Model-based Channel Estimation For Underwater Acoustic Communications

Underwater acoustic(UWA)communication is the main approach to realize underwater information exchange,and it plays an important role for human to explore ocean and exploit marine resources.The physical characteristics of acoustic wave transmission and the diversity of the actual underwater environment lead to complex characteristics of UWA communication channel,and it is challenging to achieve reliable UWA communication.Accurately estimating channel state and tracking its changes is of great significance for UWA communication.On one hand,it can provide necessary information for channel equalization and symbol detection,and on the other hand,by means of the channel estimation,the UWA communication system can be simulated close to the actual underwater environment,thus bringing convenience to the system performance evaluation.Therefore,the channel estimation is significant for the UWA communication system.This paper studies the model based UWA channel estimation technology,and the main contents and innovations are as follows.Firstly,an improved scheme is proposed for the existing autoregressive(AR)state model based Kalman filter UWA channel estimation.The existing scheme uses the least square(LS)method to obtain the channel response estimation sequence based on the pilot sequence,then estimates the channel autocorrelation and obtains the AR model coefficients through the Yule-Walker equation.The disadvantage of the above scheme is that the channel needs to be assumed to remain unchanged within the data block and the complexity increases with the cubic of the channel length,which is not suitable for UWA channel.The normalized least mean square(NLMS)adaptive algorithm is proposed to replace the LS algorithm,which reduces the computational complexity and does not depend on the assumption of channel invariance within a block,thus improving the AR coefficient estimation accuracy.Besides,the AR coefficients are dynamically updated using the estimated channel response sequence to improve the model accuracy and estimation performance.Noise initialization has an important impact on the performance of Kalman filter channel estimation and the existing method set it based on experience.In view of the shortcoming of existing method,an improved method based on normalized estimation error squared(NEES)and normalized innovation squared(NIS)is proposed to evaluate the filtering consistency,which is able to obtain reliable initialization of noise parameters.Secondly,two new Kalman filter channel estimation schemes based on the characteristics of UWA communication channels are proposed.First,the actual UWA channel often exhibits sparse characteristic,and inspired by the proportionate updating sparse adaptive filtering algorithm,a proportionate matrix is introduced in the update step of the standard Kalman filter algorithm to obtain a proportional Kalman filter(PKF)algorithm.The simulation and sea experiment results show that the UWA channel estimation scheme based on PKF can effectively use the sparse characteristic of the channel to obtain better performance than the standard Kalman filter channel estimation.Second,the observation noise in the received signal of the actual UWA communication includes environmental noise,external interference and sensor measurement error,which generally does not follow the Gaussian distribution and needs to be described by the distribution with heavy tail.Therefore an outlier robust Kalman filter algorithm is proposed to replace the standard Kalman filter algorithm for channel estimation.The simulation and sea experiment results show that the above scheme can effectively suppress the interference and improve the channel estimation performance.Finally,in view of the limitation of AR state model,the Gaussian process(GP)based UWA channel estimation is explored.As a non-parametric Bayesian learning method in machine learning,GP can better characterize the channel state transition in complex scenes.In the designed channel estimation scheme,GP is used to realize channel state prediction,and Kalman filter is used to realize channel state updating.The simulation result shows that channel estimation scheme which combines the GP prediction with the Kalman filter updating is better than the AR state model based Kalman filter channel estimation.