Research On Two Kinds Of Aliasing Signals Translation Based On Recurrent Neural Network

In recent years,signal separation has become a hot topic in the fields of acoustic signal processing,communication,and optical signal processing.It has a wide range of applications,such as ultrasonic flaw detection technology and Faster than Nyquist(FTN)communication technology.Ultrasonic flaw detection is widely used in industrial and medical fields due to its non-destructive and simple implementation characteristics.However,when the actual bandwidth is limited,ultrasonic echoes often produce aliasing and affect the detection results.Therefore,improving the distance resolution is the key for ultrasonic flaw detection.On the other hand,FTN communication technology transmits information at a faster rate to improve spectrum utilization,which meets the requirements of increasingly higher frequency band in modern mobile communications and attracts widespread attention.However,transmission at FTN rate is often accompanied by inter-symbol interference.The waveform shows severe aliasing.However,most of the current methods applied to these two kinds of aliasing signal processing are high in complexity,sensitive to parameters,and low in stability,and it is difficult to have a significant effect.In view of the above problems,this thesis innovatively introduces a recurrent neural network(RNN)to separately study the methods for solving these two types of aliasing signals.First of all,in order to solve the problems of over-reliance on dictionary design,parameter sensitivity,slow detection speed,and inability to separate severe aliasing waves in traditional algorithms,a high-resolution ultrasonic echo detection method based on RNN was proposed.Based on the idea of Two-stage,an aliasing ultrasonic echo detection network was constructed which including classification and regression.Firstly,a sequence-to-sequence RNN was used to establish a classification module to detect whether the corresponding position is an echo.That is,the aliasing echo signal was divided into sequences according to a certain time step and then perform binary classification.Then,we combined the classification result with the original input sequence to pick out all the sequences that are classified as echoes.These echoes were feed into an RNN regression module to predict the amplitude of the echo at the corresponding position.Thus we completed the amplitude and position detection of all echoes in the ultrasonic aliasing echo signal.The proposed method draws on the idea of Two-stage which filters a large number of interfering signals through classification and then perform amplitude prediction.It obtains high-precision detection results,and also performs well under a large degree of aliasing.It realized high-resolution ultrasonic echo detection,while greatly improving the detection speed.Secondly,according to the advantages of RNN processing time series data and its successful application in aliasing ultrasonic echo detection,this thesis innovatively introduced a bidirectional long short-term memory network(BLSTM)to solve the decoding problem of FTN received signals.Focus on the characteristics of inter-symbol interference introduced by FTN transmission signals,a BLSTM-based FTN decoding network was proposed.It took advantage of the ability of BLSTM to have both "long-term" and "short-term" memory and the ability to combine past and future information of bi-directional loop.Under the condition that the data transmission rate is doubled than the limit Nyquist rate,the system guarantees a certain bit error rate performance under different signal-to-noise ratios,that is,the symbol pulse wave amplitude detection performance.In order to further explore the decoding ability of the proposed method,amplitude modulation was performed.The experimental results proved that the proposed method has good decoding ability for FTN received signals,which also provided a new idea for related research on FTN decoding.In summary,this thesis utilized the unique advantages of RNN to process time series data,and according to characteristics of aliasing ultrasonic echo detection and FTN received signal decoding,proposed corresponding aliasing signal deciphering methods based on suitable RNNs respectively.