| There are two parts in this paper.In the first part,channel inverse analysis is studied in free space optical communications field,especially under rainy condition.In the second part,channel encoding and decoding are studied.Neural network technology is adopted to achieve intelligent coding.Specific researches are as follows:In chapter 1,main difficulties of analysis atmosphere laser communication channel(ALCC)is introduced,especially under rainy condition by means of traditional mathematics methods.Furthermore,a brief description of traditional channel encoding and decoding is introduced,including LDPC,RS.In chapter 2,weights feature extracting technology(WFET),as a novel data analysis tool to extract training data sets' feature information from weight sets in the trained multi-layer perceptron network(MLPN),is proposed,which can achieve in-depth analysis in data sets.Firstly,atmosphere laser communication experimental data sets are adopted to construct training data sets of MLPN.Secondly,train the MLPN with the data sets,and the feature information will transfer into MLPN adequately after training process.Thirdly,extract feature information from trained MLPN with WFET,which can calculate weights energy channels and correlations in data sets.Actual tests show that season's correlation is stronger than rainfall,and the rain intensity's correlation is stronger than season and rainfall factors.The result is an important reference of ALCC under rainy condition.In chapter 3,a channel inverse analysis of ALCC is accomplished in the Tibetan Plateau.Due to lack of raw experimental data sets and proper data analysis method,investigations on ALCC,especially under rainy condition,are rarely concerned by researchers.Neural network group and optimal weight initialization technology(OWIT)are adopted in the analysis process.Firstly,construct neural network group according to different season's conditions.Secondly,utilize existed raw data sets of ALCC under rainy condition to choose matching initial weight sets with OWIT.Thirdly,train neural network group until expected requirement is met.Finally,load rain data sets from the Tibetan Plateau(Lhasa for example)on trained neural network group to achieve the ultimate channel quality of ALCC.Actual results show that spring rain has the best quality of ALCC,followed by winter rain,summer rain and autumn rain.In chaper 4,a channel inverse analysis of low-density parity-check(LDPC)is achieved using neural network technology.Owing to the lack of appropriate data analysis tools,study on the performance of LDPC code has been rarely done or published in ALCC,especially under rainy condition.Forward weights feature extraction(FWFE)is proposed to solve the problem,which is a new way of extracting feature information from trained MLPN.Firstly,raw experimental data sets,which are obtained from ALCC under rainy condition adopting LDPC code,should be transferred into training data sets of MLPN in preprocessing stage.Secondly,construct MLPN according to the size of training data sets and choose initial weight sets matched with training data sets using OWIT.Thirdly,train the MLPN with error back-propagation algorithm until total error is qualified.Finally extract feature information from the trained MLPN using FWFE,which can achieve the correlations for each element over the performance of LDPC.Experiments show that rain density has the biggest impact on the LDPC code's performance(its correlation Re is 39),which is consistent with the result from intuitive analysis of the experiment data;season element has the second biggest impact(its Re is 24);channel quality element has the third impact(its Re is 19),and the rainfall has the least impact(its Re is 18).In chapter 5,a novel general neural network decoder is proposed in the form of symmetrical self-organizing map(SSOM),which can achieve decoding function to any error correcting codes.The SSOM decoder is tested by decoding LDPC code.The performance comparison of SSOM decoder and traditional decoder is accomplished by simulation.Actual results show that the SSOM decoder can implement both learning and decoding in the same time regardless of any encoding rules.And higher possibility of the codeword emergence means greater probability of correct error correction.Compare to traditional error correcting decoder,it is easy to construct,more intelligent to different codeword sets,which has certain prospect in future communication channel coding field.In chapter 6,a forward neural network encoder is proposed,which adopts self-organizing map(SOM)neural network as main structure.Firstly,built the forward neural network according to dimension of source bits and codeword bits.Secondly,initialize the weight sets with certain algorithm.Finally,check uniqueness of the codeword sets until qualified.In decoding process,MLPN is adopted as the neural network decoder.First of all,construct the MLPN according to the dimension of source bits and codeword bits.And secondly,train the MLPN with the codeword sets generated by the forward neural network.Then,stop the training process until total error is qualified.Finally,begin to accept and decode codeword sets.Actual simulation tests show that both neural network encoding and decoding are feasible.And the better performance will be achieved in the condition of proper forward neural network structure and the degree of output node ?.Above all,the codeword sets generated by neural network encoder cannot be decoded by traditional mathematic method,which has certain market prospect for security communication.In chapter 7,an improved neural network encoder is constructed,which adopts SOM neural network to generate check bits individually.And N source bits and K check bits are composed a complete codeword.In the project,both all-connection model and part connection model is simulated.In the decoding port,MLPN is utilized to implement decoding function.The specific steps are as follows:(1)Constructing the MLPN according to the size of codeword sets and source bits;(2)Training the MLPN with codeword sets generated by neural network decoder until qualified;(3)Accepting and decoding codeword sets via trained MLPN.Actual tests show that:(1)There exist no evident performance differences between all-connection model and part-connection model;(2)The connection of weight sets is similar to Tanner graph in part-connection model,which reduce the computational complex greatly and remain good performance at the same time.In sum,the neural network encoding has several advantages:(1)The structure of neural network and corresponding weight sets are the encoding algorithms,which are confidential to some extent;(2)The structure of neural network is simplified in part-connection model,which has less computational complex and good performance.(3)The method of encoding and decoding has certain market prospect in confidential communication field.In chapter 8,a summary of the channel inverse algorithms is implemented.Suitable applications are summarized to different channel inverse analysis algorithms respectively.Both WFET and FWFE have similar features,which extract key information from trained neural network.And neural network group technology is suitable for limited raw experiment data sets,which are lack of source key information.Furthermore,comparisons are outlined among different algorithms.The relationship in the data sets is inversed by menas of weight iteration bias in WFET.Whereas forward calculation is adopted to build the relationship among data sets in FWFE.And the neural network groups are divided in advance according to indirect causal relationship in the data sets.On the other hand,a summary of different encoding and decoding algorithms is accomplished according to their characteristics.Every algorithm has its unique application.On the other hand,comparisons among the algorithms are studied.SSOM decoder is suitable for low dimension,whereas MLPN decoder has better performance to actual application.In neural network encoder field,improved part-connection model has less computational complex and good performance,which has good prospect in future application. |
PDF Full Text Request