Acoustic signals provide characteristic information about vehicles. However, those characteristics vary with engine speed, load, range, gearing, and acoustic environment. The complexity of these variations makes standard parametric classifiers either excessively tedious to construct or unreliable under off-nominal conditions. A new method of classifier construction, Generalized Potential Function Neural Nets, or GP nets, has demonstrated excellent acoustic classification performance under nominal conditions. In addition, GP nets do not require a priori specification of the net structure, number of nodes, or learning parameters. In this effort, GP net classifiers will be constructed and tested under off-nominal conditions using data from the Wide Area Mine program. In addition, the stability of feature extraction will be analyzed under off-nominal conditions. Inconsistent feature extraction algorithms will be improved, if possible. Additional features, which can be accommodated by nonparametric neural net classifiers will be added. In Phase II, longer term line tracking and association algorithms will be investigated to more reliably characterize the engine line harmonics of individual vehicles under low signal conditions or when multiple vehicles are present. This will provide additional features. In addition, the efficiency of the GP net classifier will be optimized in terms of computational requirements. Classification of sounds produced by machinery can effect the solution of many problems. In addition to military ground and airborne vehicle detection and classification, aircraft may be classified to detect covert operations. Machine operations can be monitored and wear may be assessed. More generally, the same general approach can be used for speech recognition, and speaker identification and verification.
