Application of neural networks to bearing estimation

This study presents an application of a feedforward neural network (NN) structure to the bearing estimation problem. Using N snapshots from M sensors, the NN estimates the sensor-to-sensor propagation delays, which yield the far-field source location. The proposed network has only one output, which is the direction-of-arrival (DOA) angle. Thus, the network does not require any preprocessing. The NN buffers the sensor data, treats them as multidimensional delayed patterns and gives the location of a sinusoidal signal source in a noisy environment as output. Networks with various hidden nodes are tried with various sensor and snapshot numbers to find the best performance network structure. The effect of intersensor spacing on the performance is investigated. Using the best performance giving structure, the network is trained with various signal to noise ratios (SNRs) and then tested for various SNR levels.