As urban areas expand toward military training and testing facilities, a number of noise-related concerns have arisen. The high noise levels associated with military training exercises, specifically impulse-type noises, have impacted land use, wildlife habitats, and the well-being of domestic animals. Humans also have been impacted through physiological damage, psychological discomfort, and property damage. The resulting damage claims and restrictions on the scope and frequency of training and testing have been costly to the military. Although current monitoring systems are greatly improved over original designs, they have a relatively high false positive rate (10%) and require extensive data analysis and interpretation by human operators. A method of detecting military noise sources autonomously and accurately is needed.

The long-range goal of this project was to develop a novel, real-time impulse-noise monitoring system capable of discerning between military impulse noise (cannon, explosives, howitzer, etc.) and non-impulse noise (wind noise, aircraft noise, and other sources) and of detecting events with much lower peak levels. Initial SERDP Exploratory Development (SEED) efforts focused on developing the software algorithms that constitute an automated noise classification system.

Technical Approach

The classification algorithm consists of an artificial neural network (ANN) structure whose inputs are conventional acoustic metrics used to indicate impulsiveness of a signal and additional novel frequency-domain metrics. Because no waveforms were available to develop a noise classifier, they were measured as part of the study. Specific steps for executing the project included collecting a library of military impulse noise and non-impulse noise (potential false positives) field measurements and developing software tools that together constitute the noise classification system. Recorded acoustic time histories are necessary for the training of the noise classifier. Measurements were performed at two representative military bases—Camp Lejeune, North Carolina and Fort Indiantown Gap, Pennsylvania—under differing conditions to include environmental effects on noise propagation. Wind noise, which is believed to be responsible for the majority of false positives in the current monitoring systems, was recorded along with aircraft and other loud potential sources of false positives. The software consisted of two parts: one that computed signal metrics and a second that performed the classification based upon these metrics. The classification algorithm consisted of a multi-layer ANN, whose topology was heuristically optimized. The development of the classifier followed an iterative process, in which approximately two-thirds of the recorded waveforms were used for training and the other third for evaluation.


Under SERDP Project WP-1585, follow-on research is under way to expand the noise measurement library, refine the noise classifier algorithms, establish hardware requirements, create a real-time laboratory demonstration, and develop and demonstrate a prototype noise classifier system.


A fully trained noise classification system would be capable of autonomous, real-time recognition of signal sources and could even be incorporated into military training exercises to take the guesswork out of determining whether training conditions are favorable. The developed classifier achieves essentially 100% classification accuracy and can detect signals with much lower peak acoustic levels (<100 dB). Because it does not depend upon a particular shape in the impulse waveform, simultaneous impulse noise and other mixed noise environments are more accurately classified. Detection of events with lower peak levels would be beneficial to civilian community outreach programs concerning noise complaints. (SEED Project Completed - 2006)