Noise Signal Identification by Modified Self- Organizing Maps

This research uses a modified self-organized map to look for similarities and differences between noise signals and provides a context for unknown ones. The program first divides preprocessed data into quadrants in joint time-frequency domain and obtains the features. Features are also extracted in time domain, frequency domain, and joint time-frequency domain from noise files representing different abstract noise colors. These features are able to be input independently and classified accordingly. In addition, several types of distance metrics are tested to find best matching units including the p-norm distance formula and determining which node has the smallest weight. The output is displayed on a two-dimensional map, and the vector is colored according to the noise that has the smallest weight. A context tree allows individual features to be input and classifies the vector to the most similar noise. With the results that were gathered, the clusters have shown that distinct differences between the engine sets contribute to the study of sound discrimination by distinguishing useful sounds that humans have difficulty telling apart. Keywordsself-organizing feature map; sounds; unsupervised learning; classification; feature selection; watershed tranformation