Research On Signal Processing And Detection Of Optical Communication Based On Machine Learning

Benefiting from the thriving machine learning,machine learning provides powerful tools to deal with many areas such as natural language processing,data mining,speech recognition and image recognition.Its typical characteristics are self-learning and evolution,as long as there is With new data,new mapping networks can be built by adjusting structures and parameters to further create new capabilities.Signal equalization and optical spectroscopy are important issues in signal processing in optical communications,but there are few studies using machine learning to solve this problem.In optical communication,due to the multipath effect and the noise caused by the channel,the transmission characteristics are not ideal.The existing equalization technology requires the processing speed of the training sequence to be reduced.The complex optical communication system application requires the equalizer to reduce the error.Significantly,processing high rate signals is more timely.As a common expression of a signal in the frequency domain,the spectrogram conveys another dimension of optical information.Traditional optical spectrum analysis techniques are applied to the spectrum analyzer,although a higher resolution is achieved in spectral detection.The rate and wider wavelength range have greater sensitivity and greater dynamic range for power detection,but these solutions are based on different hardware implementation techniques and cannot achieve lower complexity applications at the software level.This thesis focuses on signal quality improvement and performance parameter monitoring in optical communication,and applies machine learning to visible light communication equalization and spectrum analysis.Combining machine learning algorithms with equalization techniques,enhancing the ability to track channel characteristics,and intelligent learning and updating of equalizers,another part of the main work is to introduce machine learning into spectrum analysis and make performance parameters from the spectrum.More accurate qualitative analysis and more accurate quantitative analysis.The main work is as follows:First,in order to meet the demand of higher speed,the unsupervised learning method in machine learning is creatively applied to the WDM-DCO-OFDM visible light communication offline system combining wavelength division multiplexing and orthogonal frequency division multiplexing,and fuzzy C is adopted.The mean clustering method softly classifies the received signals,and obtains the membership degree of the cluster center and each signal,which reduces the computational complexity and achieves a better equalization effect;then applies the gradient descent algorithm to the equalizer.The coefficient vector is adjusted and updated to obtain the signal after the equalization processing,and the purpose of achieving convergence can be automatically adjusted without training.The transmission distance of the system is 20cm,the transmission rate reaches 1Gbps,and the bit error rate is reduced to 10-5 orders of magnitude.It can automatically adjust to achieve convergence without training,reduce inter-symbol interference,and greatly improve the accuracy of signal transmission.It has a good application prospect in the non-cooperative acceptance environment.Secondly,in view of the lack of intelligent learning in the current decision algorithm,the machine learning technology is introduced into the spectrum analysis of optical communication.It is proposed that the machine learning method can be used to extract the spectrum of the discrete data for spectrum analysis.The spectrum map trains the data set,and then selects four typical machine learning algorithms:artificial neural network,support vector machine(SVM),decision tree and K nearest neighbor(KNN),etc.,and inputs the data to be trained to the selected machine learning.The algorithm is trained,and then the new spectral data is analyzed and input into the trained convolutional neural network module for feature extraction and performance analysis.The support vector machine(SVM)algorithm has the best effect on the wavelength and signal to noise ratio.The identification accuracy of the bandwidth and bandwidth is up to 100%,and the recognition time is also the shortest,which is 0.238s,0.338s and 0.443s respectively.Thirdly,for the traditional optical performance analysis module,an algorithm can only identify one parameter,and the system has high complexity and low fitness.A new method is proposed to solve the problem that the discrete data dimension is too high or uncertain.The resulting model structure does not have the problem of versatility.An intelligent spectrogram analysis method for deep learning of convolutional neural networks that automatically adapts to new scenes and new demands by automatically detecting and extracting features to self-learn and evolve,and when there are new recognition targets,according to data and training New recognition ability is added,and since the input information format is an image,the dimension of the information is determined,the structure of the model is also fixed,and both function expansion and versatility are combined.