Research On Shipborne PNT Data Fusion Filter Algorithm Based On Deep Learning

With the emergence of intelligent ships and the development of its core technologies,higher requirements are put forward for positioning,navigation and timing(PNT)information in the maritime field.At present,Global Navigation Satellite System(GNSS)is the primary PNT source of most shipboard equipment,as it has the advantages of no error accumulation,etc.However,GNSS signals are susceptible to obscuration,interference,and deception,which makes it difficult to provide stable and reliable positioning services for ship navigation.Combining various PNT information sources on the ship side by multi-sensor information fusion has become a hot research topic in the maritime field,which can provide comprehensive PNT information with anti-interference,continuous availability,stability,and reliability.The traditional combined navigation usually uses a filtering method to fuse the information of Inertial Navigation System(INS)and GNSS.Due to the complex and changing environment during the ship sailing,the measurement noise in the filtering method is challenging to be modeled to estimate accurately.The noise uncertainty becomes a key factor in limiting the filtering accuracy of combined navigation.This paper introduces a Stacked Sparse Auto encoder(SSAE)into the INS/GNSS combined navigation structure.The excellent fiting ability and high precision on data fusion of deep learning make it easy to estimate the noises in the shipborne PNT data fusion algorithm.It is in line with the development direction of intelligent ship PNT.The research includes the following parts.INS/GNSS combined navigation structure and the SSAE neural network model algorithm are studied.Because the measurement noise in the combined navigation information fusion filtering process is difficult to model and estimate accurately,an SSAE-based EKF filtering algorithm for INS/GNSS combined navigation is proposed to solve the above problem.And the SSAE-based INS/GNSS loose-combined and tight-combined structure mathematical models are derived.The SSAE neural network model is trained using historical sensor data during navigation.The original measurements input in real-time is preprocessed based on the trained model to reconstruct the noise-reduced sensor data.The dynamic scaling multiplication factor of the measurement noise covariance is calculated by comparing reconstructed data with the original data and then delivering it into the filtering iteration afterward.On the basis of the above theoretical study,the SSAE-based EKF algorithms for INS/GNSS loose-combination and tight-combination navigation are simulated and validated.The experimental results show that the performance of the SSAE-based INS/GNSS combined navigation algorithm is better than the traditional algorithm.Under the looser combined structure,the horizontal position error is decreased by 14.15%.Under the tight combined structure,the horizontal position error is reduced by 10.73%.