Abstract: Based on Perception Linear Prediction (PLP), an approach to extract features from underwater target signal is presented. The method is the simulation of hearing property of human beings. Through the auditory psychology, three auditory spectrums are estimated, and they are Critical band, Equal-loudness curve and Intensity-loudness power. Then, a twelve-dimension feature vector is obtained. The vector is also a twelve-order all-pole model and it is robust. With the feature vector , the training and recognition processes are performed. The real sea experiments prove that human ear is at different level of sensitivity with different frequency bands where six kinds of radiated noises exist respectively, that the underwater target features are robust, that the dimension is relatively lower and the computation is less expensive and the recognition ratio may arrive 91% to six kinds of underwater target noise.