple scenarios under different conditions. As a result, these solutions are difficult to embed in standard real-time sound analyzers and require extensive training to adapt to each scenario. In this paper, we propose both an automatic event detector that attempts to reproduce the expertise of a human operator, and a domain adaptation procedure that is able to transfer the knowledge acquired in a favorable context to a more challenging scenario. We have extensively tested our solution in a railway pass-by detection application, and its results outperform those of state-of-the-art solutions with significantly lower computational cost. The proposed event detector is robust, interpretable, easily deployable, and uses features that are directly obtained from the measurement systems, allowing its implementation into the sound analyzer software or a low-cost portable device.