Research On Blind Signal Separation Algorithm Based On Deep Learning

In many practical applications,it is not possible to obtain independent observations of the original signal directly,and often only mixed signals are available,making it difficult to process and analyse each signal independently.How to recover the source signal from the mixed signal has become a key concern.The traditional blind signal separation technique requires mining the mathematical properties of the signal,using the higherorder statistical properties of the signal,sparse properties,etc.to complete the separation of the signal.However,this separation technique requires the probability density function of the source signal to correspond to the properties of the non-linear function applied in the algorithm,resulting in a limitation on the type of source signal to which the algorithm is applicable and the best results that the algorithm can achieve.Deep learning techniques,as a supervised learning method,have better function fitting and adaptation capabilities that can compensate for the lack of expressiveness in traditional blind signal separation algorithms.Therefore,this thesis focuses on the Independent Component Analysis(ICA)algorithm based on higher-order statistical properties,hoping to combine neural networks with blind signal separation algorithms,so that the separation effect of the algorithm can be improved.The main research work is summarised as follows:1)In this thesis,an ICA-enhanced network structure based on interpretable neural networks is proposed in conjunction with the properties of the algorithm itself to solve the problem of fixed mathematical forms of the correlation functions of iterative expressions and restricted types of separated signals in blind separation algorithms solved by higher-order statistical properties.The method uses interpretable networks instead of non-linear functions and introduces a deep learning process in the iterative computation to design a data-driven neural network structure.The proposed ICA-enhanced network can optimise the computational process of the blind signal separation algorithm and improve the separation effect of the algorithm.Experimental results show that the algorithm can effectively improve the accuracy of separating mixed signals composed of different waveforms compared to the blind signal separation algorithm with minimum BER criterion.In addition,this thesis further demonstrates the good separation effect of the algorithm from a mathematical perspective by using cosine similarity as the evaluation criterion.2)This thesis proposes the use of polynomial substitution for nonlinear functions and the design of parameter-optimised ICA neural networks to reduce the time required to train the parameters of the blind signal separation algorithm while ensuring the accuracy of the algorithm’s separation.The method reduces the execution time of the algorithm by using a small number of parameters involved in the iterative computation.The higher the order of the polynomial,the more accurate approximation to the non-linear function can be achieved.Compared to ICA-enhanced networks,the algorithm is able to reduce the time required to train the algorithm while ensuring that the separation waveforms are accurate.In addition,the separation effectiveness of the algorithm is evaluated using cosine similarity to verify the separation effectiveness of the algorithm.3)This thesis designs a fully connected form of neural network to approximate the complete iterative process of the ICA algorithm to achieve an equivalent replacement of the iterative algorithm and reduce the number of calculations of the blind signal separation algorithm.The method first designs the order of the multivariate polynomial based on the variable parameters of the minimum BER criterion algorithm and the number of dependent variables,and calculates the number of layers and neurons of the neural network according to the method of approximating polynomials based on neural networks.The experiments show that the separation results of the proposed neural network greatly reduce the number of iterations of the algorithm and improve the efficiency of the algorithm’s computation at the expense of certain accuracy.