Research On Data-Driven And Model-Guided High-Resolution Time-Frequency Distribution Methods

Most real-life signals have non-stationary characteristics.As a powerful tool in the analysis of non-stationary signals,time-frequency(TF)analysis is helpful for obtaining accurate instantaneous frequency estimation,separation and detection results,which is of great significance in many practical applications,e.g.,radar,communications,and medicine.However,the analysis of non-stationary signals has been hampered by issues of cross-terms(CTs),low resolution,noise and manually optimized parameters,which make it difficult to visually interpret TF distributions(TFD)of the signal composed of closely-located and overlapped components.Kernel function design and sparse TF reconstruction algorithms often have the trade-off between negligible CTs and high-resolution,which is the bottleneck of restricting the time-frequency analysis.Benefitting from sufficient training data,TF analysis has found a new way inspired by deep learning.Therefore,combined with the data-driven and modelguided strategies,this thesis is devoted to achieving CT-free and high-resolution TFDs.The main research contents and contributions are as follows:1.Considering that kernel function design algorithms have limitations in manually selected parameters and the trade-off between CTs reduction and high-resolution,a data-driven model of learnable high-resolution TFD is proposed.Firstly,we employ typical convolutional layers to simulate traditional kernel functions,removing the CTs coarsely and meanwhile extracting TF features.Then,taking the various directional features of auto-terms(ATs)and CTs into consideration,weighted blocks are employed to explore dependencies among spaces and channels.It can refine coarse features extracted by the convolutional layers to obtain high-resolution TFD results.Compared with existing kernel function design and reassigned algorithms,the proposed model breaks through the trade-off between negligible CTs and highresolution in the cases of synthetic and real-life signals under high noise levels.Additionally,the accurate instantaneous frequency estimation results are obtained.2.Considering that the sparse TF reconstruction algorithm performs poorly on the signals with closely-located or overlapped components and it has a disadvantage in CTs reduction,a model-guided structure sparse TF reconstruction algorithm is proposed.Firstly,the compressive sensing model is built up in the TF domain,utilizing iterative shrinkage-thresholding algorithm(ISTA)to optimize this model.Then,the ISTA is unfolded as a fixed feed-forward network,and fast convergence rate as well as low error rate can be gained by learnable parameters.Further,as the ATs are smooth components with continuous structure and the cross-terms have oscillatory characteristic,the adaptive structure-aware threshold block is introduced to exploit the structured sparsity of signal’s TFDs.With a proper selection of threshold,ATs will be retained while CTs will be reduced,thus recovery error can be further reduced.Compared with existing sparse TF reconstruction,kernel function design,and reassigned algorithms,the proposed algorithm improves the performance on CTs reduction,achieving desirable TFD results in the cases of synthetic with closely-located and overlapped components.Based on the TFD,accurate instantaneous frequency estimation results are obtained.Besides,the CTfree and high-resolution TFD result is gained in the case of a real-life bat echolocation signal,which examines that the proposed algorithm has generalization performance to some extent.