Self-organizing clustering neural networks: Comparative study and data fusion applications

The self-organizing clustering neural network, DIGNET, is compared with ART2 and self-organizing feature maps (SOFM). DIGNET is used for designing a multi-sensor data fusion to achieve multi-radar moving target detection.; SOFM is not applicable in situations where the number of clusters is determined by the system in an autonomous way. Comparative experiments are used to investigate the performance of DIGNET and ART2 on statistical data clustering and detection problems. DIGNET exhibits faster learning and better clustering. With simpler dynamics, DIGNET is more flexible in choosing different metrics as measures of similarity. System parameters in DIGNET are determined from the self-adjusting process. Threshold value in DIGNET can be determined from a lower bound of the desirable signal-to-noise ratio. A simplified ART2 model (SART2), which adopts the structural concepts from DIGNET, exhibits faster learning than ART2 and does not suffer from a "false conviction" syndrome which seems to exist in the "fast learning" ART2.; A two-stage parallel multi-sensor data fusion with DIGNET is applied to a moving target indication (MTI) system. The MTI system consists of three radars with different carrier frequencies. Features of the received data are extracted via digital signal processing. Pulse compression, clutter cancelling, and fast Fourier transform are used to transform data from time-range domain to range-Doppler domain. Map regularization, circular metric, and contrast enhancement are used to resolve the feature misalignment problems. DIGNET performs feature extraction and filtering, and eliminate spurious feature patterns. DIGNET with different metrics, inner-product and hypercube, are used in two approaches. The clustering results from DIGNET on each channel are passed to the fusion DIGNET for a second stage clustering. The well depths in DIGNET act as the indicators of confidence levels. In one approach, well depths are used in the weighted averaging for data combination. In the other approach, well depths are used to determine whether the pattern associated with a winning cluster should be passed to the following processing stage. Experiments show that data fusion with DIGNET successfully detect the moving target embedded in clutter.