As urban areas encroach on once-remote military training bases, noise complaints have increased. Claims against the military for both physical and psychological damage are common and have resulted in the rescheduling, curtailing, or even canceling of military training exercises. Aside from the monetary costs, direct impacts on military preparedness also have resulted. Noise monitoring systems are in place at some bases; however, they tend to suffer from inaccuracies (greater than 10%) and can require extensive data analysis and interpretation by human operators.
The objective of this project was to develop an improved, autonomous military impulse noise detector. Building on the results of the initial SERDP Exploratory Development (SEED) effort (WP-1436), this project sought to (1) expand the noise measurement library; (2) refine the noise classifier; (3) establish hardware requirements for the classifier algorithms; (4) conduct a real-time laboratory demonstration; and (5) develop and demonstrate a prototype noise classifier system.
The project investigated the use of a mix of artificial neural network (ANN) and Bayesian statistical methods to create a robust noise classifier. Researchers collaborated with Applied Physical Sciences Corporation working under SERDP project WP-1427 to develop the Noise Bearing and Amplitude Measurement and Analysis System (BAMAS). Algorithms were developed that work independently of the BAMAS microphone array and in concert with it. Performance was maximized by combining the systems.
The classifier was generalized by collecting large amounts of algorithm training data under various conditions and locations from military installations across the country. The initial key step in this project was collecting a library of high-amplitude operational noise around military bases, including military impulse noise. The library contains actual recorded waveform data, sampled at a sufficient frequency to observe all important features and containing a wide variety of military impulse noise sources. Measurements were conducted for a wide variety of environmental conditions.
With the waveform library in place, the data set was investigated for suitable means of discerning military impulse noise from other high level noise sources. These methods started with the development of signal metrics that were used to train, validate, and test classification algorithms for identifying military impulse noise.
When suitable metrics and an accurate classification algorithm were in place, two noise classifier prototypes were developed and demonstrated in the field. Performance of BAMAS, the UPitt Classifier, and a combined system was recorded through remote controls and contrasted to the performance of existing noise monitors.
In all, 11,600 waveforms were recorded. Several ANN structures were investigated, including multi-layer perceptron (MLP), self-organizing map (SOM), support vector machine, and an image recognition network. A Bayesian classifier was also created that used Gaussian mixture models to fit the multivariate data. The MLP, SOM, and Bayesian classifiers were found to perform to nearly 100% accuracy during development. The UPitt algorithm was combined with the BAMAS in a PC/104-based hardware platform.
Both algorithms performed well individually. While the overall accuracy of the UPitt classifier alone was somewhat lower (89%) than desired, nonblast noise was rejected at over a 99% rate (which meets the objective to reduce false positives). When the UPitt and BAMAS algorithms were combined, blast noise was identified at over 98% accuracy, while aircraft, wind, and vehicle noise were rejected at over 98-99% accuracy. The device is at Technology Readiness Level 7, ready for demonstration and validation, before being commercialized.
The noise detectors developed can support military readiness by providing better management of noise encroachment issues and will extend the state of the science for noise monitoring. Noise monitoring accuracy was improved while requiring less human interpretation. This improvement will facilitate responses to damage claims and provide instant feedback if severe noise levels are produced.